This paper is not motivated to seek innovation within the attention mechanism. Instead, it focuses on overcoming the existing trade-offs between accuracy and efficiency within the context of point cloud processing, leveraging the power of scale. Drawing inspiration from recent advances in 3D large-scale representation learning, we recognize that model performance is more influenced by scale than by intricate design. Therefore, we present Point Transformer V3 (PTv3), which prioritizes simplicity and efficiency over the accuracy of certain mechanisms that are minor to the overall performance, such as replacing the precise neighbor search by KNN with an efficient serialized neighbor mapping of point clouds organized with specific patterns. This principle enables significant scaling, expanding the receptive field from 16 to 1024 points while remaining efficient (a 3$\times$ increase in processing speed and a 10$\times$ improvement in memory efficiency compared with its predecessor, PTv2). PTv3 attains state-of-the-art results on over 20 downstream tasks that span both indoor and outdoor scenarios. Further enhanced with multi-dataset joint training, PTv3 pushes these results to a higher level. Code will be available.

Given two images, we can estimate the relative camera pose between them by establishing image-to-image correspondences. Usually, correspondences are 2D-to-2D and the pose we estimate is defined only up to scale. Some applications, aiming at instant augmented reality anywhere, require scale-metric pose estimates, and hence, they rely on external depth estimators to recover the scale. We present MicKey, a keypoint matching pipeline that is able to predict metric correspondences in 3D camera space. By learning to match 3D coordinates across images, we are able to infer the metric relative pose without depth measurements. Depth measurements are also not required for training, nor are scene reconstructions or image overlap information. MicKey is supervised only by pairs of images and their relative poses. MicKey achieves state-of-the-art performance on the Map-Free Relocalisation benchmark while requiring less supervision than competing approaches. The code will be made publicly available.

The reflective nature of the human eye is an underappreciated source of information about what the world around us looks like. By imaging the eyes of a moving person, we capture multiple views of a scene outside the camera's direct line of sight through the reflections in the eyes. In this paper, we reconstruct a radiance field beyond the camera's line of sight using portrait images containing eye reflections. This task is challenging due to (1) the difficulty of accurately estimating eye poses and (2) the entangled appearance of the iris textures and the scene reflections. To address these, our method jointly optimizes the cornea poses, the radiance field depicting the scene, and the observer's eye iris texture. We further present a regularization prior on the iris texture to improve scene reconstruction quality. Through various experiments on synthetic and real-world captures featuring people with varied eye colors, and lighting conditions, we demonstrate the feasibility of our approach to recover the radiance field using cornea reflections.

Depth completion is a vital task for autonomous driving, as it involves reconstructing the precise 3D geometry of a scene from sparse and noisy depth measurements. However, most existing methods either rely only on 2D depth representations or directly incorporate raw 3D point clouds for compensation, which are still insufficient to capture the fine-grained 3D geometry of the scene. To address this challenge, we introduce Tri-Perspective View Decomposition (TPVD), a novel framework that can explicitly model 3D geometry. In particular, (1) TPVD ingeniously decomposes the original point cloud into three 2D views, one of which corresponds to the sparse depth input. (2) We design TPV Fusion to update the 2D TPV features through recurrent 2D-3D-2D aggregation, where a Distance-Aware Spherical Convolution (DASC) is applied. (3) By adaptively choosing TPV affinitive neighbors, the newly proposed Geometric Spatial Propagation Network (GSPN) further improves the geometric consistency. As a result, our TPVD outperforms existing methods on KITTI, NYUv2, and SUN RGBD benchmarks. Furthermore, we build a novel depth completion dataset named TOFDC, which is acquired by the time-of-flight (TOF) sensor and the color camera on smartphones. Project page.

Image keypoint descriptions that are discriminative and matchable over large changes in viewpoint are vital for 3D reconstruction. However, descriptions output by learned descriptors are typically not robust to camera rotation. While they can be made more robust by, e.g., data augmentation, this degrades performance on upright images. Another approach is test-time augmentation, which incurs a significant increase in runtime. Instead, we learn a linear transform in description space that encodes rotations of the input image. We call this linear transform a steerer since it allows us to transform the descriptions as if the image was rotated. From representation theory, we know all possible steerers for the rotation group. Steerers can be optimized (A) given a fixed descriptor, (B) jointly with a descriptor or (C) we can optimize a descriptor given a fixed steerer. We perform experiments in these three settings and obtain state-of-the-art results on the rotation invariant image matching benchmarks AIMS and Roto-360.

Recent innovations on text-to-3D generation have featured Score Distillation Sampling (SDS), which enables the zero-shot learning of implicit 3D models (NeRF) by directly distilling prior knowledge from 2D diffusion models. However, current SDS-based models still struggle with intricate text prompts and commonly result in distorted 3D models with unrealistic textures or cross-view inconsistency issues. In this work, we introduce a novel Visual Prompt-guided text-to-3D diffusion model (VP3D) that explicitly unleashes the visual appearance knowledge in 2D visual prompt to boost text-to-3D generation. Instead of solely supervising SDS with text prompt, VP3D first capitalizes on 2D diffusion model to generate a high-quality image from input text, which subsequently acts as visual prompt to strengthen SDS optimization with explicit visual appearance. Meanwhile, we couple the SDS optimization with additional differentiable reward function that encourages rendering images of 3D models to better visually align with 2D visual prompt and semantically match with text prompt. Through extensive experiments, we show that the 2D Visual Prompt in our VP3D significantly eases the learning of visual appearance of 3D models and thus leads to higher visual fidelity with more detailed textures. It is also appealing in view that when replacing the self-generating visual prompt with a given reference image, VP3D is able to trigger a new task of stylized text-to-3D generation.

Generalizable NeRF can directly synthesize novel views across new scenes, eliminating the need for scene-specific retraining in vanilla NeRF. A critical enabling factor in these approaches is the extraction of a generalizable 3D representation by aggregating source-view features. In this paper, we propose an Entangled View-Epipolar Information Aggregation method dubbed EVE-NeRF. Different from existing methods that consider cross-view and along-epipolar information independently, EVE-NeRF conducts the view-epipolar feature aggregation in an entangled manner by injecting the scene-invariant appearance continuity and geometry consistency priors to the aggregation process. Our approach effectively mitigates the potential lack of inherent geometric and appearance constraint resulting from one-dimensional interactions, thus further boosting the 3D representation generalizablity. EVE-NeRF attains state-of-the-art performance across various evaluation scenarios. Extensive experiments demonstate that, compared to prevailing single-dimensional aggregation, the entangled network excels in the accuracy of 3D scene geometry and appearance reconstruction. Our code is publicly available at https://github.com/tatakai1/EVENeRF.

Self-supervised 3D representation learning aims to learn effective representations from large-scale unlabeled point clouds.Most existing approaches adopt point discrimination as the pretext task, which assigns matched points in two distinct views as positive pairs and unmatched points as negative pairs. However, this approach often results in semantically identical points having dissimilar representations, leading to a high number of false negatives and introducing a semantic conflict problem. To address this issue, we propose GroupContrast, a novel approach that combines segment grouping and semantic-aware contrastive learning. Segment grouping partitions points into semantically meaningful regions, which enhances semantic coherence and provides semantic guidance for the subsequent contrastive representation learning. Semantic-aware contrastive learning augments the semantic information extracted from segment grouping and helps to alleviate the issue of semantic conflict. We conducted extensive experiments on multiple 3D scene understanding tasks. The results demonstrate that GroupContrast learns semantically meaningful representations and achieves promising transfer learning performance.

iToF is a prevalent, cost-effective technology for 3D perception. While its reliance on multi-measurement commonly leads to reduced performance in dynamic environments. Based on the analysis of the physical iToF imaging process, we propose the iToF flow, composed of crossmode transformation and uni-mode photometric correction, to model the variation of measurements caused by different measurement modes and 3D motion, respectively. We propose a local linear transform (LLT) based cross-mode transfer module (LCTM) for mode-varying and pixel shift compensation of cross-mode flow, and uni-mode photometric correct module (UPCM) for estimating the depth-wise motion caused photometric residual of uni-mode flow. The iToF flow-based depth extraction network is proposed which could facilitate the estimation of the 4-phase measurements at each individual time for high framerate and accurate depth estimation. Extensive experiments, including both simulation and real-world experiments, are conducted to demonstrate the effectiveness of the proposed methods. Compared with the SOTA method, our approach reduces the computation time by 75% while improving the performance by 38%. The code and database are available at https://github.com/ComputationalPerceptionLab/iToF_flow.

In this paper, we propose the first generalizable view synthesis approach that specifically targets multi-view stereo-camera images. Since recent stereo matching has demonstrated accurate geometry prediction, we introduce stereo matching into novel-view synthesis for high-quality geometry reconstruction. To this end, this paper proposes a novel framework, dubbed StereoNeRF, which integrates stereo matching into a NeRF-based generalizable view synthesis approach. StereoNeRF is equipped with three key components to effectively exploit stereo matching in novel-view synthesis: a stereo feature extractor, a depth-guided plane-sweeping, and a stereo depth loss. Moreover, we propose the StereoNVS dataset, the first multi-view dataset of stereo-camera images, encompassing a wide variety of both real and synthetic scenes. Our experimental results demonstrate that StereoNeRF surpasses previous approaches in generalizable view synthesis.

While image diffusion models have made significant progress in text-driven 3D content creation, they often fail to accurately capture the intended meaning of text prompts, especially for view information. This limitation leads to the Janus problem, where multi-faced 3D models are generated under the guidance of such diffusion models. In this paper, we propose a robust high-quality 3D content generation pipeline by exploiting orthogonal-view image guidance. First, we introduce a novel 2D diffusion model that generates an image consisting of four orthogonal-view sub-images based on the given text prompt. Then, the 3D content is created using this diffusion model. Notably, the generated orthogonal-view image provides strong geometric structure priors and thus improves 3D consistency. As a result, it effectively resolves the Janus problem and significantly enhances the quality of 3D content creation. Additionally, we present a 3D synthesis fusion network that can further improve the details of the generated 3D contents. Both quantitative and qualitative evaluations demonstrate that our method surpasses previous text-to-3D techniques. Project page: https://efficientdreamer.github.io.

In Structure-from-Motion (SfM), the underlying viewgraphs of unordered image collections generally have a highly redundant set of edges that can be sparsified for efficiency without significant loss of reconstruction quality. Often, there are also false edges due to incorrect image retrieval and repeated structures (symmetries) that give rise to ghosting and superimposed reconstruction artifacts. We present a unified method to simultaneously sparsify the viewgraph and remove false edges. We propose a scoring mechanism based on camera triplets that identifies edge redundancy as well as false edges. Our edge selection is formulated as an optimization problem which can be provably solved using a simple thresholding scheme. This results in a highly efficient algorithm which can be incorporated as a pre-processing step into any SfM pipeline, making it practically usable. We demonstrate the utility of our method on generic and ambiguous datasets that cover the range of small, medium and large-scale datasets, all with different statistical properties. Sparsification of generic datasets using our method significantly reduces reconstruction time while maintaining the accuracy of the reconstructions as well as removing ghosting artifacts. For ambiguous datasets, our method removes false edges, thereby avoiding incorrect superimposed reconstructions.

Due to the omnipresence of Neural Radiance Fields (NeRFs), the interest towards editable implicit 3D representations has surged over the last years. However, editing implicit or hybrid representations as used for NeRFs is difficult due to the entanglement of appearance and geometry encoded in the model parameters. Despite these challenges, recent research has shown first promising steps towards photorealistic and non-photorealistic appearance edits. The main open issues of related work include limited interactivity, a lack of support for local edits and large memory requirements, rendering them less useful in practice. We address these limitations with LAENeRF, a unified framework for photorealistic and non-photorealistic appearance editing of NeRFs. To tackle local editing, we leverage a voxel grid as starting point for region selection. We learn a mapping from expected ray terminations to final output color, which can optionally be supervised by a style loss, resulting in a framework which can perform photorealistic and non-photorealistic appearance editing of selected regions. Relying on a single point per ray for our mapping, we limit memory requirements and enable fast optimization. To guarantee interactivity, we compose the output color using a set of learned, modifiable base colors, composed with additive layer mixing. Compared to concurrent work, LAENeRF enables recoloring and stylization while keeping processing time low. Furthermore, we demonstrate that our approach surpasses baseline methods both quantitatively and qualitatively.

Joint camera pose and dense geometry estimation from a set of images or a monocular video remains a challenging problem due to its computational complexity and inherent visual ambiguities. Most dense incremental reconstruction systems operate directly on image pixels and solve for their 3D positions using multi-view geometry cues. Such pixel-level approaches suffer from ambiguities or violations of multi-view consistency (e.g. caused by textureless or specular surfaces). We address this issue with a new image representation which we call a SuperPrimitive. SuperPrimitives are obtained by splitting images into semantically correlated local regions and enhancing them with estimated surface normal directions, both of which are predicted by state-of-the-art single image neural networks. This provides a local geometry estimate per SuperPrimitive, while their relative positions are adjusted based on multi-view observations.We demonstrate the versatility of our new representation by addressing three 3D reconstruction tasks: depth completion, few-view structure from motion, and monocular dense visual odometry. Project page: https://makezur.github.io/SuperPrimitive/

Many problems in computer vision can be formulated as geometric estimation problems, i.e. given a collection of measurements (e.g. point correspondences) we wish to fit a model (e.g. an essential matrix) that agrees with our observations.This necessitates some measure of how much an observation "agrees" with a given model. A natural choice is to consider the smallest perturbation that makes the observation exactly satisfy the constraints. However, for many problems, this metric is expensive or otherwise intractable to compute. The so-called Sampson error approximates this geometric error through a linearization scheme. For epipolar geometry, the Sampson error is a popular choice and in practice known to yield very tight approximations of the corresponding geometric residual (the reprojection error).In this paper we revisit the Sampson approximation and provide new theoretical insights as to why and when this approximation works, as well as provide explicit bounds on the tightness under some mild assumptions. Our theoretical results are validated in several experiments on real data and in the context of different geometric estimation tasks.

3D object generation has undergone significant advancements, yielding high-quality results. However, fall short in achieving precise user control, often yielding results that do not align with user expectations, thus limiting their applicability. User-envisioning 3D object generation faces significant challenges in realizing its concepts using current generative models due to limited interaction capabilities. Existing methods mainly offer two approaches: (i) interpreting textual instructions with constrained controllability, or (ii) reconstructing 3D objects from 2D images. Both of them limit customization to the confines of the 2D reference, and potentially introduce undesirable artifacts during the 3D lifting process, restricting the scope for direct and versatile 3D modifications. In this work, we introduce Interactive3D, an innovative framework for interactive 3D generation that grants users precise control over the generative process through extensive 3D interaction capabilities. Interactive3D is constructed in two cascading stages, utilizing distinct 3D representations. The first stage employs Gaussian Splatting for direct user interaction, allowing modifications and guidance of the generative direction at any intermediate step through (i) Adding and Removing components, (ii) Deformable and Rigid Dragging, (iii) Geometric Transformations, and (iv) Semantic Editing. Subsequently, the Gaussian splats are transformed into InstantNGP. We introduce a novel (v) Interactive Hash Refinement module to further add details and extract the geometry in the second stage. Our experiments demonstrate that proposed Interactive3D markedly improves the controllability and quality of 3D generation. Our project webpage is available at https://interactive-3d.github.io/.

Neural Radiance Fields (NeRF) have been successfully applied in various aerial scenes, yet they face challenges with sparse views due to limited supervision. The acquisition of dense aerial views is often prohibitive, as unmanned aerial vehicles (UAVs) may encounter constraints in perspective range and energy constraints. In this work, we introduce Multiplane Prior guided NeRF (MPNeRF), a novel approach tailored for few-shot aerial scene rendering—marking a pioneering effort in this domain. Our key insight is that the intrinsic geometric regularities specific to aerial imagery could be leveraged to enhance NeRF in sparse aerial scenes. By investigating NeRF's and Multiplane Image (MPI)'s behavior, we propose to guide the training process of NeRF with a Multiplane Prior. The proposed Multiplane Prior draws upon MPI's benefits and incorporates advanced image comprehension through a SwinV2 Transformer, pre-trained via SimMIM. Our extensive experiments demonstrate that MPNeRF outperforms existing state-of-the-art methods applied in non-aerial contexts, by tripling the performance in SSIM and LPIPS even with three views available. We hope our work offers insights into the development of NeRF-based applications in aerial scenes with limited data.

We introduce the first approach that creates animatable human avatars from monocular videos using 3D Gaussian Splatting (3DGS). Existing methods based on neural radiance fields (NeRFs) achieve high-quality novel-view/novel-pose image synthesis but often require days of training, and are extremely slow at inference time.Recently, the community has explored fast grid structures for efficient training of clothed avatars. Albeit being extremely fast at training, these methods can barely achieve interactive rendering frame rate (about 15 FPS). In this paper, we use 3D Gaussian Splatting and learn a non-rigid deformation network to reconstruct animatable clothed human avatars that can be trained within 30 minutes and rendered at real-time frame rates (50+ FPS).Given the explicit nature of our representation, we further introduce as-isometric-as-possible regularizations on both the Gaussian mean vectors and the covariance matrices, enhancing the generalization of our model on highly articulated unseen poses. Experimental results show that our method achieves comparable and even better performance compared to state-of-the-art approaches on animatable avatar creation from a monocular input, while being 400x and 250x faster in training and inference respectively.

Addressing the intricate challenge of modeling and re-rendering dynamic scenes, most recent approaches have sought to simplify these complexities using plane-based explicit representations, overcoming the slow training time issues associated with methods like Neural Radiance Fields (NeRF) and implicit representations. However, the straightforward decomposition of 4D dynamic scenes into multiple 2D plane-based representations proves insufficient for re-rendering high-fidelity scenes with complex motions. In response, we present a novel direction-aware representation (DaRe) approach that captures scene dynamics from six different directions. This learned representation undergoes an inverse dual-tree complex wavelet transformation (DTCWT) to recover plane-based information. DaReNeRF computes features for each space-time point by fusing vectors from these recovered planes. Combining DaReNeRF with a tiny MLP for color regression and leveraging volume rendering in training yield state-of-the-art performance in novel view synthesis for complex dynamic scenes. Notably, to address redundancy introduced by the six real and six imaginary direction-aware wavelet coefficients, we introduce a trainable masking approach, mitigating storage issues without significant performance decline. Moreover, DaReNeRF maintains a 2× reduction in training time compared to prior art while delivering superior performance.

3D asset generation is getting massive amounts of attention inspired by the recent success on text-guided 2D content creation. Existing text-to-3D methods use pretrained text-to-image diffusion models in an optimization problem or fine-tune them on synthetic data, which often results in non-photorealistic 3D objects without backgrounds. In this paper, we present a method that leverages pretrained text-to-image models as a prior, and learn to generate multi-view images in a single denoising process from real-world data. Concretely, we propose to integrate 3D volume-rendering and cross-frame-attention layers into each block of the existing U-Net network of the text-to-image model. Moreover, we design an autoregressive generation that renders more 3D-consistent images at any viewpoint. We train our model on real-world datasets of objects and showcase its capabilities to generate instances with a variety of high-quality shapes and textures in authentic surroundings. Compared to the existing methods, the results generated by our method are consistent, and have favorable visual quality (-30% FID, -37% KID).

Advancements in neural signed distance fields (SDFs) have enabled modeling 3D surface geometry from a set of 2D images of real-world scenes. Baking neural SDFs can extract explicit mesh with appearance baked into texture maps as neural features. The baked meshes still have a large memory footprint and require a powerful GPU for real-time rendering. Neural optimization of such large meshes with differentiable rendering pose significant challenges. We propose a method to produce optimized meshes for large unbounded scenes with low triangle budget and high fidelity of geometry and appearance. We achieve this by combining advancements in baking neural SDFs with classical mesh simplification techniques and proposing a joint appearance-geometry refinement step. The visual quality is comparable to or better than state-of-the-art neural meshing and baking methods with high geometric accuracy despite significant reduction in triangle count, making the produced meshes efficient for storage, transmission, and rendering on mobile hardware. We validate the effectiveness of the proposed method on large unbounded scenes from mip-NeRF 360, Tanks & Temples, and Deep Blending datasets, achieving at-par rendering quality with 73× reduced triangles and 11× reduction in memory footprint.

We introduce a new family of minimal problems for reconstruction from multiple views. Our primary focus is a novel approach to autocalibration, a long-standing problem in computer vision. Traditional approaches to this problem, such as those based on Kruppa's equations or the modulus constraint, rely explicitly on the knowledge of multiple fundamental matrices or a projective reconstruction. In contrast, we consider a novel formulation involving constraints on image points, the unknown depths of 3D points, and a partially specified calibration matrix $K$. For $2$ and $3$ views, we present a comprehensive taxonomy of minimal autocalibration problems obtained by relaxing some of these constraints. These problems are organized into classes according to the number of views and any assumed prior knowledge of $K$. Within each class, we determine problems with the fewest---or a relatively small number of---solutions. From this zoo of problems, we devise three practical solvers. Experiments with synthetic and real data and interfacing our solvers with COLMAP demonstrate that we achieve superior accuracy compared to state-of-the-art calibration methods. The code is available at https://github.com/andreadalcin/MinimalPerspectiveAutocalibration.

Numerous prior studies predominantly emphasize constructing relation vectors for individual neighborhood points and generating dynamic kernels for each vector and embedding these into high-dimensional spaces to capture implicit local structures. However, we contend that such implicit high-dimensional structure modeling approch inadequately represents the local geometric structure of point clouds due to the absence of explicit structural information. Hence, we introduce X-3D, an explicit 3D structure modeling approach. X-3D functions by capturing the explicit local structural information within the input 3D space and employing it to produce dynamic kernels with shared weights for all neighborhood points within the current local region. This modeling approach introduces effective geometric prior and significantly diminishes the disparity between the local structure of the embedding space and the original input point cloud, thereby improving the extraction of local features. Experiments show that our method can be used on a variety of methods and achieves state-of-the-art performance on segmentation, classification, detection tasks with lower extra computational cost, such as \textbf{90.7\%} on ScanObjectNN for classification, \textbf{79.2\%} on S3DIS 6 fold and \textbf{74.3\%} on S3DIS Area 5 for segmentation, \textbf{76.3\%} on ScanNetV2 for segmentation and \textbf{64.5\%} mAP$_{25}$, \textbf{46.9\%} mAP$_{50}$ on SUN RGB-D and \textbf{69.0\%} mAP$_{25}$, \textbf{51.1\%} mAP$_{50}$ on ScanNetV2 . Our code is available at \href{https://github.com/sunshuofeng/X-3D}{https://github.com/sunshuofeng/X-3D}.

Recently, building on the foundation of neural radiance field, various techniques have emerged to learn unsigned distance fields (UDF) to reconstruct 3D non-watertight models from multi-view images. Yet, a central challenge in UDF-based volume rendering is formulating a proper way to convert unsigned distance values into volume density, ensuring that the resulting weight function remains unbiased and sensitive to occlusions. Falling short on these requirements often results in incorrect topology or large reconstruction errors in resulting models. This paper addresses this challenge by presenting a novel two-stage algorithm, 2S-UDF, for learning a high-quality UDF from multi-view images. Initially, the method applies an easily trainable density function that, while slightly biased and transparent, aids in coarse reconstruction. The subsequent stage then refines the geometry and appearance of the object to achieve a high-quality reconstruction by directly adjusting the weight function used in volume rendering to ensure that it is unbiased and occlusion-aware. Decoupling density and weight in two stages makes our training stable and robust, distinguishing our technique from existing UDF learning approaches. Evaluations on the DeepFashion3D, DTU, and BlendedMVS datasets validate the robustness and effectiveness of our proposed approach. In both quantitative metrics and visual quality, the results indicate our superior performance over other UDF learning techniques in reconstructing 3D non-watertight models from multi-view images. Our code is available at https://bitbucket.org/jkdeng/2sudf/.

Generalizable neural implicit surface reconstruction aims to obtain an accurate underlying geometry given a limited number of multi-view images from unseen scenes. However, existing methods select only informative and relevant views using predefined scores for training and testing phases. This constraint renders the model impractical in real-world scenarios, where the availability of favorable combinations cannot always be ensured. We introduce and validate a view-combination score to indicate the effectiveness of the input view combination. We observe that previous methods output degenerate solutions under arbitrary and unfavorable sets. Building upon this finding, we propose UFORecon, a robust view-combination generalizable surface reconstruction framework. To achieve this, we apply cross-view matching transformers to model interactions between source images and build correlation frustums to capture global correlations. Additionally, we explicitly encode pairwise feature similarities as view-consistent priors. Our proposed framework significantly outperforms previous methods in terms of view-combination generalizability and also in the conventional generalizable protocol trained with favorable view-combinations. The code is available at https://github.com/Youngju-Na/UFORecon.

We present GenN2N, a unified NeRF-to-NeRF translation framework for various NeRF translation tasks such as text-driven NeRF editing, colorization, super-resolution, inpainting, etc. Unlike previous methods designed for individual translation tasks with task-specific schemes, GenN2N achieves all these NeRF editing tasks by employing a plug-and-play image-to-image translator to perform editing in the 2D domain and lifting 2D edits into the 3D NeRF space. Since the 3D consistency of 2D edits may not be assured, we propose to model the distribution of the underlying 3D edits through a generative model that can cover all possible edited NeRFs. To model the distribution of 3D edited NeRFs from 2D edited images, we carefully design a VAE-GAN that encodes images while decoding NeRFs. The latent space is trained to align with a Gaussian distribution and the NeRFs are supervised through an adversarial loss on its renderings. To ensure the latent code does not depend on 2D viewpoints but truly reflects the 3D edits, we also regularize the latent code through a contrastive learning scheme. Extensive experiments on various editing tasks show GenN2N, as a universal framework, performs as well or better than task-specific specialists while possessing flexible generative power.

Most 3D generation research focuses on up-projecting 2D foundation models into the 3D space, either by minimizing 2D Score Distillation Sampling (SDS) loss or fine-tuning on multi-view datasets. Without explicit 3D priors, these methods often lead to geometric anomalies and multi-view inconsistency. Recently, researchers have attempted to improve the genuineness of 3D objects by directly training on 3D datasets, albeit at the cost of low-quality texture generation due to the limited texture diversity in 3D datasets. To harness the advantages of both approaches, we propose Bidirectional Diffusion (BiDiff), a unified framework that incorporates both a 3D and a 2D diffusion process, to preserve both 3D fidelity and 2D texture richness, respectively. Moreover, as a simple combination may yield inconsistent generation results, we further bridge them with novel bidirectional guidance. In addition, our method can be used as an initialization of optimization-based models to further improve the quality of 3D model and efficiency of optimization, reducing the process from 3.4 hours to 20 minutes. Experimental results have shown that our model achieves high-quality, diverse, and scalable 3D generation. Project website https://bidiff.github.io/.

Two-view homography estimation is a classic and fundamental problem in computer vision.While conceptually simple, the problem quickly becomes challenging when multiple planes are visible in the image pair.Even with correct matches, each individual plane (homography) might have a very low number of inliers when comparing to the set of all correspondences.In practice, this requires a large number of RANSAC iterations to generate a good model hypothesis.The current state-of-the-art methods therefore seek to reduce the sample size, from four point correspondences originally, by including additional information such as keypoint orientation/angles or local affine information.In this work, we continue in this direction and propose a novel one-point solver that leverages different approximate constraints derived from the same auxiliary information.In experiments we obtain state-of-the-art results, with execution time speed-ups, on large benchmark datasets and show that it is more beneficial for the solver to be sample efficient compared to generating more accurate homographies.

Despite the great success of deep learning in stereo matching, recovering accurate disparity maps is still challenging. Currently, L1 and cross-entropy are the two most widely used losses for stereo network training. Compared with the former, the latter usually performs better thanks to its probability modeling and direct supervision to the cost volume. However, how to accurately model the stereo ground-truth for cross-entropy loss remains largely under-explored. Existing works simply assume that the ground-truth distributions are uni-modal, which ignores the fact that most of the edge pixels can be multi-modal. In this paper, a novel adaptive multi-modal cross-entropy loss (ADL) is proposed to guide the networks to learn different distribution patterns for each pixel. Moreover, we optimize the disparity estimator to further alleviate the bleeding or misalignment artifacts in inference. Extensive experimental results on public datasets show that our method is general and can help classic stereo networks regain state-of-the-art performance. In particular, GANet with our method ranks $1^{st}$ on both the KITTI 2015 and 2012 benchmarks among the published methods. Meanwhile, excellent synthetic-to-realistic generalization performance can be achieved by simply replacing the traditional loss with ours.

Although neural radiance fields (NeRFs) have achieved triumphs in image novel view synthesis (NVS), LiDAR NVS remains largely unexplored. Previous LiDAR NVS methods employ a simple shift from image NVS methods while ignoring the dynamic nature and the large-scale reconstruction problem of LiDAR point clouds. In light of this, we propose LiDAR4D, a differentiable LiDAR-only framework for novel space-time LiDAR view synthesis. In consideration of the sparsity and large-scale characteristics, we design a 4D hybrid representation combined with multi-planar and grid features to achieve effective reconstruction in a coarse-to-fine manner. Furthermore, we introduce geometric constraints derived from point clouds to improve temporal consistency. For the realistic synthesis of LiDAR point clouds, we incorporate the global optimization of ray-drop probability to preserve cross-region patterns. Extensive experiments on KITTI-360 and NuScenes datasets demonstrate the superiority of our method in accomplishing geometry-aware and time-consistent dynamic reconstruction.

State-of-the-art neural implicit surface representations have achieved impressive results in indoor scene reconstruction by incorporating monocular geometric priors as additional supervision. However, we have observed that multi-view inconsistency between such priors poses a challenge for high-quality reconstructions. In response, we present NC-SDF, a neural signed distance field (SDF) 3D reconstruction framework with view-dependent normal compensation (NC). Specifically, we integrate view-dependent biases in monocular normal priors into the neural implicit representation of the scene. By adaptively learning and correcting the biases, our NC-SDF effectively mitigates the adverse impact of inconsistent supervision, enhancing both the global consistency and local details in the reconstructions. To further refine the details, we introduce an informative pixel sampling strategy to pay more attention to intricate geometry with higher information content. Additionally, we design a hybrid geometry modeling approach to improve the neural implicit representation. Experiments on synthetic and real-world datasets demonstrate that NC-SDF outperforms existing approaches in terms of reconstruction quality.

Existing NeRF-based methods for large scene reconstruction often have limitations in visual quality and rendering speed. While the recent 3D Gaussian Splatting works well on small-scale and object-centric scenes, scaling it up to large scenes poses challenges due to limited video memory, long optimization time, and noticeable appearance variations. To address these challenges, we present VastGaussian, the first method for high-quality reconstruction and real-time rendering on large scenes based on 3D Gaussian Splatting. We propose a progressive partitioning strategy to divide a large scene into multiple cells, where the training cameras and point cloud are properly distributed with an airspace-aware visibility criterion. These cells are merged into a complete scene after parallel optimization. We also introduce decoupled appearance modeling into the optimization process to reduce appearance variations in the rendered images. Our approach outperforms existing NeRF-based methods and achieves state-of-the-art results on multiple large scene datasets, enabling fast optimization and high-fidelity real-time rendering.

Neural radiance field (NeRF) is an emerging technique for 3D scene reconstruction and modeling. However, current NeRF-based methods are limited in the capabilities of adding or removing objects. This paper fills the aforementioned gap by proposing a new language-driven method for object manipulation in NeRFs through dataset updates. Specifically, to insert an object represented by a set of multi-view images into a background NeRF, we use a text-to-image diffusion model to blend the object into the given background across views. The generated images are then used to update the NeRF so that we can render view-consistent images of the object within the background. To ensure view consistency, we propose a dataset update strategy that prioritizes the radiance field training based on camera poses in a pose-ordered manner. We validate our method in two case studies: object insertion and object removal. Experimental results show that our method can generate photo-realistic results and achieves state-of-the-art performance in NeRF editing.

Despite the success of recent upsampling approaches, generating high-resolution point sets with uniform distribution and meticulous structures is still challenging. Unlike existing methods that only take spatial information of the raw data into account, we regard point cloud upsampling as generating dense point clouds from deformable topology. Motivated by this, we present SPU-PMD, a self-supervised topological mesh deformation network, for 3D densification. As a cascaded framework, our architecture is formulated by a series of coarse mesh interpolator and mesh deformers. At each stage, the mesh interpolator first produces the initial dense point clouds via mesh interpolation, which allows the model to perceive the primitive topology better. Meanwhile, the deformer infers the morphing by estimating the movements of mesh nodes and reconstructs the descriptive topology structure. By associating mesh deformation with feature expansion, this module progressively refines point clouds' surface uniformity and structural details. To demonstrate the effectiveness of the proposed method, extensive quantitative and qualitative experiments are conducted on synthetic and real-scanned 3D data. Also, we compare it with state-of-the-art techniques to further illustrate the superiority of our network. The project page is: https://github.com/lyz21/SPU-PMD

We present Intrinsic Image Diffusion, a generative model for appearance decomposition of indoor scenes. Given a single input view, we sample multiple possible material explanations represented as albedo, roughness, and metallic maps. Appearance decomposition poses a considerable challenge in computer vision due to the inherent ambiguity between lighting and material properties and the lack of real datasets. To address this issue, we advocate for a probabilistic formulation, where instead of attempting to directly predict the true material properties, we employ a conditional generative model to sample from the solution space. Furthermore, we show that utilizing the strong learned prior of recent diffusion models trained on large-scale real-world images can be adapted to material estimation and highly improves the generalization to real images. Our method produces significantly sharper, more consistent, and more detailed materials, outperforming state-of-the-art methods by 1.5dB on PSNR and by 45% better FID score on albedo prediction. We demonstrate the effectiveness of our approach through experiments on both synthetic and real-world datasets.

Recent works in implicit representations, such as Neural Radiance Fields (NeRF), have advanced the generation of realistic and animatable head avatars from video sequences. These implicit methods are still confronted by visual artifacts and jitters, since the lack of explicit geometric constraints poses a fundamental challenge in accurately modeling complex facial deformations. In this paper, we introduce Dynamic Tetrahedra (DynTet), a novel hybrid representation that encodes explicit dynamic meshes by neural networks to ensure geometric consistency across various motions and viewpoints. DynTet is parameterized by the coordinate-based networks which learn signed distance, deformation, and material texture, anchoring the training data into a predefined tetrahedra grid. Leveraging Marching Tetrahedra, DynTet efficiently decodes textured meshes with a consistent topology, enabling fast rendering through a differentiable rasterizer and supervision via a pixel loss. To enhance training efficiency, we incorporate classical 3D Morphable Models to facilitate geometry learning and define a canonical space for simplifying texture learning. These advantages are readily achievable owing to the effective geometric representation employed in DynTet. Compared with prior works, DynTet demonstrates significant improvements in fidelity, lip synchronization, and real-time performance according to various metrics. Beyond producing stable and visually appealing synthesis videos, our method also outputs the dynamic meshes which is promising to enable many emerging applications.

The problem of self-calibration of two cameras from a given fundamental matrix is one of the basic problems in geometric computer vision. Under the assumption of known principal points and square pixels, the Bougnoux formula offers a means to compute the two unknown focal lengths. However, in many practical situations, the formula yields inaccurate results due to commonly occurring singularities. Moreover, the estimates are sensitive to noise in the computed fundamental matrix and to the assumed positions of the principal points. In this paper, we therefore propose an efficient and robust iterative method to estimate the focal lengths along with the principal points of the cameras given a fundamental matrix and priors for the estimated camera intrinsics. In addition, we study a computationally efficient check of models generated within RANSAC that improves the accuracy of the estimated models while reducing the total computational time. Extensive experiments on real and synthetic data show that our iterative method brings significant improvements in terms of the accuracy of the estimated focal lengths over the Bougnoux formula and other state-of-the-art methods, even when relying on inaccurate priors. The code for the methods and experiments is available at https://github.com/kocurvik/robustselfcalibration

This paper introduces a versatile paradigm for integrating multi-view reflectance and normal maps acquired through photometric stereo. Our approach employs a pixel-wise joint re-parameterization of reflectance and normal, considering them as a vector of radiances rendered under simulated, varying illumination. This re-parameterization enables the seamless integration of reflectance and normal maps as input data in neural volume rendering-based 3D reconstruction while preserving a single optimization objective. In contrast, recent multi-view photometric stereo (MVPS) methods depend on multiple, potentially conflicting objectives. Despite its apparent simplicity, our proposed approach outperforms state-of-the-art approaches in MVPS benchmarks across F-score, Chamfer distance, and mean angular error metrics. Notably, it significantly improves the detailed 3D reconstruction of areas with high curvature or low visibility.

We present Neural 3D Strokes, a novel technique to generate stylized images of a 3D scene at arbitrary novel views from multi-view 2D images. Different from existing methods which apply stylization to trained neural radiance fields at the voxel level, our approach draws inspiration from image-to-painting methods, simulating the progressive painting process of human artwork with vector strokes. We develop a palette of stylized 3D strokes from basic primitives and splines, and consider the 3D scene stylization task as a multi-view reconstruction process based on these 3D stroke primitives. Instead of directly searching for the parameters of these 3D strokes, which would be too costly, we introduce a differentiable renderer that allows optimizing stroke parameters using gradient descent, and propose a training scheme to alleviate the vanishing gradient issue. The extensive evaluation demonstrates that our approach effectively synthesizes 3D scenes with significant geometric and aesthetic stylization while maintaining a consistent appearance across different views. Our method can be further integrated with style loss and image-text contrastive models to extend its applications, including color transfer and text-driven 3D scene drawing. Results and code are available at http://buaavrcg.github.io/Neural3DStrokes.

Unsupervised point cloud shape correspondence aims to establish point-wise correspondences between source and target point clouds. Existing methods obtain correspondences directly by computing point-wise feature similarity between point clouds. However, non-rigid objects possess strong deformability and unusual shapes, making it a longstanding challenge to directly establish correspondences between point clouds with unconventional shapes. To address this challenge, we propose an unsupervised Template-Assisted point cloud shape correspondence Network, termed TANet, including a template generation module and a template assistance module. The proposed TANet enjoys several merits. Firstly, the template generation module establishes a set of learnable templates with explicit structures. Secondly, we introduce a template assistance module that extensively leverages the generated templates to establish more accurate shape correspondences from multiple perspectives. Extensive experiments on four human and animal datasets demonstrate that TANet achieves favorable performance against state-of-the-art methods.

Group synchronization plays a crucial role in global pipelines for Structure from Motion (SfM). Its formulation is nonconvex and it is faced with highly corrupted measurements. Cycle consistency has been effective in addressing these challenges. However, computationally efficient solutions are needed for cycles longer than three, especially in practical scenarios where 3-cycles are unavailable. To overcome this computational bottleneck, we propose an algorithm for group synchronization that leverages information from cycles of lengths ranging from three to six with a complexity of $O(n^3)$ (or $O(n^{2.373})$ when using a faster matrix multiplication algorithm). We establish non-trivial theory for this and related methods that achieves competitive sample complexity, assuming the uniform corruption model. To advocate the practical need for our method, we consider distributed group synchronization, which requires at least 4-cycles, and we illustrate state-of-the-art performance by our method in this context.

Extracting planes from a 3D scene is useful for downstream tasks in robotics and augmented reality. In this paper we tackle the problem of estimating the planar surfaces in a scene from posed images. Our first finding is that a surprisingly competitive baseline results from combining popular clustering algorithms with recent improvements in 3D geometry estimation. However, such purely geometric methods are understandably oblivious to plane semantics, which are crucial to discerning distinct planes. To overcome this limitation, we propose a method that predicts multi-view consistent plane embeddings that complement geometry when clustering points into planes. We show through extensive evaluation on the ScanNetV2 dataset that our new method outperforms existing approaches and our strong geometric baseline for the task of plane estimation.

Automated driving fundamentally requires knowledge about the surrounding geometry of the scene.Modern approaches use only captured images to predict occupancy maps that represent the geometry.Training these approaches requires accurate data that may be acquired with the help of LiDAR scanners.We show that the techniques used for current benchmarks and training datasets to convert LiDAR scans into occupancy grid maps yield very low quality,and subsequently present a novel approach using evidence theory that yields more accurate reconstructions.We demonstrate that these are superior by a large margin, both qualitatively and quantitatively, and that we additionally obtain meaningful uncertainty estimates. When converting the occupancy maps back to depth estimates and comparing them with the original LiDAR measurements from the nuScenes dataset, our method yields an MAE improvement over 55 cm (30\%) over the baseline Occ3D and 98 cm (52%) over the baseline OpenOccupancy.Finally, we use the improved occupancy maps to train a state-of-the-art occupancy prediction method and demonstrate that it improves by 47 cm (25%).

In this paper, we showcase the effectiveness of optimizing monocular camera poses as a continuous function of time. The camera poses are represented using an implicit neural function which maps the given time to the corresponding camera pose. The mapped camera poses are then used for the downstream tasks where joint camera pose optimization is also required. While doing so, the network parameters -- that implicitly represent camera poses -- are optimized. We exploit the proposed method in four diverse experimental settings, namely,(1) NeRF from noisy poses; (2) NeRF from asynchronous Events; (3) Visual Simultaneous Localization and Mapping (vSLAM); and (4) vSLAM with IMUs. In all four settings, the proposed method performs significantly better than the compared baselines and the state-of-the-art methods. Additionally, with the assumption of continuous motion, changes in pose may actually live in a manifold that has lower than 6 degrees of freedom (DOF). We call this low DOF motion representation as the \emph{intrinsic motion} and use the approach in vSLAM settings, showing impressive camera tracking performance.

We present a method for reconstructing 3D shape of arbitrary Lambertian objects based on measurements by miniature, energy-efficient, low-cost single-photon cameras. These cameras, operating as time resolved image sensors, illuminate the scene with a very fast pulse of diffuse light and record the shape of that pulse as it returns back from the scene at a high temporal resolution. We propose to model this image formation process, account for its non-idealities, and adapt neural rendering to reconstruct 3D geometry from a set of spatially distributed sensors with known poses. We show that our approach can successfully recover complex 3D shapes from simulated data. We further demonstrate 3D object reconstruction from real-world captures, utilizing measurements from a commodity proximity sensor. Our work draws an interesting connection between image-based modeling and active range scanning and is a step towards 3D vision with single-photon cameras.

Inliers estimation constitutes a pivotal step in partially overlapping point cloud registration. Existing methods broadly obey coordinate-based scheme, where inlier confidence is scored through simply capturing coordinate differences in the context. However, this scheme results in massive inlier misinterpretation readily, consequently affecting the registration performance. In this paper, we explore to extend a new definition called inlier confidence calibration (ICC) to alleviate the above issues. Firstly, we provide finely initial correspondences for ICC in order to generate high quality reference point cloud copy corresponding to the source point cloud. In particular, we develop a soft assignment matrix optimization theorem that offers faster speed and greater precision compared to Sinkhorn. Benefiting from the high quality reference copy, we argue the neighborhood patch formed by inlier and its neighborhood should have consistency between source point cloud and its reference copy. Based on this insight, we construct transformation-invariant geometric constraints and capture geometric structure consistency to calibrate inlier confidence for estimated correspondences between source point cloud and its reference copy. Finally, transformation is further calculated by the weighted SVD algorithm with the calibrated inlier confidence. Our model is trained in an unsupervised manner, and extensive experiments on synthetic and real-world datasets illustrate the effectiveness of the proposed method.

The advent of neural 3D Gaussians has recently brought about a revolution in the field of neural rendering, facilitating the generation of high-quality renderings at real-time speeds. However, the explicit and discrete representation encounters challenges when applied to scenes featuring reflective surfaces. In this paper, we present GaussianShader, a novel method that applies a simplified shading function on 3D Gaussians to enhance the neural rendering in scenes with reflective surfaces while preserving the training and rendering efficiency. The main challenge in applying the shading function lies in the accurate normal estimation on discrete 3D Gaussians. Specifically, we proposed a novel normal estimation framework based on the shortest axis directions of 3D Gaussians with a delicately designed loss to make the consistency between the normals and the geometries of Gaussian spheres. Experiments show that GaussianShader strikes a commendable balance between efficiency and visual quality. Our method surpasses Gaussian Splatting in PSNR on specular object datasets, exhibiting an improvement of 1.57dB. When compared to prior works handling reflective surfaces, such as Ref-NeRF, our optimization time is significantly accelerated (23h vs. 0.58h).

Open-vocabulary querying in 3D space is challenging but essential for scene understanding tasks such as object localization and segmentation. Language-embedded scene representations have made progress by incorporating language features into 3D spaces. However, their efficacy heavily depends on neural networks that are resource-intensive in training and rendering. Although recent 3D Gaussians offer efficient and high-quality novel view synthesis, directly embedding language features in them leads to prohibitive memory usage and decreased performance. In this work, we introduce Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary query tasks. Instead of embedding high-dimensional raw semantic features on 3D Gaussians, we propose a dedicated quantization scheme that drastically alleviates the memory requirement, and a novel embedding procedure that achieves smoother yet high accuracy query, countering the multi-view feature inconsistencies and the high-frequency inductive bias in point-based representations. Our comprehensive experiments show that our representation achieves the best visual quality and language querying accuracy across current language-embedded representations, while maintaining real-time rendering frame rates on a single desktop GPU.

Despite the emergence of successful NeRF inpainting methods built upon explicit RGB and depth 2D inpainting supervisions, these methods are inherently constrained by the capabilities of their underlying 2D inpainters. This is due to two key reasons: (i) independently inpainting constituent images results in view-inconsistent imagery, and (ii) 2D inpainters struggle to ensure high-quality geometry completion and alignment with inpainted RGB images.To overcome these limitations, we propose a novel approach called MVIP-NeRF that harnesses the potential of diffusion priors for NeRF inpainting, addressing both appearance and geometry aspects.MVIP-NeRF performs joint inpainting across multiple views to reach a consistent solution, which is achieved via an iterative optimization process based on Score Distillation Sampling (SDS).Apart from recovering the rendered RGB images, we also extract normal maps as a geometric representation and define a normal SDS loss that motivates accurate geometry inpainting and alignment with the appearance.Additionally, we formulate a multi-view SDS score function to distill generative priors simultaneously from different view images, ensuring consistent visual completion when dealing with large view variations.Our experimental results demonstrate the superiority of our approach over previous methods for NeRF inpainting, offering superior appearance and geometry recovery.

We propose a method to allow precise and extremely fast mesh extraction from 3D Gaussian Splatting. Gaussian Splatting has recently become very popular as it yields realistic rendering while being significantly faster to train than NeRFs. It is however challenging to extract a mesh from the millions of tiny 3D Gaussians as these Gaussians tend to be unorganized after optimization and no method has been proposed so far. Our first key contribution is a regularization term that encourages the Gaussians to align well with the surface of the scene. We then introduce a method that exploits this alignment to extract a mesh from the Gaussians using Poisson reconstruction, which is fast, scalable, and preserves details, in contrast to the Marching Cubes algorithm usually applied to extract meshes from Neural SDFs. Finally, we introduce an optional refinement strategy that binds Gaussians to the surface of the mesh, and jointly optimizes these Gaussians and the mesh through Gaussian splatting rendering. This enables easy editing, sculpting, animating, and relighting of the Gaussians by manipulating the mesh instead of the Gaussians themselves. Retrieving such an editable mesh for realistic rendering is done within minutes with our method, compared to hours with the state-of-the-art method on SDFs, while providing a better rendering quality.

3D generation has raised great attention in recent years. With the success of text-to-image diffusion models, the 2D-lifting technique becomes a promising route to controllable 3D generation. However, these methods tend to present inconsistent geometry, which is also known as the Janus problem. We observe that the problem is caused mainly by two aspects, i.e., viewpoint bias in 2D diffusion models and overfitting of the optimization objective. To address it, we propose a two-stage 2D-lifting framework, namely DreamControl, which optimizes coarse NeRF scenes as 3D self-prior and then generates fine-grained objects with control-based score distillation. Specifically, adaptive viewpoint sampling and boundary integrity metric are proposed to ensure the consistency of generated priors. The priors are then regarded as input conditions to maintain reasonable geometries, in which conditional LoRA and weighted score are further proposed to optimize detailed textures. DreamControl can generate high-quality 3D content in terms of both geometry consistency and texture fidelity. Moreover, our control-based optimization guidance is applicable to more downstream tasks, including user-guided generation and 3D animation.

Data-driven three-dimensional parametric shape models of the human body have gained enormous popularity both for the analysis of visual data and for the generation of synthetic humans. Following a similar approach for animals does not scale to the multitude of existing animal species, not to mention the difficulty of accessing subjects to scan in 3D. However, we argue that for domestic species of great importance, like the horse, it is a highly valuable investment to put effort into gathering a large dataset of real 3D scans, and learn a realistic 3D articulated shape model. We introduce VAREN, a novel 3D articulated parametric shape model learned from 3D scans of many real horses. VAREN bridges synthesis and analysis tasks, as the generated model instances have unprecedented realism, while being able to represent horses of different sizes and shapes. Differently from previous body models, VAREN has two resolutions, an anatomical skeleton, and interpretable, learned pose-dependent deformations, which are related to the body muscles. We show with experiments that this formulation has superior performance with respect to previous strategies for modeling pose-dependent deformations in the human body case, while also being more compact and allowing an analysis of the relationship between articulation and muscle deformation during articulated motion. The VAREN model and data are available at https://varen.is.tue.mpg.de.

In this paper, we address the challenge of reconstructing general articulated 3D objects from a single video. Existing works employing dynamic neural radiance fields have advanced the modeling of articulated objects like humans and animals from videos, but face challenges with piece-wise rigid general articulated objects due to limitations in their deformation models. To tackle this, we propose Quasi-Rigid Blend Skinning, a novel deformation model that enhances the rigidity of each part while maintaining flexible deformation of the joints. Our primary insight combines three distinct approaches: 1) an enhanced bone rigging system for improved component modeling, 2) the use of quasi-sparse skinning weights to boost part rigidity and reconstruction fidelity, and 3) the application of geodesic point assignment for precise motion and seamless deformation. Our method outperforms previous works in producing higher-fidelity 3D reconstructions of general articulated objects, as demonstrated on both real and synthetic datasets.

We propose a novel concept of dual and integrated latent topologies (DITTO in short) for implicit 3D reconstruction from noisy and sparse point clouds. Most existing methods predominantly focus on single latent type, such as point or grid latents. In contrast, the proposed DITTO leverages both point and grid latents (i.e., dual latent) to enhance their strengths, the stability of grid latents and the detail-rich capability of point latents. Concretely, DITTO consists of dual latent encoder and integrated implicit decoder. In the dual latent encoder, a dual latent layer, which is the key module block composing the encoder, refines both latents in parallel, maintaining their distinct shapes and enabling recursive interaction. Notably, a newly proposed dynamic sparse point transformer within the dual latent layer effectively refines point latents. Then, the integrated implicit decoder systematically combines these refined latents, achieving high-fidelity 3D reconstruction and surpassing previous state-of-the-art methods on object- and scene-level datasets, especially in thin and detailed structures.

Neural Radiance Fields (NeRF) exhibit remarkable performance for Novel View Synthesis (NVS) given a set of 2D images. However, NeRF training requires accurate camera pose for each input view, typically obtained by Structure-from-Motion (SfM) pipelines. Recent works have attempted to relax this constraint, but they still often rely on decent initial poses which they can refine. Here we aim at removing the requirement for pose initialization. We present Incremental CONfidence (ICON), an optimization procedure for training NeRFs from 2D video frames. ICON only assumes smooth camera motion to estimate initial guess for poses. Further, ICON introduces "confidence": an adaptive measure of model quality used to dynamically reweight gradients. ICON relies on high-confidence poses to learn NeRF, and high-confidence 3D structure (as encoded by NeRF) to learn poses. We show that ICON, without prior pose initialization, achieves superior performance in both CO3D and HO3D versus methods which use SfM pose.

Rotation invariance is an important requirement for point shape analysis. To achieve this, current state-of-the-art methods attempt to construct the local rotation-invariant representation through learning or defining the local reference frame (LRF). Although efficient, these LRF-based methods suffer from perturbation of local geometric relations, resulting in suboptimal local rotation invariance. To alleviate this issue, we propose a Local-consistent Transformation (LocoTrans) learning strategy. Specifically, we first construct the local-consistent reference frame (LCRF) by considering the symmetry of the two axes in LRF. In comparison with previous LRFs, our LCRF is able to preserve local geometric relationships better through performing local-consistent transformation. However, as the consistency only exists in local regions, the relative pose information is still lost in the intermediate layers of the network. We mitigate such a relative pose issue by developing a relative pose recovery (RPR) module. RPR aims to restore the relative pose between adjacent transformed patches. Equipped with LCRF and RPR, our LocoTrans is capable of learning local-consistent transformation and preserving local geometry, which benefits rotation invariance learning. Competitive performance under arbitrary rotations on both shape classification and part segmentation tasks and ablations can demonstrate the effectiveness of our method. Code will be available publicly at https://github.com/wdttt/LocoTrans.

With photo-realistic image generation, Neural Radiance Field (NeRF) is widely used for large-scale dynamic scene reconstruction as autonomous driving simulator. However, large-scale scene reconstruction still suffers from extremely long training time and rendering time. Low-resolution(LR) rendering combined with upsampling can alleviate this problem but it degrades image quality. In this paper, we design a lightweight reference decoder which exploits prior information from known views to improve image reconstruction quality of new views. In addition, to speed up prior information search, we propose an optical flow and structural similarity based prior information search method. Results on KITTI and VKITTI2 datasets show that our method significantly outperforms the baseline method in terms of training speed, rendering speed and rendering quality.

Affine subspaces of Euclidean spaces are also referred to as flats. A standard task in computer vision, or more generally in engineering and applied sciences, is fitting a flat to a set of points, which is commonly solved using the PCA. We generalize this technique to enable fitting a flat to a set of other flats, possibly of varying dimensions, based on representing the flats as squared distance fields. Compared to previous approaches such as Riemannian centers of mass in the manifold of affine Grassmannians, our approach is conceptually much simpler and computationally more efficient, yet offers desirable properties such as respecting symmetries and being equivariant to rigid transformations, leading to more intuitive and useful results in practice. We demonstrate these claims in a number of synthetic experiments and a multi-view reconstruction task of line-like objects.

Recent progress in human shape learning, shows that neural implicit models are effective in generating 3D human surfaces from limited number of views, and even from a single RGB image. However, existing monocular approaches still struggle to recover fine geometric details such as face, hands or cloth wrinkles. They are also easily prone to depth ambiguities that result in distorted geometries along the camera optical axis.In this paper, we explore the benefits of incorporating depth observations in the reconstruction process by introducing ANIM, a novel method that reconstructs arbitrary 3D human shapes from single-view RGB-D images with an unprecedented level of accuracy.Our model learns geometric details from both multi-resolution pixel-aligned and voxel-aligned features to leverage depth information and enable spatial relationships, mitigating depth ambiguities.We further enhance the quality of the reconstructed shape by introducing a depth-supervision strategy, which improves the accuracy of the signed distance field estimation of points that lie on the reconstructed surface.Experiments demonstrate that ANIM outperforms state-of-the-art works that use RGB, surface normals, point cloud or RGB-D data as input. In addition, we introduce ANIM-Real, a new multi-modal dataset comprising high-quality scans paired with consumer-grade RGB-D camera, and our protocol to fine-tune ANIM, enabling high-quality reconstruction from real-world human capture.

Stereo matching is a core task for many computer vision and robotics applications. Despite their dominance in traditional stereo methods, the hand-crafted Markov Random Field (MRF) models lack sufficient modeling accuracy compared to end-to-end deep models. While deep learning representations have greatly improved the unary terms of the MRF models, the overall accuracy is still severely limited by the hand-crafted pairwise terms and message passing. To address these issues, we propose a neural MRF model, where both potential functions and message passing are designed using data-driven neural networks. Our fully data-driven model is built on the foundation of variational inference theory, to prevent convergence issues and retain stereo MRF's graph inductive bias. To make the inference tractable and scale well to high-resolution images, we also propose a Disparity Proposal Network (DPN) to adaptively prune the search space of disparity. The proposed approach ranks $1^{st}$ on both KITTI 2012 and 2015 leaderboards among all published methods while running faster than 100 ms. This approach significantly outperforms prior global methods, e.g., lowering D1 metric by more than 50% on KITTI 2015. In addition, our method exhibits strong cross-domain generalization and can recover sharp edges. The codes at https://github.com/aeolusguan/NMRF.

Geometry-agnostic system identification is a technique for identifying the geometry and physical properties of an object from video sequences without any geometric assumptions. Recently, physics-augmented continuum neural radiance fields (PAC-NeRF) has demonstrated promising results for this technique by utilizing a hybrid Eulerian–Lagrangian representation, in which the geometry is represented by the Eulerian grid representations of NeRF, the physics is described by a material point method (MPM), and they are connected via Lagrangian particles. However, a notable limitation of PAC-NeRF is that its performance is sensitive to the learning of the geometry from the first frames owing to its two-step optimization. First, the grid representations are optimized with the first frames of video sequences, and then the physical properties are optimized through video sequences utilizing the fixed first-frame grid representations. This limitation can be critical when learning of the geometric structure is difficult, for example, in a few-shot (sparse view) setting. To overcome this limitation, we propose Lagrangian particle optimization (LPO), in which the positions and features of particles are optimized through video sequences in Lagrangian space. This method allows for the optimization of the geometric structure across the entire video sequence within the physical constraints imposed by the MPM. The experimental results demonstrate that the LPO is useful for geometric correction and physical identification in sparse-view settings.

DiffusionAvatars synthesizes a high-fidelity 3D head avatar of a person, offering intuitive control over both pose and expression. We propose a diffusion-based neural renderer that leverages generic 2D priors to produce compelling images of faces. For coarse guidance of the expression and head pose, we render a neural parametric head model (NPHM) from the target viewpoint, which acts as a proxy geometry of the person. Additionally, to enhance the modeling of intricate facial expressions, we condition DiffusionAvatars directly on the expression codes obtained from NPHM via cross-attention. Finally, to synthesize consistent surface details across different viewpoints and expressions, we rig learnable spatial features to the head’s surface via TriPlane lookup in NPHM’s canonical space. We train DiffusionAvatars on RGB videos and corresponding fitted NPHM meshes of a person and test the obtained avatars in both self-reenactment and animation scenarios. Our experiments demonstrate that DiffusionAvatars generates temporally consistent and visually appealing videos for novel poses and expressions of a person, outperforming existing approaches.

Although Vision Transformer (ViT) has achieved significant success in computer vision, it does not perform well in dense prediction tasks due to the lack of inner-patch information interaction and the limited diversity of feature scale. Most existing studies are devoted to designing vision-specific transformers to solve the above problems, which introduce additional pre-training costs. Therefore, we present a plain, pre-training-free, and feature-enhanced ViT backbone with Convolutional Multi-scale feature interaction, named ViT-CoMer, which facilitates bidirectional interaction between CNN and transformer. Compared to the state-of-the-art, ViT-CoMer has the following advantages: (1) We inject spatial pyramid multi-receptive field convolutional features into the ViT architecture, which effectively alleviates the problems of limited local information interaction and single-feature representation in ViT. (2) We propose a simple and efficient CNN-Transformer bidirectional fusion interaction module that performs multi-scale fusion across hierarchical features, which is beneficial for handling dense prediction tasks. (3) We evaluate the performance of ViT-CoMer across various dense prediction tasks, different frameworks, and multiple advanced pre-training. Notably, our ViT-CoMer-L achieves 64.3% AP on COCO val2017 without extra training data, and 62.1% mIoU on ADE20K val, both of which are comparable to state-of-the-art methods. We hope ViT-CoMer can serve as a new backbone for dense prediction tasks to facilitate future research. The code will be released at https://github.com/Traffic-X/ViT-CoMer.

Predicting 3D point trajectory is a fundamental learning task which commonly should be equivariant under Euclidean transformation, e.g., SE(3). The existing equivariant models are commonly based on the group equivariant convolution, equivariant message passing, vector neuron, frame averaging, etc. In this paper, we propose a novel pose-transformed equivariant network, in which the points are firstly uniquely normalized and then transformed by the learned pose transformations, upon which the points after motion are predicted and aggregated. Under each transformed pose, we design the point position predictor consisting of multiple Pose-Transformed Points Prediction blocks, in which the global and local motions are estimated and aggregated. This framework can be proven to be equivariant to SE(3) transformation over 3D points. We evaluate the pose-transformed equivariant network on extensive datasets including human motion capture, molecular dynamics modeling and dynamics simulation. Extensive experimental comparisons demonstrated our SOTA performance compared with the existing equivariant networks for 3D point trajectory prediction.

Large-kernel convolutional neural networks (ConvNets) have recently received extensive research attention, but two unresolved and critical issues demand further investigation. 1) The architectures of existing large-kernel ConvNets largely follow the design principles of conventional ConvNets or transformers, while the architectural design for large-kernel ConvNets remains under-addressed. 2) As transformers have dominated multiple modalities, it remains to be investigated whether ConvNets also have a strong universal perception ability in domains beyond vision. In this paper, we contribute from two aspects. 1) We propose four architectural guidelines for designing large-kernel ConvNets, the core of which is to exploit the essential characteristics of large kernels that distinguish them from small kernels - they can see wide without going deep. Following such guidelines, our proposed large-kernel ConvNet shows leading performance in image recognition (ImageNet accuracy of 88.0%, ADE20K mIoU of 55.6%, and COCO box AP of 56.4%), demonstrating better performance and higher speed than the recent powerful competitors. 2) We discover large kernels are the key to unlocking the exceptional performance of ConvNets in domains where they were originally not proficient. With certain modality-related preprocessing approaches, the proposed model achieves state-of-the-art performance on time-series forecasting and audio recognition tasks even without modality-specific customization to the architecture. All the code and models are publicly available on GitHub and Huggingface.

In the field of deep point cloud understanding, KPConv is a unique architecture that uses kernel points to locate convolutional weights in space, instead of relying on multi-layer perceptron encodings. While it initially achieved success, it has since been surpassed by recent MLP networks that employ updated designs and training strategies. Building upon the kernel point principle, we present two novel designs: KPConvD (depthwise KPConv), a lighter design that enables the use of deeper architectures, and KPConvX, an innovative design that scales the depthwise convolutional weights of KPConvD with kernel attention values. Using KPConvX with a modern architecture and training strategy, we are able to outperform current state-of-the-art approaches on the ScanObjectNN, Scannetv2, and S3DIS datasets. We validate our design choices through ablation studies and will release our code and models.

As foundation models become more popular, there is a growing need to efficiently finetune them for downstream tasks. Although numerous adaptation methods have been proposed, they are designed to be efficient only in terms of how many parameters are trained. They, however, typically still require backpropagating gradients throughout the model, meaning that their training-time and -memory cost does not reduce as significantly.We propose an adaptation method which does not backpropagate gradients through the backbone. We achieve this by designing a lightweight network in parallel that operates on features from the frozen, pretrained backbone. As a result, our method is efficient not only in terms of parameters, but also in training-time and memory usage. Our approach achieves state-of-the-art accuracy-parameter trade-offs on the popular VTAB benchmark, and we further show how we outperform prior works with respect to training-time and -memory usage too. We further demonstrate the training efficiency and scalability of our method by adapting a vision transformer backbone of 4 billion parameters for the computationally demanding task of video classification, without any intricate model parallelism. Here, we outperform a prior adaptor-based method which could only scale to a 1 billion parameter backbone, or fully-finetuning a smaller backbone, with the same GPU and less training time.

Convolutional neural networks benefit from translation equivariance, achieving tremendous success. Equivariant networks further extend this property to other transformation groups. However, most existing methods require discretization or sampling of groups, leading to increased model sizes for larger groups, such as the affine group. In this paper, we build affine equivariant networks based on differential invariants from the viewpoint of symmetric PDEs, without discretizing or sampling the group. To address the division-by-zero issue arising from fractional differential invariants of the affine group, we construct a new kind of affine invariants by normalizing polynomial relative differential invariants to replace classical differential invariants. For further flexibility, we design an equivariant layer, which can be directly integrated into convolutional networks of various architectures. Moreover, our framework for the affine group is also applicable to its continuous subgroups. We implement equivariant networks for the scale group, the rotation-scale group, and the affine group. Numerical experiments demonstrate the outstanding performance of our framework across classification tasks involving transformations of these groups. Remarkably, under the out-of-distribution setting, our model achieves a 3.37% improvement in accuracy over the main counterpart affConv on the affNIST dataset.

Recently, some large kernel convnets strike back with appealing performance and efficiency. However, given the square complexity of convolution, scaling up kernels can bring about an enormous amount of parameters and the proliferated parameters can induce severe optimization problem. Due to these issues, current CNNs compromise to scale up to 51×51 in the form of stripe convolution (i.e., 51×5+5×51) and start to saturate as the kernel size continues growing. In this paper, we delve into addressing these vital issues and explore whether we can continue scaling up kernels for more performance gains. Inspired by human vision, we propose a human-like peripheral convolution that efficiently reduces over 90% parameter count of dense grid convolution through parameter sharing, and manage to scale up kernel size to extremely large. Our peripheral convolution behaves highly similar to human, reducing the complexity of convolution from O(K^2) to O(logK) without backfiring performance. Built on this, we propose Parameter-efficient Large Kernel Network (PeLK). Our PeLK outperforms modern vision Transformers and ConvNet architectures like Swin, ConvNeXt, RepLKNet and SLaK on various vision tasks including ImageNet classification, semantic segmentation on ADE20K and object detection on MS COCO. For the first time, we successfully scale up the kernel size of CNNs to an unprecedented 101×101 and demonstrate consistent improvements.

In the field of computer vision, Vision Transformers (ViTs) have emerged as a prominent deep learning architecture. Despite being inspired by Convolutional Neural Networks (CNNs), ViTs are susceptible to small spatial shifts in the input data – they lack shift-equivariance. To address this shortcoming, we introduce novel data-adaptive designs for each of the ViT modules that break shift-equivariance, such as tokenization, self-attention, patch merging, and positional encoding. With our proposed modules, we achieve perfect circular shift-equivariance across four prominent ViT architectures: Swin, SwinV2, CvT, and MViTv2. Additionally, we leverage our design to further enhance consistency under standard shifts. We evaluate our adaptive ViT models on image classification and semantic segmentation tasks. Our models achieve competitive performance across three diverse datasets, showcasing perfect (100%) circular shift consistency while improving standard shift consistency.

Recent Vision Transformer Compression (VTC) works mainly follow a two-stage scheme, where the importance score of each model unit is first evaluated or preset in each submodule, followed by the sparsity score evaluation according to the target sparsity constraint. Such a separate evaluation process induces the gap between importance and sparsity score distributions, thus causing high search costs for VTC. In this work, for the first time, we investigate how to integrate the evaluations of importance and sparsity scores into a single stage, searching the optimal subnets in an efficient manner. Specifically, we present OFB, a cost-efficient approach that simultaneously evaluates both importance and sparsity scores, termed Once for Both (OFB), for VTC. First, a bi-mask scheme is developed by entangling the importance score and the differentiable sparsity score to jointly determine the pruning potential (prunability) of each unit.Such a bi-mask search strategy is further used together with a proposed adaptive one-hot loss to realize the progressive-and-efficient search for the most important subnet. Finally, Progressive Masked Image Modeling (PMIM) is proposed to regularize the feature space to be more representative during the search process, which may be degraded by the dimension reduction. Extensive experiments demonstrate that OFB can achieve superior compression performance over state-of-the-art searching-based and pruning-based methods under various Vision Transformer architectures, meanwhile promoting search efficiency significantly, e.g., costing one GPU search day for the compression of DeiT-S on ImageNet-1K.

Quantization for model compression can efficiently reduce the network complexity and storage requirement, but the original training data is necessary to remedy the performance loss caused by quantization. The Data-Free Quantization (DFQ) methods have been proposed to handle the absence of original training data with synthetic data. However, there are differences between the synthetic and original training data, which affects the performance of the quantized network, but none of the existing methods considers the differences. In this paper, we propose an efficient data-free quantization via pseudo-label filtering, which is the first to evaluate the synthetic data before quantization. We design a new metric for evaluating synthetic data using self-entropy, which indicates the reliability of synthetic data. The synthetic data can be categorized with the metric into high- and low-reliable datasets for the following training process. Besides, the multiple pseudo-labels are designed to label the synthetic data with different reliability, which can provide valuable supervision information and avoid misleading training by low-reliable samples. Extensive experiments are implemented on several datasets, including CIFAR-10, CIFAR-100, and ImageNet with various models. The experimental results show that our method can perform excellently and outperform existing methods in accuracy.

Federated Learning (FL) enables joint training across distributed clients using their local data privately. Federated Multi-Task Learning (FMTL) builds on FL to handle multiple tasks, assuming model congruity that identical model architecture is deployed in each client. To relax this assumption and thus extend real-world applicability, we introduce a novel problem setting, Hetero-Client Federated Multi-Task Learning (HC-FMTL), to accommodate diverse task setups. The main challenge of HC-FMTL is the model incongruity issue that invalidates conventional aggregation methods. It also escalates the difficulties in accurate model aggregation to deal with data and task heterogeneity inherent in FMTL. To address these challenges, we propose the FedHCA$^2$ framework, which allows for federated training of personalized models by modeling relationships among heterogeneous clients. Drawing on our theoretical insights into the difference between multi-task and federated optimization, we propose the Hyper Conflict-Averse Aggregation scheme to mitigate conflicts during encoder updates. Additionally, inspired by task interaction in MTL, the Hyper Cross Attention Aggregation scheme uses layer-wise cross attention to enhance decoder interactions while alleviating model incongruity. Moreover, we employ learnable Hyper Aggregation Weights for each client to customize personalized parameter updates. Extensive experiments demonstrate the superior performance of FedHCA$^2$ in various HC-FMTL scenarios compared to representative methods. Our code will be made publicly available.

The remarkable success of Vision Transformers in Artificial Neural Networks (ANNs) has led to a growing interest in incorporating the self-attention mechanism and transformer-based architecture into Spiking Neural Networks (SNNs). While existing methods propose spiking self-attention mechanisms that are compatible with SNNs, they lack reasonable scaling methods, and the overall architectures proposed by these methods suffer from a bottleneck in effectively extracting local features. To address these challenges, we propose a novel spiking self-attention mechanism named Dual Spike Self-Attention (DSSA) with a reasonable scaling method. Based on DSSA, we propose a novel spiking Vision Transformer architecture called SpikingResformer, which combines the ResNet-based multi-stage architecture with our proposed DSSA to improve both performance and energy efficiency while reducing parameters. Experimental results show that SpikingResformer achieves higher accuracy with fewer parameters and lower energy consumption than other spiking Vision Transformer counterparts. Notably, our SpikingResformer-L achieves 79.40% top-1 accuracy on ImageNet with 4 time-steps, which is the state-of-the-art result in the SNN field.

In many practical applications, 3D point cloud analysis requires rotation invariance. In this paper, we present a learnable descriptor invariant under 3D rotations and reflections, i.e., the O(3) actions, utilizing the recently introduced steerable 3D spherical neurons and vector neurons. Specifically, we propose an embedding of the 3D spherical neurons into 4D vector neurons, which leverages end-to-end training of the model. In our approach, we perform TetraTransform—an equivariant embedding of the 3D input into 4D, constructed from the steerable neurons—and extract deeper O(3)-equivariant features using vector neurons. This integration of the TetraTransform into the VN-DGCNN framework, termed TetraSphere, negligibly increases the number of parameters by less than 0.0002%. TetraSphere sets a new state-of-the-art performance classifying randomly rotated real-world object scans of the challenging subsets of ScanObjectNN. Additionally, TetraSphere outperforms all equivariant methods on randomly rotated synthetic data: classifying objects from ModelNet40 and segmenting parts of the ShapeNet shapes. Thus, our results reveal the practical value of steerable 3D spherical neurons for learning in 3D Euclidean space. The code is available at https://github.com/pavlo-melnyk/tetrasphere.

Sharpness-Aware Minimization (SAM) has been instrumental in improving deep neural network training by minimizing both training loss and loss sharpness. Despite the practical success, the mechanisms behind SAM's generalization enhancements remain elusive, limiting its progress in deep learning optimization. In this work, we investigate SAM's core components for generalization improvement and introduce "Friendly-SAM" (F-SAM) to further enhance SAM's generalization. Our investigation reveals the key role of batch-specific stochastic gradient noise within the adversarial perturbation, i.e., the current minibatch gradient, which significantly influences SAM's generalization performance. By decomposing the adversarial perturbation in SAM into full gradient and stochastic gradient noise components, we discover that relying solely on the full gradient component degrades generalization while excluding it leads to improved performance. The possible reason lies in the full gradient component's increase in sharpness loss for the entire dataset, creating inconsistencies with the subsequent sharpness minimization step solely on the current minibatch data. Inspired by these insights, F-SAM aims to mitigate the negative effects of the full gradient component. It removes the full gradient estimated by an exponentially moving average (EMA) of historical stochastic gradients, and then leverages stochastic gradient noise for improved generalization. Moreover, we provide theoretical validation for the EMA approximation and prove the convergence of F-SAM on non-convex problems. Extensive experiments demonstrate the superior generalization performance and robustness of F-SAM over vanilla SAM. Code is available at https://github.com/nblt/F-SAM.

Vision Transformer (ViT) has gained increasing attention in the computer vision community in recent years. However, the core component of ViT, Self-Attention, lacks explicit spatial priors and bears a quadratic computational complexity, thereby constraining the applicability of ViT. To alleviate these issues, we draw inspiration from the recent Retentive Network (RetNet) in the field of NLP, and propose RMT, a strong vision backbone with explicit spatial prior for general purposes. Specifically, we extend the RetNet's temporal decay mechanism to the spatial domain, and propose a spatial decay matrix based on the Manhattan distance to introduce the explicit spatial prior to Self-Attention. Additionally, an attention decomposition form that adeptly adapts to explicit spatial prior is proposed, aiming to reduce the computational burden of modeling global information without disrupting the spatial decay matrix. Based on the spatial decay matrix and the attention decomposition form, we can flexibly integrate explicit spatial prior into the vision backbone with linear complexity. Extensive experiments demonstrate that RMT exhibits exceptional performance across various vision tasks. Specifically, without extra training data, RMT achieves 84.8% and 86.1% top-1 acc on ImageNet-1k with 27M/4.5GFLOPs and 96M/18.2GFLOPs. For downstream tasks, RMT achieves 54.5 box AP and 47.2 mask AP on the COCO detection task, and 52.8 mIoU on the ADE20K semantic segmentation task.

We introduce Deformable ConvNets v4 (DCNv4), a highly efficient and effective operator for a broad spectrum of vision applications featuring an advanced sparse attention mechanism. DCNv4 addresses the limitations of its predecessor, DCNv3, with two key enhancements: 1. removing softmax normalization in spatial aggregation to enhance its dynamic property and expressive power and 2. optimizing memory access to minimize redundant operations for speedup. These improvements result in a significantly faster convergence compared to DCNv3 and a substantial increase in processing speed, with DCNv4 achieving more than three times the forward speed.Our evaluation demonstrates DCNv4's superior performance in various tasks, including image classification, instance and semantic segmentation, and notably in image generation. When integrated into generative models like U-Net in the latent diffusion model, DCNv4 outperforms baselines, underscoring its potential to enhance generative models. In practical applications, replacing DCNv3 with DCNv4 in the InternImage model to create FlashInternImage results in up to an 80\% speed increase without necessitating further modifications.DCNv4's advancements in speed and efficiency, combined with its robust performance across diverse vision tasks, position it as a foundational building block for future efficient and effective vision models.

Supernet is a core component in many recent Neural Architecture Search (NAS) methods. It not only helps embody the search space but also provides a (relative) estimation of the final performance of candidate architectures. Thus, it is critical that the top architectures ranked by a supernet should be consistent with those ranked by true performance, which is known as the order-preserving ability.In this work, we analyze the order-preserving ability on the whole search space (global) and a sub-space of top architectures (local), and empirically show that the local order-preserving for current two-stage NAS methods still need to be improved.To rectify this, we propose a novel concept of $\textbf{Supernet Shifting}$, a refined search strategy combining architecture searching with supernet fine-tuning. Specifically, apart from evaluating, the training loss is also accumulated in searching and the supernet is updated every iteration. Since superior architectures are sampled more frequently in evolutionary searching, the supernet is encouraged to focus on top architectures, thus improving local order-preserving.Besides, a pre-trained supernet is often un-reusable for one-shot methods. We show that Supernet Shifting can fulfill transferring supernet to a new dataset. Specifically, the last classifier layer will be unset and trained through evolutionary searching. Comprehensive experiments show that our method has better order-preserving ability and can find a dominating architecture. Moreover, the pre-trained supernet can be easily transferred into a new dataset with no loss of performance. Source code will be made public available.

Context. Our understanding of the generalization capabilities of neural networks (NNs) is incomplete. The prevailing explanation is based on implicit biases of gradient descent (GD) but it cannot account for recent findings of the capabilities of models found by gradient-free methods nor the `simplicity bias' observed even in untrained networks. This study seeks the source of inherent properties of NNs.Findings. To characterize inductive biases provided by architectures independently from GD, we examine networks of random weights and show that they do not correspond to random functions. We characterize the functions implemented by various architectures using decompositions in Fourier and polynomial bases and compressed representations. Even simple MLPs have strong inductive biases:uniform sampling in parameter space yields a strongly biased sampling of functions in frequency, order, and compressibility. Popular components including ReLUs, residual connections, and normalizations induce a bias toward the lower end of these measures,accounting for the ``simplicity bias'' frequently attributed to (S)GD. We also show that transformer-based sequence models inherit similar properties from their building blocks.Implications. We provide a fresh explanation for the success of deep learning compatible with recent observations, complementing those based on gradient-based optimization. This also points at future avenues for controlling the solutions implemented in trained models.

Inspired by the long-range modeling ability of ViTs, large-kernel convolutions are widely studied and adopted recently to enlarge the receptive field and improve model performance, like the remarkable work ConvNeXt which employs 7x7 depthwise convolution. Although such depthwise operator only consumes a few FLOPs, it largely harms the model efficiency on powerful computing devices due to the high memory access costs. For example, ConvNeXt-T has similar FLOPs with ResNet-50 but only achieves ~60% throughputs when trained on A100 GPUs with full precision. Although reducing the kernel size of ConvNeXt can improve speed, it results in significant performance degradation, which poses a challenging problem: How to speed up large-kernel-based CNN models while preserving their performance. To tackle this issue, inspired by Inceptions, we propose to decompose large-kernel depthwise convolution into four parallel branches along channel dimension, i.e., small square kernel, two orthogonal band kernels, and an identity mapping. With this new Inception depthwise convolution, we build a series of networks, namely IncepitonNeXt, which not only enjoy high throughputs but also maintain competitive performance. For instance, InceptionNeXt-T achieves 1.6x higher training throughputs than ConvNeX-T, as well as attains 0.2% top-1 accuracy improvement on ImageNet-1K. We anticipate InceptionNeXt can serve as an economical baseline for future architecture design to reduce carbon footprint.

Given a well-behaved neural network, is possibleto identify its parent, based on which it was tuned?In this paper, we introduce a novel task known as {neural lineage} detection, aiming at discovering lineage relationships between parent and child models. Specifically, from a set of parent models, neural lineage detection predicts which parent model a child model has been fine-tuned from. We propose two approaches to address this task. (1) For practical convenience, we introduce a learning-free approach, which integrates an approximation of the finetuning process into the neural network representation similarity metrics, leading to a similarity-based lineage detection scheme. (2) For the pursuit of accuracy, we introduce a learning-based lineage detector comprising encoders and a transformer detector. Through experimentation, we have validated that our proposed learning-free and learning-based methods outperform the baseline in various learning settings and are adaptable to a variety of visual models. Moreover, they also exhibit the ability to trace cross-generational lineage, effectively identifying not only parent models but also their ancestors.

Quantized neural networks employ reduced precision representations for both weights and activations. This quantization process significantly reduces the memory requirements and computational complexity of the network. Binary Neural Networks (BNNs) are the extreme quantization case, representing values with just one bit. Since the sign function is typically used to map real values to binary values, smooth approximations are introduced to mimic the gradients during error backpropagation. Thus, the mismatch between the forward and backward models corrupts the direction of the gradient causing training inconsistency problems and performance degradation. In contrast to current BNN approaches, we propose to employ a binary periodic (BiPer) function during binarization. Specifically, we use a square wave for the forward pass to obtain the binary values and employ the trigonometric sine function with the same period of the square wave as a differentiable surrogate during the backward pass. We demonstrate that this approach can control the quantization error by using the frequency of the periodic function and improves network performance. Extensive experiments validate the effectiveness of BiPer in benchmark datasets and network architectures, with improvements of up to 1% and 0.69% with respect to state-of-the-art methods in the classification task over CIFAR-10 and ImageNet, respectively. Our code is publicly available at https://github.com/edmav4/BiPer.

Recent studies have drawn attention to the untapped potential of the "star operation" (element-wise multiplication) in network design. While intuitive explanations abound, the foundational rationale behind its application remains largely unexplored. Our study attempts to reveal the star operation's ability of mapping inputs into high-dimensional, non-linear feature spaces—akin to kernel tricks—without widening the network. We further introduce StarNet, a simple yet powerful prototype, demonstrating impressive performance and low latency under compact network structure and efficient budget. Like stars in the sky, the star operation appears unremarkable but holds a vast universe of potential. Our work encourages further exploration across tasks, with codes available for transparency and reproducibility.

Binary neural networks utilize 1-bit quantized weights and activations to reduce both the model's storage demands and computational burden. However, advanced binary architectures still incorporate millions of inefficient and nonhardware-friendly full-precision multiplication operations. A&B BNN is proposed to directly remove part of the multiplication operations in a traditional BNN and replace the rest with an equal number of bit operations, introducing the mask layer and the quantized RPReLU structure based on the normalizer-free network architecture. The mask layer can be removed during inference by leveraging the intrinsic characteristics of BNN with straightforward mathematical transformations to avoid the associated multiplication operations. The quantized RPReLU structure enables more efficient bit operations by constraining its slope to be integer powers of 2. Experimental results achieved 92.30%, 69.35%, and 66.89% on the CIFAR-10, CIFAR-100, and ImageNet datasets, respectively, which are competitive with the state-of-the-art. Ablation studies have verified the efficacy of the quantized RPReLU structure, leading to a 1.14% enhancement on the ImageNet compared to using a fixed slope RLeakyReLU. The proposed add&bit-operation-only BNN offers an innovative approach for hardware-friendly network architecture.

We investigate a fundamental aspect of machine vision: the measurement of features, by revisiting clustering, one of the most classic approaches in machine learning and data analysis. Existing visual feature extractors, including ConvNets, ViTs, and MLPs, represent an image as rectangular regions. Though prevalent, such a grid-style paradigm is built upon engineering practice and lacks explicit modeling of data distribution. In this work, we propose feature extraction with clustering (FEC), a conceptually elegant yet surprisingly ad-hoc interpretable neural clustering framework, which views feature extraction as a process of selecting representatives from data and thus automatically captures the underlying data distribution. Given an image, FEC alternates between grouping pixels into individual clusters to abstract representatives and updating the deep features of pixels with current representatives. Such an iterative working mechanism is implemented in the form of several neural layers and the final representatives can be used for downstream tasks. The cluster assignments across layers, which can be viewed and inspected by humans, make the forward process of FEC fully transparent and empower it with promising ad-hoc interpretability. Extensive experiments on various visual recognition models and tasks verify the effectiveness, generality, and interpretability of FEC. We expect this work will provoke a rethink of the current de facto grid-style paradigm.

Neural Architecture Search is a costly practice. The fact that a search space can span a vast number of design choices with each architecture evaluation taking nontrivial overhead makes it hard for an algorithm to sufficiently explore candidate networks. In this paper, we propose AutoBuild, a scheme which learns to align the latent embeddings of operations and architecture modules with the ground-truth performance of the architectures they appear in. By doing so, AutoBuild is capable of assigning interpretable importance scores to architecture modules, such as individual operation features and larger macro operation sequences such that high-performance neural networks can be constructed without any need for search. Through experiments performed on state-of-the-art image classification, segmentation, and Stable Diffusion models, we show that by mining a relatively small set of evaluated architectures, AutoBuild can learn to build high-quality architectures directly or help to reduce search space to focus on relevant areas, finding better architectures that outperform both the original labeled ones and ones found by search baselines. Code available at https://github.com/Ascend-Research/AutoBuild

A diffusion model, which is formulated to produce an image using thousands of denoising steps, usually suffers from a slow inference speed. Existing acceleration algorithms simplify the sampling by skipping most steps yet observe considerable performance degradation. By viewing the generation of diffusion models as a discretized integrating process, we argue that the quality drop is partly caused by applying an inaccurate integral direction to a timestep interval. To rectify such inaccuracy, we propose a $\textbf{timestep aligner}$ that helps find a more accurate integral direction for a particular interval at the minimum cost. Specifically, at each denoising step, we replace the original parameterization by conditioning the network on a new timestep, which is obtained by aligning the sampling distribution to the real distribution. Extensive experiments show that our plug-in design can be trained efficiently and boost the inference performance of various state-of-the-art acceleration methods, especially for the one with few denoising steps. For example, when using 10 denoising steps on the popular LSUN Bedroom dataset, we improve the FID of DDIM from 9.65 to 6.07, simply by adopting our method for a more appropriate set of timesteps. Code will be made publicly available.

Post-training Sparsity (PTS) is a recently emerged avenue that chases efficient network sparsity with limited data in need. Existing PTS methods, however, undergo significant performance degradation compared with traditional methods that retrain the sparse networks via the whole dataset, especially at high sparsity ratios. In this paper, we attempt to reconcile this disparity by transposing three cardinal factors that profoundly alter the performance of conventional sparsity into the context of PTS. Our endeavors particularly comprise (1) A base-decayed sparsity objective that promotes efficient knowledge transferring from dense network to the sparse counterpart. (2) A reducing regrowing search algorithm designed to ascertain the optimal sparsity distribution while circumventing overfitting to the small calibration set in PTS. (3) The employment of dynamic sparse training predicated on the preceding aspects, aimed at comprehensively optimizing the sparsity structure while ensuring training stability. Our proposed framework, termed UniPTS, is validated to be much superior to existing PTS methods across extensive benchmarks. As an illustration, it amplifies the performance of POT, a recently proposed recipe, from 3.9% to 68.6% when pruning ResNet50 at 90% sparsity ratio on ImageNet. We release the code of our paper at https://github.com/xjjxmu/UniPTS.

In this paper we tackle the problem of learning Structure-from-Motion (SfM) through the use of graph attention networks. SfM is a classic computer vision problem that is solved though iterative minimization of reprojection errors, referred to as Bundle Adjustment (BA), starting from a good initialization. In order to obtain a good enough initialization to BA, conventional methods rely on a sequence of sub-problems (such as pairwise pose estimation, pose averaging or triangulation) which provides an initial solution that can then be refined using BA. In this work we replace these sub-problems by learning a model that takes as input the 2D keypoints detected across multiple views, and outputs the corresponding camera poses and 3D keypoint coordinates. Our model takes advantage of graph neural networks to learn SfM-specific primitives, and we show that it can be used for fast inference of the reconstruction for new and unseen sequences. The experimental results show that the proposed model outperforms competing learning-based methods, and challenges COLMAP while having lower runtime.

Recently, efficient Vision Transformers have shown great performance with low latency on resource-constrained devices.Conventionally, they use 4$\times$4 patch embeddings and a 4-stage structure at the macro level, while utilizing sophisticated attention with multi-head configuration at the micro level.This paper aims to address computational redundancy at all design levels in a memory-efficient manner.We discover that using larger-stride patchify stem not only reduces memory access costs but also achieves competitive performance by leveraging token representations with reduced spatial redundancy from the early stages.Furthermore, our preliminary analyses suggest that attention layers in the early stages can be substituted with convolutions, and several attention heads in the latter stages are computationally redundant.To handle this, we introduce a single-head attention module that inherently prevents head redundancy and simultaneously boosts accuracy by parallelly combining global and local information.Building upon our solutions, we introduce SHViT, a Single-Head Vision Transformer that obtains the state-of-the-art speed-accuracy tradeoff.For example, on ImageNet-1k, our SHViT-S4 is 3.3$\times$, 8.1$\times$, and 2.4$\times$ faster than MobileViTv2 $\times$1.0 on GPU, CPU, and iPhone12 mobile device, respectively, while being 1.3\% more accurate.For object detection and instance segmentation on MS COCO using Mask-RCNN head, our model achieves performance comparable to FastViT-SA12 while exhibiting 3.8$\times$ and 2.0$\times$ lower backbone latency on GPU and mobile device, respectively.

Point clouds frequently contain noise and outliers, presenting obstacles for downstream applications. In this work, we introduce a novel denoising method for point clouds. By leveraging the latent space, we explicitly uncover noise components, allowing for the extraction of a clean latent code. This, in turn, facilitates the restoration of clean points via inverse transformation. A key component in our network is a new multi-level graph convolution network for capturing rich geometric structural features at various scales from local to global. These features are then integrated into the invertible neural network which bijectively maps the latent space, to guide the noise disentanglement process. Additionally, we employ an invertible monotone operator to model the transformation process, effectively enhancing the representation of integrated geometric features. This enhancement allows our network to precisely differentiate between noise factors and the intrinsic clean points in the latent code by projecting them onto separate channels. Both qualitative and quantitative evaluations demonstrate that our method outperforms state-of-the-art methods at various noise levels. The source code is available at https://github.com/yanbiao1/PD-LTS.

Gradient sparsification and quantization offer a promising prospect to alleviate the communication overhead problem in distributed learning. However, direct combination of the two results in suboptimal solutions, due to the fact that sparsification and quantization haven't been learned together. In this paper, we propose Joint Sparsification-Quantization (JointSQ) inspired by the discovery that sparsification can be treated as 0-bit quantization, regardless of architectures. Specifically, we mathematically formulate JointSQ as a mixed-precision quantization problem, expanding the solution space. It can be solved by the designed MCKP-Greedy algorithm. Theoretical analysis demonstrates the minimal compression noise of JointSQ, and extensive experiments on various network architectures, including CNN, RNN, and Transformer, also validate this point. Under the introduction of computation overhead consistent with or even lower than previous methods, JointSQ achieves a compression ratio of 1000× on different models while maintaining near-lossless accuracy and brings 1.4× to 2.9× speedup over existing methods.

Out-of-distribution (OOD) detection has attracted a large amount of attention from the machine learning research community in recent years due to its importance in deployed systems. Most of the previous studies focused on the detection of OOD samples in the multi-class classification task. However, OOD detection in the multi-label classification task, a more common real-world use case, remains an underexplored domain. In this research, we propose YolOOD - a method that utilizes concepts from the object detection domain to perform OOD detection in the multi-label classification task. Object detection models have an inherent ability to distinguish between objects of interest (in-distribution data) and irrelevant objects (OOD data) in images that contain multiple objects belonging to different class categories. These abilities allow us to convert a regular object detection model into an image classifier with inherent OOD detection capabilities with just minor changes. We compare our approach to state-of-the-art OOD detection methods and demonstrate YolOOD's ability to outperform these methods on a comprehensive suite of in-distribution and OOD benchmark datasets.

The simulated annealing algorithm aims to improve model convergence through multiple restarts of training. However, existing annealing algorithms overlook the correlation between different cycles, neglecting the potential for incremental learning. We contend that a fixed network structure prevents the model from recognizing distinct features at different training stages. To this end, we propose RepAn, redesigning the irreversible re-parameterization (Rep) method and integrating it with annealing to enhance training. Specifically, the network goes through Rep, expansion, restoration, and backpropagation operations during training, and iterating through these processes in each annealing round. Such a method exhibits good generalization and is easy to apply, and we provide theoretical explanations for its effectiveness. Experiments demonstrate that our method improves baseline performance by $6.38\%$ on the CIFAR-100 dataset and $2.80\%$ on ImageNet, achieving state-of-the-art performance in the Rep field. The code is available in our supplementary material.

Dataset distillation offers a lightweight synthetic dataset for fast network training with promising test accuracy. To imitate the performance of the original dataset, most approaches employ bi-level optimization and the distillation space relies on the matching architecture. Nevertheless, these approaches either suffer significant computational costs on large-scale datasets or experience performance decline on cross-architectures. We advocate for designing an economical dataset distillation framework that is independent of the matching architectures, ensuring the versatility of the datasets. With empirical observations, we argue that constraining the consistency of the real and synthetic image spaces will enhance the cross-architecture generalization. Motivated by this, we introduce Dataset Distillation via Disentangled Diffusion Model (D$^4$M), an efficient framework for dataset distillation on large-scale datasets. Compared to architecture-dependent methods, D$^4$M employs diffusion model with an autoencoder to guarantee consistency and incorporates label information into category prototypes, which not only reduces computational overhead but also endows the synthetic image with superior representation capability. The distilled datasets are versatile, eliminating the need for repeated generation of distinct datasets for various architectures. We implement D$^4$M on the ImageNet as well as several other benchmarks. Through comprehensive experiments, D$^4$M demonstrates superior performance and robust generalization, surpassing the SOTA methods across most aspects. Code and distilled datasets will be public.

Today, state-of-the-art deep neural networks that process event-camera data first convert a temporal window of events into dense, grid-like input representations. As such, they exhibit poor generalizability when deployed at higher inference frequencies (i.e., smaller temporal windows) than the ones they were trained on.We address this challenge by introducing state-space models (SSMs) with learnable timescale parameters to event-based vision. This design adapts to varying frequencies without the need to retrain the network at different frequencies. Additionally, we investigate two strategies to counteract aliasing effects when deploying the model at higher frequencies.We comprehensively evaluate our approach against existing methods based on RNN and Transformer architectures across various benchmarks, including Gen1 and 1 Mpx event camera datasets. Our results demonstrate that SSM-based models train 33\% faster and also exhibit minimal performance degradation when tested at higher frequencies than the training input. Traditional RNN and Transformer models exhibit performance drops of more than 20 mAP, with SSMs having a drop of 3.31 mAP, highlighting the effectiveness of SSMs in event-based vision tasks.

The landscape of deep learning research is moving towards innovative strategies to harness the true potential of data. Traditionally, emphasis has been on scaling model architectures, resulting in large and complex neural networks, which can be difficult to train with limited computational resources. However, independently of the model size, data quality (i.e. amount and variability) is still a major factor that affects model generalization.In this work, we propose a novel technique to exploit available data through the use of automatic data augmentation for the tasks of image classification and semantic segmentation. We introduce the first Differentiable Augmentation Search method (DAS) to generate variations of images that can be processed as videos. Compared to previous approaches, DAS is extremely fast and flexible, allowing the search on very large search spaces in less than a GPU day.Our intuition is that the increased receptive field in the temporal dimension provided by DAS could lead to benefits also to the spatial receptive field. More specifically, we leverage DAS to guide the reshaping of the spatial receptive field by selecting task-dependant transformations.As a result, compared to standard augmentation alternatives, we improve in terms of accuracy on ImageNet, Cifar10, Cifar100, Tiny-ImageNet, Pascal-VOC-2012 and CityScapes datasets when plugging-in our DAS over different light-weight video backbones.

In this paper, we address the limitations of the DETR-based semi-supervised object detection (SSOD) framework, particularly focusing on the challenges posed by the quality of object queries. In DETR-based SSOD, the one-to-one assignment strategy provides inaccurate pseudo-labels, while the one-to-many assignments strategy leads to overlapping predictions. These issues compromise training efficiency and degrade model performance, especially in detecting small or occluded objects. We introduce Sparse Semi-DETR, a novel transformer-based, end-to-end semi-supervised object detection solution to overcome these challenges. Sparse Semi-DETR incorporates a Query Refinement Module to enhance the quality of object queries, significantly improving detection capabilities for small and partially obscured objects. Additionally, we integrate a Reliable Pseudo-Label Filtering Module that selectively filters high-quality pseudo-labels, thereby enhancing detection accuracy and consistency. On the MS-COCO and Pascal VOC object detection benchmarks, Sparse Semi-DETR achieves a significant improvement over current state-of-the-art methods that highlight Sparse Semi-DETR's effectiveness in semi-supervised object detection, particularly in challenging scenarios involving small or partially obscured objects.

Robust segmentation is critical for deriving quantitative measures from large-scale, multi-center, and longitudinal medical scans. Manually annotating medical scans, however, is expensive and labor-intensive and may not always be available in every domain. Unsupervised domain adaptation (UDA) is a well-studied technique that alleviates this label-scarcity problem by leveraging available labels from another domain. In this study, we introduce Masked Autoencoding and Pseudo-Labeling Segmentation (MAPSeg), a $\textbf{unified}$ UDA framework with great versatility and superior performance for heterogeneous and volumetric medical image segmentation. To the best of our knowledge, this is the first study that systematically reviews and develops a framework to tackle four different domain shifts in medical image segmentation. More importantly, MAPSeg is the first framework that can be applied to $\textbf{centralized}$, $\textbf{federated}$, and $\textbf{test-time}$ UDA while maintaining comparable performance. We compare MAPSeg with previous state-of-the-art methods on a private infant brain MRI dataset and a public cardiac CT-MRI dataset, and MAPSeg outperforms others by a large margin (10.5 Dice improvement on the private MRI dataset and 5.7 on the public CT-MRI dataset). MAPSeg poses great practical value and can be applied to real-world problems. GitHub: https://github.com/XuzheZ/MAPSeg/.

Federated learning is a promising framework to train neural networks with widely distributed data. However, performance degrades heavily with heterogeneously distributed data. Recent work has shown this is due to the final layer of the network being most prone to local bias, some finding success freezing the final layer as an orthogonal classifier. We investigate the training dynamics of the classifier by applying SVD to the weights motivated by the observation that freezing weights results in constant singular values. We find that there are differences when training in IID and non-IID settings. Based on this finding, we introduce two regularization terms for local training to continuously emulate IID settings: (1) variance in the dimension-wise probability distribution of the classifier and (2) hyperspherical uniformity of representations of the encoder. These regularizations promote local models to act as if it were in an IID setting regardless of the local data distribution, thus offsetting proneness to bias while being flexible to the data. On extensive experiments in both label-shift and feature-shift settings, we verify that our method achieves highest performance by a large margin especially in highly non-IID cases in addition to being scalable to larger models and datasets.

We introduce Florence-2, a novel vision foundation model with a unified, prompt-based representation for a variety of computer vision and vision-language tasks. While existing large vision models excel in transfer learning, they struggle to perform a diversity of tasks with simple instructions, a capability that implies handling the complexity of various spatial hierarchy and semantic granularity. Florence-2 was designed to take text-prompt as task instructions and generate desirable results in text forms, whether it be captioning, object detection, grounding or segmentation. This multi-task learning setup demands large-scale, high-quality annotated data. To this end, we co-developed FLD-5B that consists of 5.4 billion comprehensive visual annotations on 126 million images, using an iterative strategy of automated image annotation and model refinement. We adopted a sequence-to-sequence structure to train Florence-2 to perform versatile and comprehensive vision tasks. Extensive evaluations on numerous tasks demonstrated Florence-2 to be a strong vision foundation model contender with unprecedented zero-shot and fine-tuning capabilities.

Channel pruning approaches for convolutional neural networks (ConvNets) deactivate the channels, statically or dynamically, and require special implementation. In addition, channel squeezing in representative ConvNets is carried out via 1 × 1 convolutions which dominates a large portion of computations and network parameters. Given these challenges, we propose an effective multi-purpose module for dynamic channel sampling, namely Pick-or-Mix (PiX), which does not require special implementation. PiX divides a set of channels into subsets and then picks from them, where the picking decision is dynamically made per each pixel based on the input activations. We plug PiX into prominent ConvNet architectures and verify its multi-purpose utilities. After replacing 1 × 1 channel squeezing layers in ResNet with PiX, the network becomes 25% faster without losing accuracy. We show that PiX allows ConvNets to learn better data representation than widely adopted approaches to enhance networks’ representation power (e.g., SE, CBAM, AFF, SKNet, and DWP). We also show that PiX achieves state-of-the-art performance on network downscaling and dynamic channel pruning applications.

Fine-tuning is the process of extending the training of pre-trained models on specific target tasks, thereby significantly enhancing their performance across various applications.However, fine-tuning often demands large memory consumption, posing a challenge for low-memory devices that some previous memory-efficient fine-tuning methods attempted to mitigate by pruning activations for gradient computation, albeit at the cost of significant computational overhead from the pruning processes during training.To address these challenges, we introduce PreBackRazor, a novel activation pruning scheme offering both computational and memory efficiency through a sparsified backpropagation strategy, which judiciously avoids unnecessary activation pruning and storage and gradient computation.Before activation pruning, our approach samples a probability of selecting a portion of parameters to freeze, utilizing a bandit method for updates to prioritize impactful gradients on convergence. During the feed-forward pass, each model layer adjusts adaptively based on parameter activation status, obviating the need for sparsification and storage of redundant activations for subsequent backpropagation. Benchmarking on fine-tuning foundation models, our approach maintains baseline accuracy across diverse tasks, yielding over 20\% speedup and around 10\% memory reduction. Moreover, integrating with an advanced CUDA kernel achieves up to 60\% speedup without extra memory costs or accuracy loss, significantly enhancing the efficiency of fine-tuning foundation models on memory-constrained devices.

Training-free network architecture search (NAS) aims to discover high-performing networks with zero-cost proxies, capturing network characteristics related to the final performance. However, network rankings estimated by previous training-free NAS methods have shown weak correlations with the performance. To address this issue, we propose AZ-NAS, a novel approach that leverages the ensemble of various zero-cost proxies to enhance the correlation between a predicted ranking of networks and the ground truth substantially in terms of the performance. To achieve this, we introduce four novel zero-cost proxies that are complementary to each other, analyzing distinct traits of architectures in the views of expressivity, progressivity, trainability, and complexity. The proxy scores can be obtained simultaneously within a single forward and backward pass, making an overall NAS process highly efficient. In order to integrate the rankings predicted by our proxies effectively, we introduce a non-linear ranking aggregation method that highlights the networks highly-ranked consistently across all the proxies. Experimental results conclusively demonstrate the efficacy and efficiency of AZ-NAS, outperforming state-of-the-art methods on standard benchmarks, all while maintaining a reasonable runtime cost.

Deep neural networks have shown exemplary performance on semantic scene understanding tasks on source domains, but due to the absence of style diversity during training, enhancing performance on unseen target domains using only single source domain data remains a challenging task. Generation of simulated data is a feasible alternative to retrieving large style-diverse real-world datasets as it is a cumbersome and budget-intensive process. However, the large domain-specific inconsistencies between simulated and real-world data pose a significant generalization challenge in semantic segmentation. In this work, to alleviate this problem, we propose a novel Multi-Resolution Feature Perturbation (MRFP) technique to randomize domain-specific fine-grained features and perturb style of coarse features. Our experimental results on various urban-scene segmentation datasets clearly indicate that, along with the perturbation of style-information, perturbation of fine-feature components is paramount to learn domain invariant robust feature maps for semantic segmentation models. MRFP is a simple and computationally efficient, transferable module with no additional learnable parameters or objective functions, that helps state-of-the-art deep neural networks to learn robust domain invariant features for simulation-to-real semantic segmentation. Code is available at https://github.com/airl-iisc/MRFP.

Recently, model merging techniques have surfaced as a solution to combine multiple single-talent models into a single multi-talent model. However, previous endeavors in this field have either necessitated additional training or fine-tuning processes, or require that the models possess the same pre-trained initialization. In this work, we identify a common drawback in prior works w.r.t. the inconsistency of unit similarity in the weight space and the activation space. To address this inconsistency, we propose an innovative model merging framework, coined as merging under dual-space constraints (MuDSC). Specifically, instead of solely maximizing the objective of a single space, we advocate for the exploration of permutation matrices situated in a region with a unified high similarity in the dual space, achieved through the linear combination of activation and weight similarity matrices. In order to enhance usability, we have also incorporated adaptations for group structure, including Multi-Head Attention and Group Normalization. Comprehensive experimental comparisons demonstrate that MuDSC can significantly boost the performance of merged models with various task combinations and architectures. Furthermore, the visualization of the merged model within the multi-task loss landscape reveals that MuDSC enables the merged model to reside in the overlapping segment, featuring a unified lower loss for each task. Code and models will be made publicly available.

Super-resolution (SR) is an ill-posed inverse problem, where the size of the set of feasible solutions that are consistent with a given low-resolution image is very large. Many algorithms have been proposed to find a "good" solution among the feasible solutions that strike a balance between fidelity and perceptual quality. Unfortunately, all known methods generate artifacts and hallucinations while trying to reconstruct high-frequency (HF) image details. A fundamental question is: Can a model learn to distinguish genuine image details from artifacts? Although some recent works focused on the differentiation of details and artifacts, this is a very challenging problem and a satisfactory solution is yet to be found. This paper shows that the characterization of genuine HF details versus artifacts can be better learned by training GAN-based SR models using wavelet-domain loss functions compared to RGB-domain or Fourier-space losses. Although wavelet-domain losses have been used in the literature before, they have not been used in the context of the SR task. More specifically, we train the discriminator only on the HF wavelet sub-bands instead of on RGB images and the generator is trained by a fidelity loss over wavelet subbands to make it sensitive to the scale and orientation of structures. Extensive experimental results demonstrate that our model achieves better perception-distortion trade-off according to multiple objective measures and visual evaluations. Code and models will be made available.

Advancements in neural radiance fields have allowed for detailed 3D scene reconstructions and novel view synthesis. Yet, efficiently editing these representations while retaining photorealism is an emerging challenge. Recent methods face three primary limitations: they’re slow for interactive use, lack precision at object boundaries, and struggle to ensure view consistency in the edits. In this paper, we introduce IReNe to address these three key limitations, enabling swift, near real-time color editing in NeRF. Leveraging a pre-trained NeRF model and a single training image with user-applied color edits, IReNe swiftly adjusts network parameters in seconds. This adjustment allows the model to generate new scene views, accurately representing the color changes from the training image while also controlling object boundaries and view-specific effects. Enhanced object boundary control is achieved by integrating a trainable segmentation module into the model. The process gains efficiency by retraining only the weights of the last network layer. Moreover, we’ve observed that neurons in this layer can be classified into those responsible for view-dependent effects and those contributing to color rendering. We introduce an automated classification approach to identify these neuron types and exclusively finetune the weights of the color-rendering neurons. This further accelerates training and ensures consistent color edits across different views. A thorough validation on a new dataset, meticulously edited for object colors, demonstrates significant quantitative and qualitative advancements over competitors, accelerating speeds by 5× to 500×.

Nonlinearities are decisive in neural representation learning. Traditional Activation (Act) functions impose fixed inductive biases on neural networks with oriented biological intuitions. Recent methods leverage self-gated curves to compensate for the rigid traditional Act paradigms in fitting flexibility. However, substantial improvements are still impeded by the norm-induced mismatched feature re-calibrations (see Section 1), i.e., the actual importance of a feature can be inconsistent with its explicit intensity such that violates the basic intention of a direct self-gated feature re-weighting. To address this problem, we propose to learn discriminative neural feature Act with a novel prototype, namely, AdaShift, which enhances typical self-gated Act by incorporating an adaptive shift factor into the re-weighting function of Act. AdaShift casts dynamic translations on the inputs of a re-weighting function by exploiting comprehensive feature-filter context cues of different ranges in a simple yet effective manner. We obtain the new intuitions of AdaShift by rethinking the feature-filter relationships from a common Softmax-based classification and by generalizing the new observations to a common learning layer that encodes features with updatable filters. Our practical AdaShifts, built upon the new Act prototype, demonstrate significant improvements to the popular/SOTA Act functions on different vision benchmarks. By simply replacing ReLU with AdaShifts, ResNets can match advanced Transformer counterparts (e.g., ResNet-50 vs. Swin-T) with lower cost and fewer parameters.

Segmentation-based scene text detection algorithms that are accurate to the pixel level can satisfy the detection of arbitrary shape scene text and have received widespread attention. On the one hand, due to the complexity and diversity of the scene text, the convolution with a fixed kernel size has some limitations in extracting the visual features of the scene text. On the other hand, most of the existing segmentation-based algorithms only segment the center of the text, losing information such as the edges and directions of the text, with limited detection accuracy. There are also some improved algorithms that use iterative corrections or introduce other multiple information to improve text detection accuracy but at the expense of efficiency. To address these issues, this paper proposes a simple and effective scene text detection method, the Kernel Adaptive Convolution, which is designed with a Kernel Adaptive Convolution Module for scene text detection via predicting the distance map. Specifically, first, we design an extensible kernel adaptive convolution module (KACM) to extract visual features from multiple convolutions with different kernel sizes in an adaptive manner. Secondly, our method predicts the text distance map under the supervision of a priori information (including direction map, and foreground segmentation map) and completes the text detection from the predicted distance map. Experiments on four publicly available datasets prove the effectiveness of our algorithm, in which the accuracy and efficiency of both the Total-Text and TD500 outperform the state-of-the-art algorithm. The algorithm efficiency is improved while the accuracy is competitive on ArT and Ctw1500.

To defend deep neural networks from adversarial attacks, adversarial training has been drawing increasing attention for its effectiveness. However, the accuracy and robustness resulting from the adversarial training are limited by the architecture, because adversarial training improves accuracy and robustness by adjusting the weight connection affiliated to the architecture. In this work, we propose ARNAS to search for accurate and robust architectures for adversarial training. First we design an accurate and robust search space, in which the placement of the cells and the proportional relationship of the filter numbers are carefully determined. With the design, the architectures can obtain both accuracy and robustness by deploying accurate and robust structures to their sensitive positions, respectively. Then we propose a differentiable multi-objective search strategy, performing gradient descent towards directions that are beneficial for both natural loss and adversarial loss, thus the accuracy and robustness can be guaranteed at the same time. We conduct comprehensive experiments in terms of white-box attacks, black-box attacks, and transferability. Experimental results show that the searched architecture has the strongest robustness with the competitive accuracy, and breaks the traditional idea that NAS-based architectures cannot transfer well to complex tasks in robustness scenarios. By analyzing outstanding architectures searched, we also conclude that accurate and robust neural architectures tend to deploy different structures near the input and output, which has great practical significance on both hand-crafting and automatically designing of accurate and robust architectures.

Existing point cloud semantic segmentation networks cannot identify unknown classes and update their knowledge, due to a closed-set and static perspective of the real world, which would induce the intelligent agent to make bad decisions. To address this problem, we propose a Probability-Driven Framework (PDF) for open world semantic segmentation that includes (i) a lightweight U-decoder branch to identify unknown classes by estimating the uncertainties, (ii) a flexible pseudo-labeling scheme to supply geometry features along with probability distribution features of unknown classes by generating pseudo labels, and (iii) an incremental knowledge distillation strategy to incorporate novel classes into the existing knowledge base gradually. Our framework enables the model to behave like human beings, which could recognize unknown objects and incrementally learn them with the corresponding knowledge. Experimental results on the S3DIS and ScanNetv2 datasets demonstrate that the proposed PDF outperforms other methods by a large margin in both important tasks of open world semantic segmentation.

Revolutionizing the field of deep learning, Transformer-based models have achieved remarkable performance in many tasks. Recent research has recognized these models are robust to shuffling but are limited to inter-token permutation in the forward propagation. In this work, we propose our definition of permutation equivariance, a broader concept covering both inter- and intra- token permutation in the forward and backward propagation of neural networks. We rigorously proved that such permutation equivariance property can be satisfied on most vanilla Transformer-based models with almost no adaptation. We examine the property over a range of state-of-the-art models including ViT, Bert, GPT, and others, with experimental validations. Further, as a proof-of-concept, we explore how real-world applications including privacy-enhancing split learning, and model authorization, could exploit the permutation equivariance property, which implicates wider, intriguing application scenarios. The code is available at \url{https://github.com/Doby-Xu/ST}

Test-time adaptation (TTA) has emerged as a promising solution to address performance decay due to unforeseen distribution shifts between training and test data. While recent TTA methods excel in adapting to test data variations, such adaptability exposes a model to vulnerability against malicious examples, an aspect that has received limited attention. Previous studies have uncovered security vulnerabilities within TTA even when a small proportion of the test batch is maliciously manipulated. In response to the emerging threat, we propose median batch normalization (MedBN), leveraging the robustness of the median for statistics estimation within the batch normalization layer during test-time inference. Our method is algorithm-agnostic, thus allowing seamless integration with existing TTA frameworks. Our experimental results on benchmark datasets, including CIFAR10-C, CIFAR100-C and ImageNet-C, consistently demonstrate that MedBN outperforms existing approaches in maintaining robust performance across different attack scenarios, encompassing both instant and cumulative attacks. Through extensive experiments, we show that our approach sustains the performance even in the absence of attacks, achieving a practical balance between robustness and performance.

Data-free knowledge distillation is able to utilize the knowledge learned by a large teacher network to augment the training of a smaller student network without accessing the original training data, avoiding privacy, security and proprietary risks in real applications. Existing methods typically follow an inversion-and-distillation paradigm in which a generative adversarial network is trained by leveraging the pre-trained teacher network, and is used to synthesize a large-scale sample set for knowledge distillation. In this paper, we reexamine this common data-free knowledge distillation paradigm and show that there is considerable room to improve the overall training efficiency through a lens of small-scale data inversion for distillation. In light of several empirical observations indicating the importance of how to balance class distributions in terms of the synthetic sample diversity and difficulty during both data synthesis and distillation processes, we propose Small Scale Data-free Knowledge Distillation (SSD-KD). In formulation, SSD-KD introduces a modulating function to balance synthetic samples and a priority sampling function to select proper samples, facilitated by the dynamic replay buffer and reinforcement learning. As a result, SSD-KD can perform distillation training conditioned on an extremely small scale of synthetic samples, making the overall training efficiency an order of magnitude faster than current mainstream methods while retaining competitive model performance. Experiments on image classification and semantic segmentation benchmarks demonstrate the efficacy of our method. The code is provided for results reproduction.

To better understand the behavior of image classifiers, it is useful to visualize the contribution of individual pixels to the model prediction.In this study, we propose a method, MoXI (Model eXplanation by Interactions), that efficiently and accurately identifies a group of pixels with high prediction confidence. The proposed method employs game-theoretic concepts, Shapley values and interactions, taking into account the effects of individual pixels and the cooperative influence of pixels on model confidence. Theoretical analysis and experiments demonstrate that our method better identifies the pixels that are highly contributing to the model outputs than widely-used visualization methods using Grad-CAM, Attention rollout, and Shapley value. While prior studies have suffered from the exponential computational cost in the computation of Shapley value and interactions, we show that this can be reduced to linear cost for our task.

Domain shift is a formidable issue in Machine Learning that causes a model to suffer from performance degradation when tested on unseen domains. Federated Domain Generalization (FedDG) attempts to train a global model using collaborative clients in a privacy-preserving manner that can generalize well to unseen clients possibly with domain shift. However, most existing FedDG methods either cause additional privacy risks of data leakage or induce significant costs in client communication and computation, which are major concerns in the Federated Learning paradigm. To circumvent these challenges, here we introduce a novel architectural method for FedDG, namely gPerXAN, which relies on a normalization scheme working with a guiding regularizer. In particular, we carefully design Personalized eXplicitly Assembled Normalization to enforce client models selectively filtering domain-specific features that are biased towards local data while retaining discrimination of those features. Then, we incorporate a simple yet effective regularizer to guide these models in directly capturing domain-invariant representations that the global model’s classifier can leverage. Extensive experimental results on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, indicate that our proposed method outperforms other existing methods in addressing this particular problem.

Redundancy is a persistent challenge in Capsule Networks (CapsNet), leading to high computational costs and parameter counts. Although previous studies have introduced pruning after the initial capsule layer, dynamic routing's fully connected nature and non-orthogonal weight matrices reintroduce redundancy in deeper layers. Besides, dynamic routing requires iterating to converge, further increasing computational demands. In this paper, we propose an Orthogonal Capsule Network (OrthCaps) to reduce redundancy, improve routing performance and decrease parameter counts. Firstly, an efficient pruned capsule layer is introduced to discard redundant capsules. Secondly, dynamic routing is replaced with orthogonal sparse attention routing, eliminating the need for iterations and fully connected structures. Lastly, weight matrices during routing are orthogonalized to sustain low capsule similarity, which is the first approach to use Householder orthogonal decomposition to enforce orthogonality in CapsNet. Our experiments on baseline datasets affirm the efficiency and robustness of OrthCaps in classification tasks, in which ablation studies validate the criticality of each component. OrthCaps-Shallow outperforms other Capsule Network benchmarks on four datasets, utilizing only 110k parameters – a mere 1.25\% of a standard Capsule Network's total. To the best of our knowledge, it achieves the smallest parameter count among existing Capsule Networks. Similarly, OrthCaps-Deep demonstrates competitive performance across four datasets, utilizing only 1.2\% of the parameters required by its counterparts.

Transformer models developed in NLP make a great impact on computer vision fields, producing promising performance on various tasks.While multi-head attention, a characteristic mechanism of the transformer, attracts keen research interest such as for reducing computation cost, we analyze the transformer model from a viewpoint of feature transformation based on a distribution of input feature tokens.The analysis inspires us to derive a novel transformation method from mean-shift update which is an effective gradient ascent to seek a local mode of distinctive representation on the token distribution.We also present an efficient projection approach to reduce parameter size of linear projections constituting the proposed multi-head feature transformation.In the experiments on ImageNet-1K dataset, the proposed methods are embedded into various network models to exhibit favorable performance improvement in place of the transformer module.

The advent of Vision Transformers (ViTs) marks a substantial paradigm shift in the realm of computer vision. ViTs capture the global information of images through self-attention modules, which perform dot product computations among patchified image tokens. While self-attention modules empower ViTs to capture long-range dependencies, the computational complexity grows quadratically with the number of tokens, which is a major hindrance to the practical application of ViTs. Moreover, the self-attention mechanism in deep ViTs is also susceptible to the attention saturation issue. Accordingly, we argue against the necessity of computing the attention scores in every layer, and we propose the Less-Attention Vision Transformer (LaViT), which computes only a few attention operations at each stage and calculates the subsequent feature alignments in other layers via attention transformations that leverage the previously calculated attention scores. This novel approach can mitigate two primary issues plaguing traditional self-attention modules: the heavy computational burden and attention saturation. Our proposed architecture offers superior efficiency and ease of implementation, merely requiring matrix multiplications that are highly optimized in contemporary deep learning frameworks. Moreover, our architecture demonstrates exceptional performance across various vision tasks including classification, detection and segmentation.

High-resolution wide-angle fisheye images are becoming more and more important for robotics applications such as autonomous driving. However, using ordinary convolutional neural networks or vision transformers on this data is problematic due to projection and distortion losses introduced when projecting to a rectangular grid on the plane. We introduce the HEAL-SWIN transformer, which combines the highly uniform Hierarchical Equal Area iso-Latitude Pixelation (HEALPix) grid used in astrophysics and cosmology with the Hierarchical Shifted-Window (SWIN) transformer to yield an efficient and flexible model capable of training on high-resolution, distortion-free spherical data. In HEAL-SWIN, the nested structure of the HEALPix grid is used to perform the patching and windowing operations of the SWIN transformer, enabling the network to process spherical representations with minimal computational overhead. We demonstrate the superior performance of our model on both synthetic and real automotive datasets, as well as a selection of other image datasets, for semantic segmentation, depth regression and classification tasks. Our code will be made available.

Despite their exceptional performance in vision tasks, deep learning models often struggle when faced with domain shifts during testing. Test-Time Training (TTT) methods have recently gained popularity by their ability to enhance the robustness of models through the addition of an auxiliary objective that is jointly optimized with the main task. Being strictly unsupervised, this auxiliary objective is used at test time to adapt the model without any access to labels. In this work, we propose Noise-Contrastive Test-Time Training (NC-TTT), a novel unsupervised TTT technique based on the discrimination of noisy feature maps. By learning to classify noisy views of projected feature maps, and then adapting the model accordingly on new domains, classification performance can be recovered by an important margin. Experiments on several popular test-time adaptation baselines demonstrate the advantages of our method compared to recent approaches for this task. The code can be found at: https://github.com/GustavoVargasHakim/NCTTT.git

Vision Transformers (ViT) and Visual Prompt Tuning (VPT) achieve state-of-the-art performance with improved efficiency in various computer vision tasks. This suggests a promising paradigm shift of adapting pre-trained ViT models to Federated Learning (FL) settings. However, the challenge of data heterogeneity among FL clients presents a significant hurdle in effectively deploying ViT models. Existing Generalized FL (GFL) and Personalized FL (PFL) methods have limitations in balancing performance across both global and local data distributions. In this paper, we present a novel algorithm, SGPT, that integrates GFL and PFL approaches by employing a unique combination of both shared and group-specific prompts. This design enables SGPT to capture both common and group-specific features. A key feature of SGPT is its prompt selection module, which facilitates the training of a single global model capable of automatically adapting to diverse local client data distributions without the need for local fine-tuning. To effectively train the prompts, we utilize block coordinate descent (BCD), learning from common feature information (shared prompts), and then more specialized knowledge (group prompts) iteratively. Theoretically, we justify that learning the proposed prompts can reduce the gap between global and local performance. Empirically, we conduct experiments on both label and feature heterogeneity settings in comparison with state-of-the-art baselines, along with extensive ablation studies, to substantiate the superior performance of SGPT.

This research paper presents a novel class of restoration network architecture based on the Volterra series formulation. By incorporating non-linearity into the system response function through higher order convolutions instead of traditional activation functions, we introduce a general framework for image/video restoration. Through extensive experimentation, we demonstrate that our proposed architecture achieves state-of-the-art (SOTA) performance in the field of Image/Video Restoration. Moreover, we establish that the recently introduced Non-Linear Activation Free Network (NAF-NET) can be considered a special case within the broader class of Volterra Neural Networks. These findings highlight the potential of Volterra Neural Networks as a versatile and powerful tool for addressing complex restoration tasks in computer vision.

We propose to improve transformers of a specific modality with irrelevant data from other modalities, e.g. improve an ImageNet model with audio or point cloud datasets. We would like to highlight that the data samples of the target modality are irrelevant to the other modalities, which distinguishes our method from other works utilizing paired (e.g. CLIP) or interleaved data of different modalities. We propose a methodology named Multimodal Pathway - given a target modality and a transformer designed for it, we use an auxiliary transformer trained with data of another modality and construct pathways to connect components of the two models so that data of the target modality can be processed by both models. In this way, we utilize the universal sequence-to-sequence modeling abilities of transformers obtained from two modalities. As a concrete implementation, we use a modality-specific tokenizer and task-specific head as usual but utilize the transformer blocks of the auxiliary model via a proposed method named Cross-Modal Re-parameterization, which exploits the auxiliary weights without any inference costs. On the image, point cloud, video, and audio recognition tasks, we observe significant and consistent performance improvements with irrelevant data from other modalities. The code and models are available at https://github.com/AILab-CVC/M2PT.

Vision graph neural networks (ViG) offer a new avenue for exploration in computer vision. A major bottleneck in ViGs is the inefficient k-nearest neighbor (KNN) operation used for graph construction. To solve this issue, we propose a new method for designing ViGs, Dynamic Axial Graph Construction (DAGC), which is more efficient than KNN as it limits the number of considered graph connections made within an image. Additionally, we propose a novel CNN-GNN architecture, GreedyViG, which uses DAGC. Extensive experiments show that GreedyViG beats existing ViG, CNN, and ViT architectures in terms of accuracy, GMACs, and parameters on image classification, object detection, instance segmentation, and semantic segmentation tasks. Our smallest model, GreedyViG-S, achieves 81.1\% top-1 accuracy on ImageNet-1K, 2.9\% higher than Vision GNN and 2.2\% higher than Vision HyperGraph Neural Network (ViHGNN), with less GMACs and a similar number of parameters. Our largest model, GreedyViG-B obtains 83.9\% top-1 accuracy, 0.2\% higher than Vision GNN, with a 66.6\% decrease in parameters and a 69\% decrease in GMACs. GreedyViG-B also obtains the same accuracy as ViHGNN with a 67.3\% decrease in parameters and a 71.3\% decrease in GMACs. Our work shows that hybrid CNN-GNN architectures not only provide a new avenue for designing efficient models, but that they can also exceed the performance of current state-of-the-art models$^1$.

The success of a specific neural network architecture is closely tied to the dataset and task it tackles; there is no one-size-fits-all solution. Thus, considerable efforts have been made to quickly and accurately estimate the performances of neural architectures, without full training or evaluation, for given tasks and datasets. Neural architecture encoding has played a crucial role in the estimation, and graph-based methods, which treat an architecture as a graph, have shown prominent performance. For enhanced representation learning of neural architectures, we introduce FlowerFormer, a powerful graph transformer that incorporates the information flows within a neural architecture. FlowerFormer consists of two key components: (a) bidirectional asynchronous message passing, inspired by the flows; (b) global attention built on flow-based masking. Our extensive experiments demonstrate the superiority of FlowerFormer over existing neural encoding methods, and its effectiveness extends beyond computer vision models to include graph neural networks and auto speech recognition models.

While federated learning (FL) systems often utilize quantization to battle communication and computational bottlenecks, they have heretofore been limited to deploying fixed-precision quantization schemes. Meanwhile, the concept of mixed-precision quantization (MPQ), where different layers of a deep learning model are assigned varying bit-width, remains unexplored in the FL settings. We present a novel FL algorithm, FedMPQ, which introduces mixed-precision quantization to resource-heterogeneous FL systems. Specifically, local models, quantized so as to satisfy bit-width constraint, are trained by optimizing an objective function that includes a regularization term which promotes reduction of precision in some of the layers without significant performance degradation. The server collects local model updates, de-quantizes them into full-precision models, and then aggregates them into a global model. To initialize the next round of local training, the server relies on the information learned in the previous training round to customize bit-width assignments of the models delivered to different clients. In extensive benchmarking experiments on several model architectures and different datasets in both iid and non-iid settings, FedMPQ outperformed the baseline FL schemes that utilize fixed-precision quantization while incurring only a minor computational overhead on the participating devices.

Neural field is an emerging paradigm in data representation that trains a neural network to approximate the given signal. A key obstacle that prevents its widespread adoption is the encoding speed - generating neural fields requires an overfitting of a neural network, which can take a significant number of SGD steps to reach the desired fidelity level. In this paper, we delve into the impacts of data transformations on the speed of neural field training, specifically focusing on how permuting pixel locations affect the convergence speed of SGD. Counterintuitively, we find that randomly permuting the pixel locations can considerably accelerate the training. To explain this phenomenon, we examine the neural field training through the lens of PSNR curves, loss landscapes, and error patterns. Our analyses suggest that the random pixel permutations remove the easy-to-fit patterns, which facilitate easy optimization in the early stage but hinder capturing fine details of the signal.

Creating multi-view wire art (MVWA), a static 3D sculpture with diverse interpretations from different viewpoints, is a complex task even for skilled artists. In response, we present DreamWire, an AI system enabling everyone to craft MVWA easily. Users express their vision through text prompts or scribbles, freeing them from intricate 3D wire organisation. Our approach synergises 3D Bezier curves, Prim's algorithm, and knowledge distillation from diffusion models or their variants (e.g., ControlNet). This blend enables the system to represent 3D wire art, ensuring spatial continuity and overcoming data scarcity. Extensive evaluation and analysis are conducted to shed insight on the inner workings of the proposed system, including the trade-off between connectivity and visual aesthetics.

High-resolution image generation with Generative Artificial Intelligence (GenAI) has immense potential but, due to the enormous capital investment required for training, it is increasingly centralised to a few large corporations, and hidden behind paywalls. This paper aims to democratise high-resolution GenAI by advancing the frontier of high-resolution generation while remaining accessible to a broad audience. We demonstrate that existing Latent Diffusion Models (LDMs) possess untapped potential for higher-resolution image generation. Our novel DemoFusion framework seamlessly extends open-source GenAI models, employing Progressive Upscaling, Skip Residual, and Dilated Sampling mechanisms to achieve higher-resolution image generation. The progressive nature of DemoFusion requires more passes, but the intermediate results can serve as "previews", facilitating rapid prompt iteration.

Existing diffusion models for pose-guided human video generation mostly suffer from temporal inconsistency in the generated appearance and poses due to the inherent randomization nature of the generation process. In this paper, we propose a novel framework, DiffPerformer, to synthesize high-fidelity and temporally consistent human video. Without complex architecture modification or costly training, DiffPerformer finetunes a pretrained diffusion model on a single video of the target character and introduces an implicit video representation as a proxy to learn temporally consistent guidance for the diffusion model. The guidance is encoded into VAE latent space and an iterative optimization loop is constructed between the implicit video representation and the diffusion model, allowing to harness the smooth property of the implicit video representation and the generative capabilities of the diffusion model in a mutually beneficial way. Moreover, we propose 3D-aware human flow as a temporal constraint during the optimization to explicitly model the correspondence between driving poses and human appearance. This alleviates the misalignment between guided poses and target performer and therefore maintains the appearance coherence under various motions. Extensive experiments demonstrate that our method outperforms the state-of-the-art methods.

Large-scale text-to-image (T2I) diffusion models have showcased incredible capabilities in generating coherent images based on textual descriptions, enabling vast applications in content generation. While recent advancements have introduced control over factors such as object localization, posture, and image contours, a crucial gap remains in our ability to control the interactions between objects in the generated content. Well-controlling interactions in generated images could yield meaningful applications, such as creating realistic scenes with interacting characters. In this work, we study the problems of conditioning T2I diffusion models with Human-Object Interaction (HOI) information, consisting of a triplet label (person, action, object) and corresponding bounding boxes. We propose a pluggable interaction control model, called InteractDiffusion that extends existing pre-trained T2I diffusion models to enable them being better conditioned on interactions. Specifically, we tokenize the HOI information and learn their relationships via interaction embeddings. A conditioning self-attention layer is trained to map HOI tokens to visual tokens, thereby conditioning the visual tokens better in existing T2I diffusion models. Our model attains the ability to control the interaction and location on existing T2I diffusion models, which outperforms existing baselines by a large margin in HOI detection score, as well as fidelity in FID and KID. The code will be made available.

Generative models have recently exhibited exceptional capabilities in text-to-image generation, but still struggle to generate image sequences coherently. In this work, we focus on a novel, yet challenging task of generating a coherent image sequence based on a given storyline, denoted as open-ended visual storytelling.We make the following three contributions:(i) to fulfill the task of visual storytelling,we propose a learning-based auto-regressive image generation model, termed as StoryGen, with a novel vision-language context module, that enables to generate the current frame by conditioning on the corresponding text prompt and preceding image-caption pairs;(ii) to address the data shortage of visual storytelling, we collect paired image-text sequences by sourcing from online videos and open-source E-books, establishing processing pipeline for constructing a large-scale dataset with diverse characters, storylines, and artistic styles, named StorySalon;(iii) Quantitative experiments and human evaluations have validated the superiority of our StoryGen, where we show StoryGen can generalize to unseen characters without any optimization, and generate image sequences with coherent content and consistent character.The code, model, and dataset will be made publicly available to the research community.

Manually creating 3D environments for AR/VR applications is a complex process requiring expert knowledge in 3D modeling software.Pioneering works facilitate this process by generating room meshes conditioned on textual style descriptions.Yet, many of these automatically generated 3D meshes do not adhere to typical room layouts, compromising their plausibility, e.g., by placing several beds in one bedroom.To address these challenges, we present ControlRoom3D, a novel method to generate high-quality room meshes.Central to our approach is a user-defined 3D semantic proxy room that outlines a rough room layout based on semantic bounding boxes and a textual description of the overall room style.Our key insight is that when rendered to 2D, this 3D representation provides valuable geometric and semantic information to control powerful 2D models to generate 3D consistent textures and geometry that aligns well with the proxy room.Backed up by an extensive study including quantitative metrics and qualitative user evaluations, our method generates diverse and globally plausible 3D room meshes, thus empowering users to design 3D rooms effortlessly without specialized knowledge.

Diffusion models have recently revolutionized the field of image synthesis due to their ability to generate photorealistic images. However, one of the major drawbacks of diffusion models is that the image generation process is costly. A large image-to-image network has to be applied many times to iteratively refine an image from random noise. While many recent works propose techniques to reduce the number of required steps, they generally treat the underlying denoising network as a black box. In this work, we investigate the behavior of the layers within the network and find that 1) the layers' output changes smoothly over time, 2) the layers show distinct patterns of change, and 3) the change from step to step is often very small.e hypothesize that many layer computations in the denoising network are redundant. Leveraging this, we introduce block caching, in which we reuse outputs from layer blocks of previous steps to speed up inference. Furthermore, we propose a technique to automatically determine caching schedules based on each block's changes over timesteps. In our experiments, we show through FID, human evaluation and qualitative analysis that Block Caching allows to generate images with higher visual quality at the same computational cost. We demonstrate this for different state-of-the-art models (LDM and EMU) and solvers (DDIM and DPM).

Synthesizing realistic videos of talking faces under custom lighting conditions and viewing angles benefits various downstream applications like video conferencing. However, most existing relighting methods are either time-consuming or unable to adjust the viewpoints. In this paper, we present the first real-time 3D-aware method for relighting in-the-wild videos of talking faces based on Neural Radiance Fields (NeRF). Given an input portrait video, our method can synthesize talking faces under both novel views and novel lighting conditions with a photo-realistic and disentangled 3D representation. Specifically, we infer an albedo tri-plane, as well as a shading tri-plane based on a desired lighting condition for each video frame with fast dual-encoders. We also leverage a temporal consistency network to ensure smooth transitions and reduce flickering artifacts. Our method runs at 32.98 fps on consumer-level hardware and achieves state-of-the-art results in terms of reconstruction quality, lighting error, lighting instability, temporal consistency and inference speed. We demonstrate the effectiveness and interactivity of our method on various portrait videos with diverse lighting and viewing conditions.

Text-to-image diffusion models produce high quality images but do not offer control over individual instances in the image. We introduce InstanceDiffusion that adds precise instance-level control to text-to-image diffusion models. InstanceDiffusion supports free-form language conditions per instance and allows flexible ways to specify instance locations such as simple single points, scribbles, bounding boxes or intricate instance segmentation masks, and combinations thereof. We propose three major changes to text-to-image models that enable precise instance-level control. Our UniFusion block enables instance-level conditions for text-to-image models, the ScaleU block improves image fidelity, and our Multi-instance Sampler improves generations for multiple instances. InstanceDiffusion significantly surpasses specialized state-of-the-art models for each location condition. Notably, on the COCO dataset, we outperform previous state-of-the-art by 20.4\% AP$_{50}^\text{box}$ for box inputs, and 25.4\% IoU for mask inputs.

Creating and animating 3D biped cartoon characters is crucial and valuable in various applications. Compared with geometry, the diverse texture design plays an important role in making 3D biped cartoon characters vivid and charming. Therefore, we focus on automatic texture design for cartoon characters based on input instructions. This is challenging for domain-specific requirements and a lack of high-quality data. To address this challenge, we propose Make-It-Vivid, the first attempt to enable high-quality texture generation from text in UV space. We prepare a detailed text-texture paired data for 3D characters by using vision-question-answering agents. Then we customize a pretrained text-to-image model to generate texture map with template structure while preserving the natural 2D image knowledge. Furthermore, to enhance fine-grained details, we propose a novel adversarial learning scheme to shorten the domain gap between original dataset and realistic texture domain. Extensive experiments show that our approach outperforms current texture generation methods, resulting in efficient character texturing and faithful generation with prompts. Besides, we showcase various applications such as out of domain generation and texture stylization. We also provide an efficient generation system for automatic text-guided textured character generation and animation.

Recent advances in vision-language models like Stable Diffusion have shown remarkable power in creative image synthesis and editing.However, most existing text-to-image editing methods encounter two obstacles: First, the text prompt needs to be carefully crafted to achieve good results, which is not intuitive or user-friendly. Second, they are insensitive to local edits and can irreversibly affect non-edited regions, leaving obvious editing traces. To tackle these problems, we propose a Zero-shot instructiON-guided local image Editing approach, termed $\texttt{ZONE}$. We first convert the editing intent from the user-provided instruction (e.g., ``make his tie blue") into specific image editing regions through InstructPix2Pix. We then propose a Region-IoU scheme for precise image layer extraction from an off-the-shelf segment model. We further develop an edge smoother based on FFT for seamless blending between the layer and the image.Our method allows for arbitrary manipulation of a specific region with a single instruction while preserving the rest. Extensive experiments demonstrate that our $\texttt{ZONE}$ achieves remarkable local editing results and user-friendliness, outperforming state-of-the-art methods. Code is available at https://github.com/lsl001006/ZONE.

Conditional diffusion models are powerful generative models that can leverage various types of conditional information, such as class labels, segmentation masks, or text captions. However, in many real-world scenarios, conditional information may be noisy or unreliable due to human annotation errors or weak alignment. In this paper, we propose the Coherence-Aware Diffusion (CAD), a novel method to integrate confidence in conditional information into diffusion models, allowing them to learn from noisy annotations without discarding data. We assume that each data point has an associated confidence score that reflects the quality of the conditional information. We then condition the diffusion model on both the conditional information and the confidence score. In this way, the model learns to ignore or discount the conditioning when the confidence is low. We show that our method is theoretically sound and empirically effective on various conditional generation tasks. Moreover, we show that leveraging confidence generates realistic and diverse samples that respect conditional information better than models trained on cleaned datasets where samples with low confidence have been discarded.

We introduce a new task of generating “Illustrated Instructions”, i.e. visual instructions customized to a user’s needs. We identify desiderata unique to this task, and formalize it through a suite of automatic and human evaluation metrics, designed to measure the validity, consistency, and efficacy of the generations. We combine the power of large language models (LLMs) together with strong text-to-image generation diffusion models to propose a simple approach called StackedDiffusion, that generates such illustrated instructions given text as input. The resulting model strongly outperforms baseline approaches and state-of-the-art multimodal LLMs; and in 30% of cases, users even prefer it to human-generated articles. Most notably, it enables various new and exciting applications far beyond what static articles on the web can provide, such as personalized instructions complete with intermediate steps and pictures in response to a user’s individual situation.

Spike cameras, leveraging spike-based integration sampling and high temporal resolution, offer distinct advantages over standard cameras. However, existing approaches reliant on spike cameras often assume optimal illumination, a condition frequently unmet in real-world scenarios. To address this, we introduce SpikeNeRF, the first work that derives a NeRF-based volumetric scene representation from spike camera data. Our approach leverages NeRF's multi-view consistency to establish robust self-supervision, effectively eliminating erroneous measurements and uncovering coherent structures within exceedingly noisy input amidst diverse real-world illumination scenarios. The framework comprises two core elements: a spike generation model incorporating an integrate-and-fire neuron layer and parameters accounting for non-idealities, such as threshold variation, and a spike rendering loss capable of generalizing across varying illumination conditions. We describe how to effectively optimize neural radiance fields to render photorealistic novel views from the novel continuous spike stream, demonstrating advantages over other vision sensors in certain scenes. Empirical evaluations conducted on both real and novel realistically simulated sequences affirm the efficacy of our methodology. The dataset and source code are released at https://github.com/BIT-Vision/SpikeNeRF.

Recently, dataset distillation has paved the way towards efficient machine learning, especially for image datasets. However, the distillation for videos, characterized by an exclusive temporal dimension, remains an underexplored domain. In this work, we provide the first systematic study of video distillation and introduce a taxonomy to categorize temporal compression. Our investigation reveals that the temporal information is usually not well learned during distillation, and the temporal dimension of synthetic data contributes little. The observations motivate our unified framework of disentangling the dynamic and static information in the videos. It first distills the videos into still images as static memory and then compensates the dynamic and motion information with a learnable dynamic memory block. Our method achieves state-of-the-art on video datasets at different scales, with notably smaller memory storage budget. Our code is available at https://github.com/yuz1wan/video_distillation.

This paper introduces a novel unified representation of diffusion models for image generation and segmentation. Specifically, we use a colormap to represent entity-level masks, addressing the challenge of varying entity numbers while aligning the representation closely with the image RGB domain. Two novel modules, including the location-aware color palette and progressive dichotomy module, are proposed to support our mask representation. On the one hand, a location-aware palette guarantees the colors' consistency to entities' locations. On the other hand, the progressive dichotomy module can efficiently decode the synthesized colormap to high-quality entity-level masks in a depth-first binary search without knowing the cluster numbers. To tackle the issue of lacking large-scale segmentation training data, we employ an inpainting pipeline and then improve the flexibility of diffusion models across various tasks, including inpainting, image synthesis, referring segmentation, and entity segmentation. Comprehensive experiments validate the efficiency of our approach, demonstrating comparable segmentation mask quality to state-of-the-art and adaptability to multiple tasks. We will make both the code and the model available to the public.

Existing GAN-based text-to-image models treat images as 2D pixel arrays. In this paper, we approach the text-to-image task from a different perspective, where a 2D image is represented as an implicit neural representation (INR). We show that straightforward conditioning of the unconditional INR-based GAN method on text inputs is not enough to achieve good performance. We propose a word-level attention-based weight modulation operator which controls the generation process of INR-GAN based on hypernetworks. Our experiments on benchmark datasets show that HyperCGAN achieves competitive performance to existing pixel-based methods and retain the properties of continuous generative models. Code, models, and benchmarks will be made publicly available.

Text-to-image generation has witnessed significant progress with the advent of diffusion models. Despite the ability to generate photorealistic images, current text-to-image diffusion models still often struggle to accurately interpret and follow complex input text prompts. In contrast to existing models that aim to generate images only with their best effort, we introduce Self-correcting LLM-controlled Diffusion (SLD). SLD is a framework that generates an image from the input prompt, assesses its alignment with the prompt, and performs self-corrections on the inaccuracies in the generated image. Steered by an LLM controller, SLD turns text-to-image generation into an iterative closed-loop process, ensuring correctness in the resulting image. SLD is not only training-free but can also be seamlessly integrated with diffusion models behind API access, such as DALL-E 3, to further boost the performance of state-of-the-art diffusion models. Experimental results show that our approach can rectify a majority of incorrect generations, particularly in generative numeracy, attribute binding, and spatial relationships. Furthermore, by simply adjusting the instructions to the LLM, SLD can perform image editing tasks, bridging the gap between text-to-image generation and image editing pipelines. We will make our code available for future research and applications.

Despite many attempts to leverage pre-trained text-to-image models (T2I) like Stable Diffusion (SD) for controllable image editing, producing good predictable results remains a challenge. Previous approaches have focused on either fine-tuning pre-trained T2I models on specific datasets to generate certain kinds of images (e.g., with a specific object or person), or on optimizing the weights, text prompts, and/or learning features for each input image in an attempt to coax the image generator to produce the desired result. However, these approaches all have shortcomings and fail to produce good results in a predictable and controllable manner. To address this problem, we present TiNO-Edit, an SD-based method that focuses on optimizing the noise patterns and diffusion timesteps during editing, something previously unexplored in the literature. With this simple change, we are able to generate results that both better align with the original images and reflect the desired result. Furthermore, we propose a set of new loss functions that operate in the latent domain of SD, greatly speeding up the optimization when compared to prior approaches, which operate in the pixel domain. Our method can be easily applied to variations of SD including Textual Inversion and DreamBooth that encode new concepts and incorporate them into the edited results. We present a host of image-editing applications enabled by our approach.

Generative models, e.g., Stable Diffusion, have enabled the creation of photorealistic images from text prompts.Yet, the generation of 360-degree panorama images from text remains a challenge, particularly due to the dearth of paired text-panorama data and the domain gap between panorama and perspective images.In this paper, we introduce a novel dual-branch diffusion model named PanFusion to generate a 360-degree image from a text prompt.We leverage the stable diffusion model as one branch to provide prior knowledge in natural image generation and register it to another panorama branch for holistic image generation.We propose a unique cross-attention mechanism with projection awareness to minimize distortion during the collaborative denoising process.Our experiments validate that PanFusion surpasses existing methods and, thanks to its dual-branch structure, can integrate additional constraints like room layout for customized panorama outputs.

Recent years have witnessed remarkable progress in image generation task, where users can create visually astonishing images with high-quality. However, existing text-to-image diffusion models are proficient in generating concrete concepts (dogs) but encounter challenges with more abstract ones (emotions). Several efforts have been made to modify image emotions with color and style adjustments, facing limitations in effectively conveying emotions with fixed image contents. In this work, we introduce Emotional Image Content Generation (EICG), a new task to generate semantic-clear and emotion-faithful images given emotion categories. Specifically, we propose an emotion space and construct a mapping network to align it with the powerful Contrastive Language-Image Pre-training (CLIP) space, providing a concrete interpretation of abstract emotions. Attribute loss and emotion confidence are further proposed to ensure the semantic diversity and emotion fidelity of the generated images. Our method outperforms the state-of-the-art text-to-image approaches both quantitatively and qualitatively, where we derive three custom metrics, i.e., emotion accuracy, semantic clarity and semantic diversity. In addition to generation, our method can help emotion understanding and inspire emotional art design.

Multi-view diffusion models, obtained by applying Supervised Finetuning (SFT) to text-to-image diffusion models, have driven recent breakthroughs in text-to-3D research. However, due to the limited size and quality of existing 3D datasets, they still suffer from multi-view inconsistencies and Neural Radiance Field (NeRF) reconstruction artifacts. We argue that multi-view diffusion models can benefit from further Reinforcement Learning Finetuning (RLFT), which allows models to learn from the data generated by themselves and improve beyond their dataset limitations during SFT. To this end, we introduce Carve3D, an improved RLFT algorithm coupled with a novel Multi-view Reconstruction Consistency (MRC) metric, to enhance the consistency of multi-view diffusion models. To measure the MRC metric on a set of multi-view images, we compare them with their corresponding NeRF renderings at the same camera viewpoints. The resulting model, which we denote as Carve3DM, demonstrates superior multi-view consistency and NeRF reconstruction quality than existing models. Our results suggest that pairing SFT with Carve3D's RLFT is essential for developing multi-view-consistent diffusion models, mirroring the standard Large Language Model (LLM) alignment pipeline. Our code, training and testing data, and video results are available at: https://desaixie.github.io/carve-3d.

In this paper, we uncover the untapped potential of diffusion U-Net, which serves as a "free lunch" that substantially improves the generation quality on the fly. We initially investigate the key contributions of the U-Net architecture to the denoising process and identify that its main backbone primarily contributes to denoising, whereas its skip connections mainly introduce high-frequency features into the decoder module, causing the potential neglect of crucial functions intrinsic to the backbone network. Capitalizing on this discovery, we propose a simple yet effective method, termed ``\textbf{FreeU}'', which enhances generation quality without additional training or finetuning. Our key insight is to strategically re-weight the contributions sourced from the U-Net’s skip connections and backbone feature maps, to leverage the strengths of both components of the U-Net architecture. Promising results on image and video generation tasks demonstrate that our FreeU can be readily integrated to existing diffusion models, e.g., Stable Diffusion, DreamBooth and ControlNet, to improve the generation quality with only a few lines of code. All you need is to adjust two scaling factors during inference.

Diffusion models generate images with an unprecedented level of quality, but how can we freely rearrange image layouts? Recent works generate controllable scenes via learning spatially disentangled latent codes, but these methods do not apply to diffusion models due to their fixed forward process. In this work, we propose $\textbf{SceneDiffusion}$ to optimize a layered scene representation during the diffusion sampling process. Our key insight is that spatial disentanglement can be obtained by jointly denoising scene renderings at different spatial layouts. Our generated scenes support a wide range of spatial editing operations, including moving, resizing, cloning, and layer-wise appearance editing operations, including object restyling and replacing. Moreover, a scene can be generated conditioned on a reference image, thus enabling object moving for in-the-wild images. Notably, this approach is training-free, compatible with general text-to-image diffusion models, and responsive in less than a second.

Text-to-image (T2I) generative models have attracted significant attention and found extensive applications within and beyond academic research. For example, the Civitai community, a platform for T2I innovation, currently hosts an impressive array of 74,492 distinct models. However, this diversity presents a formidable challenge in selecting the most appropriate model and parameters, a process that typically requires numerous trials. Drawing inspiration from the tool usage research of large language models (LLMs), we introduce DiffAgent, an LLM agent designed to screen the accurate selection in seconds via API calls. DiffAgent leverages a novel two-stage training framework, SFTA, enabling it to accurately align T2I API responses with user input in accordance with human preferences. To train and evaluate DiffAgent's capabilities, we present DABench, a comprehensive dataset encompassing an extensive range of T2I APIs from the community. Our evaluations reveal that DiffAgent not only excels in identifying the appropriate T2I API but also underscores the effectiveness of the SFTA training framework. Codes are available at https://github.com/OpenGVLab/DiffAgent.

We concentrate on a novel human-centric image synthesis task, that is, given only one reference facial photograph, it is expected to generate specific individual images with diverse head positions, poses, and facial expressions in different contexts. To accomplish this goal, we argue that our generative model should be capable of the following favorable characteristics: (1) a strong visual and semantic understanding of our world and human society for basic object and human image generation. (2) generalizable identity preservation ability. (3) flexible and fine-grained head control. Recently, large pre-trained text-to-image diffusion models have shown remarkable results, serving as a powerful generative foundation. As a basis, we aim to unleash the above two capabilities of the pre-trained model. In this work, we present a new framework named CapHuman. We embrace the "encode then learn to align" paradigm, which enables generalizable identity preservation for new individuals without cumbersome tuning at inference. CapHuman encodes identity features and then learns to align them into the latent space. Moreover, we introduce the 3D facial prior to equip our model with control over the human head in a flexible and 3D-consistent manner. Extensive qualitative and quantitative analyses demonstrate our CapHuman can produce well-identity-preserved, photo-realistic, and high-fidelity portraits with content-rich representations and various head renditions, superior to established baselines.

Advanced video frame interpolation (VFI) algorithms approximate intermediate motions between two input frames to synthesize intermediate frame. However, they struggle to handle complex scenarios with curvilinear motions since they neglect the latent acceleration information between input frames. Moreover, the supervision of predicted motions are tricky because real ground-truth motions are available. To this end, we propose a novel framework for implicit quadratic video frame interpolation (IQ-009VFI), which explores latent acceleration information and accurate intermediate motions via knowledge distillation. Specifically, the proposed IQ-VFI consists of an implicit acceleration estimation network (IAE) and a VFI backbone, the former is to extract latent acceleration priors between two input frames to progressively upgrade linear motions from the latter in coarse-to-fine manner. Moreover, to encourage both components to distill more acceleration and motion cues oriented towards VFI, we propose a knowledge distillation strategy in which implicit acceleration distillation loss and implicit motion distillation loss are used to adaptively guide latent acceleration priors and intermediate motions learning, respectively. Extensive experiments show that our proposed IQ-VFI can achieve state-of-the-art performances on various benchmark dataset

Diffusion model is a promising approach to image generation and has been employed for Pose-Guided Person Image Synthesis (PGPIS) with competitive performance. While existing methods simply align the person appearance to the target pose, they are prone to overfitting due to the lack of a high-level semantic understanding on the source person image. In this paper, we propose a novel Coarse-to-Fine Latent Diffusion (CFLD) method for PGPIS. In the absence of image-caption pairs and textual prompts, we develop a novel training paradigm purely based on images to control the generation process of the pre-trained text-to-image diffusion model. A perception-refined decoder is designed to progressively refine a set of learnable queries and extract semantic understanding of person images as a coarse-grained prompt. This allows for the decoupling of fine-grained appearance and pose information controls at different stages, and thus circumventing the potential overfitting problem. To generate more realistic texture details, a hybrid-granularity attention module is proposed to encode multi-scale fine-grained appearance features as bias terms to augment the coarse-grained prompt. Both quantitative and qualitative experimental results on the DeepFashion benchmark demonstrate the superiority of our method over the state of the arts for PGPIS. Code is available in the supplementary materials and will be released to the public.

The rapid expansion of large-scale text-to-image diffusion models has raised growing concerns regarding their potential misuse in creating harmful or misleading content. In this paper, we introduce MACE, a finetuning framework for the task of mass concept erasure. This task aims to prevent models from generating images that embody unwanted concepts when prompted. Existing concept erasure methods are typically restricted to handling fewer than five concepts simultaneously and struggle to find a balance between erasing concept synonyms (generality) and maintaining unrelated concepts (specificity). In contrast, MACE differs by successfully scaling the erasure scope up to 100 concepts and by achieving an effective balance between generality and specificity. This is achieved by leveraging closed-form cross-attention refinement along with LoRA finetuning, collectively eliminating the information of undesirable concepts. Furthermore, MACE integrates multiple LoRAs without mutual interference. We conduct extensive evaluations of MACE against prior methods across four different tasks: object erasure, celebrity erasure, explicit content erasure, and artistic style erasure. Our results reveal that MACE surpasses prior methods in all evaluated tasks. Code is available at https://github.com/Shilin-LU/MACE.

In this study, we explore Transformer, based diffusion models for image and video generation. Despite the dominance of Transformer architectures in various fields due to their flexibility and scalability, the visual generative domain primarily utilizes CNN-based U-Net architectures, particularly in diffusion-based models. We introduce GenTron, a family of Generative models employing Transformer-based diffusion, to address this gap. Our initial step was to adapt Diffusion Transformers (DiTs) from class to text conditioning, a process involving thorough empirical exploration of the conditioning mechanism. We then scale GenTron from approximately 900M to over 3B parameters, observing improvements in visual quality. Furthermore, we extend GenTron to text-to-video generation, incorporating novel motion-free guidance to enhance video quality. In human evaluations against SDXL, GenTron achieves a 51.1% win rate in visual quality (with a 19.8% draw rate), and a 42.3% win rate in text alignment (with a 42.9% draw rate). GenTron notably performs well in T2I-CompBench, highlighting its compositional generation ability. We hope GenTron could provide meaningful insights and serve as a valuable reference for future research. Website is available at https://www.shoufachen.com/gentron_website/.

Portrait harmonization aims to composite a subject into a new background, adjusting its lighting and color to ensure harmony with the background scene. Existing harmonization techniques often only focus on adjusting the global color and brightness of the foreground and ignore crucial illumination cues from the background such as apparent lighting direction, leading to unrealistic compositions.We introduce Relightful Harmonization, a lighting-aware diffusion model designed to seamlessly harmonize sophisticated lighting effect for the foreground portrait using any background image.Our approach unfolds in three stages. First, we introduce a lighting representation module that allows our diffusion model to encode lighting information from target image background.Second, we introduce an alignment network that aligns lighting features learned from image background with lighting features learned from panorama environment maps, which is a complete representation for scene illumination.Last, to further boost the photorealism of the proposed method, we introduce a novel data simulation pipeline that generates synthetic training pairs from a diverse range of natural images, which are used to refine the model.Our method outperforms existing benchmarks in visual fidelity and lighting coherence, showing superior generalization in real-world testing scenarios, highlighting its versatility and practicality.

Diffusion models have emerged as the de facto paradigm for video generation. However, their reliance on web-scale data of varied quality often yields results that are visually unappealing and misaligned with the textual prompts. To tackle this problem, we propose InstructVideo to instruct text-to-video diffusion models with human feedback by reward fine-tuning. InstructVideo has two key ingredients: 1) To ameliorate the cost of reward fine-tuning induced by generating through the full DDIM sampling chain, we recast reward fine-tuning as editing. By leveraging the diffusion process to corrupt a sampled video, InstructVideo requires only partial inference of the DDIM sampling chain, reducing fine-tuning cost while improving fine-tuning efficiency. 2) To mitigate the absence of a dedicated video reward model for human preferences, we repurpose established image reward models, e.g., HPSv2. To this end, we propose Segmental Video Reward, a mechanism to provide reward signals based on segmental sparse sampling, and Temporally Attenuated Reward, a method that mitigates temporal modeling degradation during fine-tuning. Extensive experiments, both qualitative and quantitative, validate the practicality and efficacy of using image reward models in InstructVideo, significantly enhancing the visual quality of generated videos without compromising generalization capabilities. Code and models can be accessed through our project page https://instructvideo.github.io/.

Human-centric video frame interpolation has great potential for enhancing entertainment experiences and finding commercial applications in the sports analysis industry, e.g., synthesizing slow-motion videos. Although there are multiple benchmark datasets available for video frame interpolation in the community, none of them is dedicated to human-centric scenarios. To bridge this gap, we introduce SportsSloMo, a benchmark featuring over 130K high-resolution ($\geq$720p) slow-motion sports video clips, totaling over 1M video frames, sourced from YouTube. We re-train several state-of-the-art methods on our benchmark, and we observed a noticeable decrease in their accuracy compared to other datasets. This highlights the difficulty of our benchmark and suggests that it poses significant challenges even for the best-performing methods, as human bodies are highly deformable and occlusions are frequent in sports videos. To tackle these challenges, we propose human-aware loss terms, where we add auxiliary supervision for human segmentation in panoptic settings and keypoints detection. These loss terms are model-agnostic and can be easily plugged into any video frame interpolation approach. Experimental results validate the effectiveness of our proposed human-aware loss terms, leading to consistent performance improvement over existing models. The dataset and code will be publicly released to foster future research.

Neural Radiance Fields (NeRF) revolutionize the realm of visual media by providing photorealistic Free-Viewpoint Video (FVV) experiences, offering viewers unparalleled immersion and interactivity. However, the technology's significant storage requirements and the computational complexity involved in generation and rendering currently limit its broader application. To close this gap, this paper presents Temporal Tri-Plane Radiance Fields (TeTriRF), a novel technology that significantly reduces the storage size for Free-Viewpoint Video (FVV) while maintaining low-cost generation and rendering. TeTriRF introduces a hybrid representation with tri-planes and voxel grids to support scaling up to long-duration sequences and scenes with complex motions or rapid changes. We propose a group training scheme tailored to achieving high training efficiency and yielding temporally consistent, low-entropy scene representations on feature domain. Leveraging these properties of the representations, we introduce a compression pipeline with off-the-shelf video codecs, achieving an order of magnitude less storage size compared to the state-of-the-art. Our experiments demonstrate that TeTriRF can achieve competitive quality with a higher compression rate.

The field of generative image inpainting and object insertion has made significant progress with the recent advent of latent diffusion models. Utilizing a precise object mask can greatly enhance these applications. However, due to the challenges users encounter in creating high-fidelity masks, there is a tendency for these methods to rely on more coarse masks (e.g., bounding box) for these applications. This results in limited control and compromised background content preservation. To overcome these limitations, we introduce SmartMask, which allows any novice user to create detailed masks for precise object insertion. Combined with a ControlNet-Inpaint model, our experiments demonstrate that SmartMask achieves superior object insertion quality, preserving the background content more effectively than previous methods. Notably, unlike prior works the proposed approach can also be used even without user-mask guidance, which allows it to perform mask-free object insertion at diverse positions and scales. Furthermore, we find that when used iteratively with a novel instruction-tuning based planning model, SmartMask can be used to design detailed layouts from scratch. As compared with user-scribble based layout design, we observe that SmartMask allows for better quality outputs with layout-to-image generation methods.

Recent advancements in diffusion-based models have demonstrated significant success in generating images from text. However, video editing models have not yet reached the same level of visual quality and user control. To address this, we introduce RAVE, a zero-shot video editing method that leverages pre-trained text-to-image diffusion models without additional training. RAVE takes an input video and a text prompt to produce high-quality videos while preserving the original motion and semantic structure. It employs a novel noise shuffling strategy, leveraging spatio-temporal interactions between frames, to produce temporally consistent videos faster than existing methods. It is also efficient in terms of memory requirements, allowing it to handle longer videos. RAVE is capable of a wide range of edits, from local attribute modifications to shape transformations. In order to demonstrate the versatility of RAVE, we create a comprehensive video evaluation dataset ranging from object-focused scenes to complex human activities like dancing and typing, and dynamic scenes featuring swimming fish and boats. Our qualitative and quantitative experiments highlight the effectiveness of RAVE in diverse video editing scenarios compared to existing methods. Our code, dataset and videos can be found in our supplementary materials.

The recent advancements in text-to-3D generation mark a significant milestone in generative models, unlocking new possibilities for creating imaginative 3D assets across various real-world scenarios. While recent advancements in text-to-3D generation have shown promise, they often fall short in rendering detailed and high-quality 3D models. This problem is especially prevalent as many methods base themselves on Score Distillation Sampling (SDS). This paper identifies a notable deficiency in SDS, that it brings inconsistent and low-quality updating direction for the 3D model, causing the over-smoothing effect. To address this, we propose a novel approach called Interval Score Matching (ISM). ISM employs deterministic diffusing trajectories and utilizes interval-based score matching to counteract over-smoothing. Furthermore, we incorporate 3D Gaussian Splatting into our text-to-3D generation pipeline. Extensive experiments show that our model largely outperforms the state-of-the-art in quality and training efficiency.

Personalization has emerged as a prominent aspect within the field of generative AI, enabling the synthesis of individuals in diverse contexts and styles, while retaining high-fidelity to their identities. However, the process of personalization presents inherent challenges in terms of time and memory requirements. Fine-tuning each personalized model needs considerable GPU time investment, and storing a personalized model per subject can be demanding in terms of storage capacity.To overcome these challenges, we propose HyperDreamBooth—a hypernetwork capable of efficiently generating a small set of personalized weights from a single image of a person. By composing these weights into the diffusion model, coupled with fast finetuning, HyperDreamBooth can generate a person's face in various contexts and styles, with high subject details while also preserving the model's crucial knowledge of diverse styles and semantic modifications.Our method achieves personalization on faces in roughly 20 seconds, 25x faster than DreamBooth and 125x faster than Textual Inversion, using as few as one reference image, with the same quality and style diversity as DreamBooth. Also our method yields a model that is 10,000x smaller than a normal DreamBooth model.

Customized generation using diffusion models has made impressive progress in image generation, but remains unsatisfactory in the challenging video generation task, as it requires the controllability of both subjects and motions. To that end, we present DreamVideo, a novel approach to generating personalized videos from a few static images of the desired subject and a few videos of target motion. DreamVideo decouples this task into two stages, subject learning and motion learning, by leveraging a pre-trained video diffusion model. The subject learning aims to accurately capture the fine appearance of the subject from provided images, which is achieved by combining textual inversion and fine-tuning of our carefully designed identity adapter. In motion learning, we architect a motion adapter and fine-tune it on the given videos to effectively model the target motion pattern. Combining these two lightweight and efficient adapters allows for flexible customization of any subject with any motion. Extensive experimental results demonstrate the superior performance of our DreamVideo over the state-of-the-art methods for customized video generation. Our project page is at https://dreamvideo-t2v.github.io.

Dynamic human rendering from video sequences has achieved remarkable progress by formulating the rendering as a mapping from static poses to human images. However, existing methods focus on the human appearance reconstruction of every single frame while the temporal motion relations are not fully explored. In this paper, we propose a new 4D motion modeling paradigm, SurMo, that jointly models the temporal dynamics and human appearances in a unified framework with three key designs: 1) Surface-based motion encoding that models 4D human motions with an efficient compact surface-based triplane. It encodes both spatial and temporal motion relations on the dense surface manifold of a statistical body template, which inherits body topology priors for generalizable novel view synthesis with sparse training observations. 2) Physical motion decoding that is designed to encourage physical motion learning by decoding the motion triplane features at timestep t to predict both spatial derivatives and temporal derivatives at the next timestep t+1 in training stage. 3) 4D appearance decoding that renders the motion triplanes into images by an efficient volumetric surface-conditioned renderer that focuses on the rendering of body surfaces with motion learning conditioning. Extensive experiments validate the state-of-the-art performance of our new paradigm and illustrate the expressiveness of surface-based motion triplanes for rendering high-fidelity view-consistent humans with fast motions and even motion-dependent shadows. Our project page is at: https://taohuumd.github.io/projects/SurMo.

Existing text-to-image (T2I) diffusion models usually struggle in interpreting complex prompts, especially those with quantity, object-attribute binding, and multi-subject descriptions. In this work, we introduce a semantic panel as the middleware in decoding texts to images, supporting the generator to better follow instructions. The panel is obtained through arranging the visual concepts parsed from the input text by the aid of large language models, and then injected into the denoising network as a detailed control signal to complement the text condition. To facilitate text-to-panel learning, we come up with a carefully designed semantic formatting protocol, accompanied by a fully-automatic data preparation pipeline. Thanks to such a design, our approach, which we call Ranni, manages to enhance a pre-trained T2I generator regarding its textual controllability. More importantly, the introduction of the generative middleware brings a more convenient form of interaction (i.e., directly adjusting the elements in the panel or using language instructions) and further allows users to finely customize their generation, based on which we develop a practical system and showcase its potential in continuous generation and chatting-based editing.

We address the task of generating temporally consistent and physically plausible images of actions and object state transformations. Given an input image and a text prompt describing the targeted transformation, our generated images preserve the environment and transform objects in the initial image. Our contributions are threefold. First, we leverage a large body of instructional videos and automatically mine a dataset of triplets of consecutive frames corresponding to initial object states, actions, and resulting object transformations. Second, equipped with this data, we develop and train a conditioned diffusion model dubbed GenHowTo. Third, we evaluate GenHowTo on a variety of objects and actions and show superior performance compared to existing methods. In particular, we introduce a quantitative evaluation where GenHowTo achieves 88% and 74% on seen and unseen interaction categories, respectively, outperforming prior work by a large margin.

Diffusion-based text-to-video generation has witnessed impressive progress in the past year yet still falls behind text-to-image generation. One of the key reasons is the limited scale of publicly available data (e.g., 10M video-text pairs in WebVid10M vs. 5B image-text pairs in LAION), considering the high cost of video captioning. Instead, it could be far easier to collect unlabeled clips from video platforms like YouTube. Motivated by this, we come up with a novel text-to-video generation framework, termed TF-T2V, which can directly learn with text-free videos. The rationale behind is to separate the process of text decoding from that of temporal modeling. To this end, we employ a content branch and a motion branch, which are jointly optimized with weights shared. Following such a pipeline, we study the effect of doubling the scale of training set (i.e., video-only WebVid10M) with some randomly collected text-free videos and are encouraged to observe the performance improvement (FID from 9.67 to 8.19 and FVD from 484 to 441), demonstrating the scalability of our approach. We also find that our model could enjoy sustainable performance gain (FID from 8.19 to 7.64 and FVD from 441 to 366) after reintroducing some text labels for training. Finally, we validate the effectiveness and generalizability of our ideology on both native text-to-video generation and compositional video synthesis paradigms. Code and models will be made public.

Although diffusion models are rising as a powerful solution for blind face restoration, they are criticized for two problems: 1) slow training and inference speed, and 2) failure in preserving the original identity and fine-grained facial details. In this work, we propose WaveFace to solve the problems in the frequency domain, where low- and high-frequency components decomposed by wavelet transformation are considered individually to maximize authenticity as well as efficiency.The diffusion model is applied to recover the low-frequency component only, which presents general information of the original image but 1/16 in size. To preserve the original identity, the generation is conditioned on the low-frequency component of low-quality images at each denoising step. Meanwhile, high-frequency components at multiple decomposition levels are handled by an unified network, which recovers complex facial details in a single step. Evaluations on four benchmark datasets show that: 1) WaveFace outperforms state-of-the-art methods in the authenticity, especially in terms of identity preservation, and 2) authentic images are restored with the efficiency 10$\times$ faster than existing diffusion model-based BFR methods.

This work presents AnyDoor, a diffusion-based image generator with the power to teleport target objects to new scenes at user-specified locations with desired shapes in a harmonious way. Instead of tuning parameters for each object, our model is trained only once and effortlessly generalizes to diverse object-scene combinations at the inference stage. Such a challenging zero-shot setting requires an adequate characterization of a certain object. To this end, we complement the commonly used identity feature with detail features, which are carefully designed to maintain texture details yet allow versatile local variations (e.g., lighting, orientation, posture, etc.), supporting the object in favorably blending with different surroundings. We further propose to borrow knowledge from video datasets, where we can observe various forms (i.e., along the time axis) of a single object, leading to stronger model generalizability and robustness. Extensive experiments demonstrate the superiority of our approach over existing alternatives as well as its great potential in real-world applications, such as virtual try-on and object moving.

Diffusion models have revolutionized image generation in recent years, yet they are still limited to a few sizes and aspect ratios. We propose ElasticDiffusion, a novel training-free decoding method that enables pretrained text-to-image diffusion models to generate images with various sizes. ElasticDiffusion attempts to decouple the generation trajectory of a pretrained model into local and global signals. The local signal controls low-level pixel information and can be estimated on local patches, while the global signal is used to maintain overall structural consistency and is estimated with a reference image. We test our method on CelebA-HQ (faces) and LAION-COCO (objects/indoor/outdoor scenes). Our experiments and qualitative results show superior image coherence quality across aspect ratios compared to MultiDiffusion and the standard decoding strategy of Stable Diffusion.

Diffusion models generate high-quality images but require dozens of forward passes. We introduce Distribution Matching Distillation (DMD), a procedure to transform a diffusion model into a one-step image generator with minimal impact on image quality. We enforce the one-step image generator match the diffusion model at distribution level, by minimizing an approximate KL divergence whose gradient can be expressed as the difference between 2 score functions, one of the target distribution and the other of the synthetic distribution being produced by our one-step generator. The score functions are parameterized as two diffusion models trained separately on each distribution. Combined with a simple regression loss matching the large-scale structure of the multi-step diffusion outputs, our method outperforms all published few-step diffusion approaches, reaching 2.62 FID on ImageNet 64x64 and 11.49 FID on zero-shot COCO-30k, comparable to Stable Diffusion but orders of magnitude faster. Utilizing FP16 inference, our model can generate images at 20 FPS on modern hardware.

Diffusion models have recently achieved remarkable progress in generating realistic images. However, challenges remain in accurately understanding and synthesizing the layout requirements in the textual prompts. To align the generated image with layout instructions, we present a training-free layout calibration system SimM that intervenes in the generative process on the fly during inference time. Specifically, following a "check-locate-rectify" pipeline, the system first analyses the prompt to generate the target layout and compares it with the intermediate outputs to automatically detect errors. Then, by moving the located activations and making intra- and inter-map adjustments, the rectification process can be performed with negligible computational overhead. To evaluate SimM over a range of layout requirements, we present a benchmark SimMBench that compensates for the lack of superlative spatial relations in existing datasets. And both quantitative and qualitative results demonstrate the effectiveness of the proposed SimM in calibrating the layout inconsistencies.

Despite diffusion models having shown powerful abilities to generate photorealistic images, generating videos that are realistic and diverse still remains in its infancy. One of the key reasons is that current methods intertwine spatial content and temporal dynamics together, leading to a notably increased complexity of text-to-video generation (T2V). In this work, we propose HiGen, a diffusion model-based method that improves performance by decoupling the spatial and temporal factors of videos from two perspectives, i.e., structure level and content level. At the structure level, we decompose the T2V task into two steps, including spatial reasoning and temporal reasoning, using a unified denoiser. Specifically, we generate spatially coherent priors using text during spatial reasoning and then generate temporally coherent motions from these priors during temporal reasoning. At the content level, we extract two subtle cues from the content of the input video that can express motion and appearance changes, respectively. These two cues then guide the model's training for generating videos, enabling flexible content variations and enhancing temporal stability. Through the decoupled paradigm, HiGen can effectively reduce the complexity of this task and generate realistic videos with semantics accuracy and motion stability. Extensive experiments demonstrate the superior performance of HiGen over the state-of-the-art T2V methods. Our source codes and models will be released.

Realistic 3D human generation from text prompts is a desirable yet challenging task. Existing methods optimize 3D representations like mesh or neural fields via score distillation sampling (SDS), which suffers from inadequate fine details or excessive training time. In this paper, we propose an efficient yet effective framework, HumanGaussian, that generates high-quality 3D humans with fine-grained geometry and realistic appearance. Our key insight is that 3D Gaussian Splatting is an efficient renderer with periodic Gaussian shrinkage or growing, where such adaptive density control can be naturally guided by intrinsic human structures. Specifically, 1) we first propose a Structure-Aware SDS that simultaneously optimizes human appearance and geometry. The multi-modal score function from both RGB and depth space is leveraged to distill the Gaussian densification and pruning process. 2) Moreover, we devise an Annealed Negative Prompt Guidance by decomposing SDS into a noisier generative score and a cleaner classifier score, which well addresses the over-saturation issue. The floating artifacts are further eliminated based on Gaussian size in a prune-only phase to enhance generation smoothness. Extensive experiments demonstrate the superior efficiency and competitive quality of our framework, rendering vivid 3D humans under diverse scenarios.

We introduce WonderJourney, a modular framework for perpetual 3D scene generation. Unlike prior work on view generation that focuses on a single type of scenes, we start at any user-provided location (by a text description or an image), and generate a journey through a long sequence of diverse yet coherently connected 3D scenes. We leverage an LLM to generate textual descriptions of the scenes in this journey, a text-driven point cloud generation pipeline to make a compelling and coherent sequence of 3D scenes, and a large VLM to verify the generated scenes. We show compelling, diverse visual results across various scene types and styles, forming imaginary “wonderjourneys”. Project website: https://kovenyu.com/WonderJourney/.

Diffusion Models (DMs) have emerged as powerful generative models with unprecedented image generation capability. These models are widely used for data augmentation and creative applications. However, DMs reflect the biases present in the training datasets. This is especially concerning in the context of faces, where the DM prefers one demographic subgroup vs. others (eg. female vs male). In this work, we present a method for debiasing DMs without relying on additional data or model retraining. Specifically, we propose \textbf{Distribution Guidance}, which enforces the generated images to follow the \underline{prescribed attribute distribution}. To realize this, we build on the key insight that the latent features of denoising UNet hold rich demographic semantics, and the same can be leveraged to guide debiased generation. We train \textbf{Attribute Distribution Predictor} (ADP) - a linear head that maps the latent features to the distribution of attributes. ADP is trained with pseudo labels generated from existing attribute classifiers. The proposed Distribution Guidance with ADP enables us to do fair generation. Our method reduces bias across single/multiple attributes and outperforms the baseline by a significant margin. Further, we present a downstream task of training a fair attribute classifier by rebalancing the training set with our generated data.

Advances in image diffusion models have recently led to notable improvements in the generation of high-quality images. In combination with Neural Radiance Fields (NeRFs), they enabled new opportunities in 3D generation. However, most generative 3D approaches are object-centric and applying them to editing existing photorealistic scenes is not trivial. We propose SIGNeRF, a novel approach for fast and controllable NeRF scene editing and scene-integrated object generation. A new generative update strategy ensures 3D consistency across the edited images, without requiring iterative optimization. We find that depth-conditioned diffusion models inherently possess the capability to generate 3D consistent views by requesting a grid of images instead of single views. Based on these insights, we introduce a multi-view reference sheet of modified images. Our method updates an image collection consistently based on the reference sheet and refines the original NeRF with the newly generated image set in one go. By exploiting the depth conditioning mechanism of the image diffusion model, we gain fine control over the spatial location of the edit and enforce shape guidance by a selected region or an external mesh.

Text-driven video generation witnesses rapid progress. However, merely using text prompts is not enough to depict the desired subject appearance that accurately aligns with users’ intents, especially for customized content creation. In this paper, we study the task of video generation with image prompts, which provide more accurate and direct content control beyond the text prompts. Specifically, we propose a feed-forward framework VideoBooth, with two dedicated designs: 1) We propose to embed image prompts in a coarse-to-fine manner. Coarse visual embeddings from image encoder provide high-level encodings of image prompts, while fine visual embeddings from the proposed attention injection module provide multi-scale and detailed encoding of image prompts. These two complementary embeddings can faithfully capture the desired appearance. 2) In the attention injection module at fine level, multi-scale image prompts are fed into different cross-frame attention layers as additional keys and values. This extra spatial information refines the details in the first frame and then it is propagated to the remaining frames, which maintains temporal consistency. Extensive experiments demonstrate that VideoBooth achieves state-of-the-art performance in generating customized high-quality videos with subjects specified in image prompts. Notably, VideoBooth is a generalizable framework where a single model works for a wide range of image prompts with only feed-forward passes.

We present a method to generate full-body selfies from photographs originally taken at arms length. Because self-captured photos are typically taken close up, they have limited field of view and exaggerated perspective that distorts facial shapes. We instead seek to generate the photo some one else would take of you from a few feet away. Our approach takes as input four selfies of your face and body, a background image, and generates a full-body selfie in a desired target pose. We introduce a novel diffusion-based approach to combine all of this information into high-quality, well-composed photos of you with the desired pose and background.

In this paper, we present CCEdit, a versatile generative video editing framework based on diffusion models. Our approach employs a novel trident network structure that separates structure and appearance control, ensuring precise and creative editing capabilities. Utilizing the foundational ControlNet architecture, we maintain the structural integrity of the video during editing. The incorporation of an additional appearance branch enables users to exert fine-grained control over the edited key frame. These two side branches seamlessly integrate into the main branch, which is constructed upon existing text-to-image (T2I) generation models, through learnable temporal layers. The versatility of our framework is demonstrated through a diverse range of choices in both structure representations and personalized T2I models. To facilitate comprehensive evaluation, we introduce the BalanceCC benchmark dataset, comprising 100 videos and 4 target prompts for each video. Our extensive user studies compare CCEdit with eight state-of-the-art video editing methods. The outcomes demonstrate CCEdit's substantial superiority over all other methods, affirming its exceptional editing capability.

In the evolving landscape of digital media and video production, the precise manipulation and reproduction of visual elements like camera movements and character actions are highly desired. Existing SLAM methods face limitations in dynamic scenes and human pose estimation often focuses on 2D projections, neglecting 3D statuses. To address these issues, we first introduce a reverse filming behavior estimation technique. It optimizes camera trajectories by leveraging NeRF as a differentiable renderer and refining SMPL tracks. We then introduce a cinematic transfer pipeline that is able to transfer various shot types to a new 2D video or a 3D virtual environment. The incorporation of 3D engine workflow enables superior rendering and control abilities, which also achieves a higher rating in the user study.

In subject-driven text-to-image synthesis, the synthesis process tends to be heavily influenced by the reference images provided by users, often overlooking crucial attributes detailed in the text prompt. In this work, we propose Subject-Agnostic Guidance (SAG), a simple yet effective solution to remedy the problem. We show that through constructing a subject-agnostic condition and applying our proposed dual classifier-free guidance, one could obtain outputs consistent with both the given subjects and input text prompts. We validate the efficacy of our approach through both optimization-based and encoder-based methods. Additionally, we demonstrate its applicability in second-order customization methods, where an encoder-based model is fine-tuned with DreamBooth. Our approach is conceptually simple and requires only minimal code modifications, but leads to substantial quality improvements, as evidenced by our evaluations and user studies.

Point-based interactive editing serves as an essential tool to complement the controllability of existing generative models. A concurrent work, DragDiffusion, updates the diffusion latent map in response to user inputs, causing global latent map alterations. This results in imprecise preservation of the original content and unsuccessful editing due to gradient vanishing. In contrast, we present DragNoise, offering robust and accelerated editing without retracing the latent map. The core rationale of DragNoise lies in utilizing the predicted noise output of each U-Net as a semantic editor. This approach is grounded in two critical observations: firstly, the bottleneck features of U-Net inherently possess semantically rich features ideal for interactive editing; secondly, high-level semantics, established early in the denoising process, show minimal variation in subsequent stages. Leveraging these insights, DragNoise edits diffusion semantics in a single denoising step and efficiently propagates these changes, ensuring stability and efficiency in diffusion editing. Comparative experiments reveal that DragNoise achieves superior control and semantic retention, reducing optimization time by over 50% compared to DragDiffusion.

Current controls over diffusion models (e.g., through text or ControlNet) for image generation fall short in recognizing abstract, continuous attributes like illumination direction or non-rigid shape change. In this paper, we present an approach for allowing users of text-to-image models to have fine-grained control of several attributes in an image. We do this by engineering special sets of input tokens that can be transformed in a continuous manner -- we call them \textbf{Continuous 3D Words}. These attributes can, for example, be represented as sliders and applied jointly with text prompts for fine-grained control over image generation. Given only a single mesh and a rendering engine, we show that our approach can be adopted to provide continuous user control over several 3D-aware attributes, including time-of-day illumination, bird wing orientation, dollyzoom effect, and object poses. Our method is capable of conditioning image creation with multiple Continuous 3D Words and text descriptions simultaneously while adding no overhead to the generative process.

Generative Adversarial Networks (GANs) significantly advanced image generation but their performance heavily depends on abundant training data. In scenarios with limited data, GANs often struggle with discriminator overfitting and unstable training. Batch Normalization (BN), despite being known for enhancing generalization and training stability, has rarely been used in the discriminator of Data-Efficient GANs. Our work addresses this gap by identifying a critical flaw in BN: the tendency for gradient explosion during the centering and scaling steps. To tackle this issue, we present CHAIN (lipsCHitz continuity constrAIned Normalization), which replaces the conventional centering step with zero-mean regularization and integrates a Lipschitz continuity constraint in the scaling step. CHAIN further enhances GAN training by adaptively interpolating the normalized and unnormalized features, effectively avoiding discriminator overfitting. Our theoretical analysis firmly establishes CHAIN's effectiveness in reducing gradients in latent features and weights, leading to improved stability and generalization in GAN training. Empirical evidence supports our theory. CHAIN achieves state-of-the-art results in data-limited scenarios on CIFAR-10/100 and ImageNet, and in five low-shot and seven high-resolution few-shot image datasets.

Generating novel views of an object from a single image is a challenging task. It requires an understanding of the underlying 3D structure of the object from an image and rendering high-quality, spatially consistent new views. While recent methods for view synthesis based on diffusion have shown great progress, achieving consistency among various view estimates and at the same time abiding by the desired camera pose remains a critical problem yet to be solved. In this work, we demonstrate a strikingly simple method, where we utilize a pre-trained video diffusion model to solve this problem. Our key idea is that synthesizing a novel view could be reformulated as synthesizing a video of a camera going around the object of interest---a scanning video---which then allows us to leverage the powerful priors that a video diffusion model would have learned. Thus, to perform novel-view synthesis, we create a smooth camera trajectory to the target view that we wish to render, and denoise using both a view-conditioned diffusion model and a video diffusion model. By doing so, we obtain a highly consistent novel view synthesis, outperforming the state of the art.

Personalized text-to-image generation models empower users to create images depicting their individual possessions in diverse scenes, finding applications in various domains. To achieve the personalization capability, existing methods rely on finetuning a text-to-image foundation model on a user's custom dataset, which can be nontrivial for general users, resource-intensive, and time-consuming. Despite attempts at developing finetuning-free methods, their generation quality is much lower compared to their finetuning counterparts. In this paper, we propose Joint-Image Diffusion (JeDi), an effective technique for learning a finetuning-free personalization model. Our key idea is to learn the joint distribution of multiple related text-image pairs that share a common subject. To facilitate the learning, we propose a scalable synthetic dataset generation technique. Once trained, our model enables fast and easy personalization at test time by simply using the reference images as inputs during the sampling process. Our approach does not require any expensive optimization process or additional modules, and can faithfully preserve the identity represented by any number of reference images. Experimental results show that our model achieves state-of-the-art generation quality, both quantitatively and qualitatively, significantly outperforming the prior finetuning-free personalization baselines.

In recent times, the generation of 3D assets from text prompts has shown impressive results. Both 2D and 3D diffusion models can help generate decent 3D objects based on prompts. 3D diffusion models have good 3D consistency, but their quality and generalization are limited as trainable 3D data is expensive and hard to obtain. 2D diffusion models enjoy strong abilities of generalization and fine generation, but 3D consistency is hard to guarantee. This paper attempts to bridge the power from the two types of diffusion models via the recent explicit and efficient 3D Gaussian splatting representation. A fast 3D object generation framework, named as GaussianDreamer, is proposed, where the 3D diffusion model provides priors for initialization and the 2D diffusion model enriches the geometry and appearance. Operations of noisy point growing and color perturbation are introduced to enhance the initialized Gaussians. Our GaussianDreamer can generate a high-quality 3D instance or 3D avatar within 15 minutes on one GPU, much faster than previous methods, while the generated instances can be directly rendered in real time. Demos and code are available at https://taoranyi.com/gaussiandreamer/.

The quality of the prompts provided to text-to-image diffusion models determines how faithful the generated content is to the user's intent, often requiring `prompt engineering'.To harness visual concepts from target images without prompt engineering, current approaches largely rely on embedding inversion by optimizing and then mapping them to pseudo-tokens.However, working with such high-dimensional vector representations is challenging because they lack semantics and interpretability, and only allow simple vector operations when using them.Instead, this work focuses on inverting the diffusion model to obtain interpretable language prompts directly. The challenge of doing this lies in the fact that the resulting optimization problem is fundamentally discrete and the space of prompts is exponentially large; this makes using standard optimization techniques, such as stochastic gradient descent, difficult.To this end, we utilize a delayed projection scheme to optimize for prompts representative of the vocabulary space in the model.Further, we leverage the findings that different timesteps of the diffusion process cater to different levels of detail in an image. The later, noisy, timesteps of the forward diffusion process correspond to the semantic information, and therefore, prompt inversion in this range provides tokens representative of the image semantics. We show that our approach can identify semantically interpretable and meaningful prompts for a target image which can be used to synthesize diverse images with similar content.We further illustrate the application of the optimized prompts in evolutionary image generation and concept removal.

We present a Multi-Instance Generation (MIG) task, simultaneously generating multiple instances with diverse controls in one image. Given a set of predefined coordinates and their corresponding descriptions, the task is to ensure that generated instances are accurately at the designated locations and that all instances' attributes adhere to their corresponding description. This broadens the scope of current research on Single-instance generation, elevating it to a more versatile and practical dimension. Inspired by the idea of divide and conquer, we introduce an innovative approach named Multi-Instance Generation Controller (MIGC) to address the challenges of the MIG task. Initially, we break down the MIG task into several subtasks, each involving the shading of a single instance. To ensure precise shading for each instance, we introduce an instance enhancement attention mechanism. Lastly, we aggregate all the shaded instances to provide the necessary information for accurately generating multiple instances in stable diffusion (SD). To evaluate how well generation models perform on the MIG task, we provide a COCO-MIG benchmark along with an evaluation pipeline. Extensive experiments were conducted on the proposed COCO-MIG benchmark, as well as on various commonly used benchmarks. The evaluation results illustrate the exceptional control capabilities of our model in terms of quantity, position, attribute, and interaction. Code and demos will be released at https://migcproject.github.io/.

We are witnessing significant breakthroughs in the technology for generating 3D objects from text. Existing approaches either leverage large text-to-image models to optimize a 3D representation or train 3D generators on object-centric datasets. Generating entire scenes, however, remains very challenging as a scene contains multiple 3D objects, diverse and scattered. In this work, we introduce SceneWiz3D – a novel approach to synthesize high fidelity 3D scenes from text. We marry the locality of objects with globality of scenes by introducing a hybrid 3D representation – explicit for objects and implicit for scenes. Remarkably, an object, being represented explicitly, can be either generated from text using conventional text-to-3D approaches, or provided by users. To configure the layout of the scene and automatically place objects, we apply Particle Swarm Optimization technique during the distillation process. Furthermore, in the text-to-scene scenario it is difficult for certain parts of the scene (e.g., corners, occlusion) to receive multi-view supervision, leading to inferior geometry. To mitigate the lack of such supervision, we incorporate an RGBD panorama diffusion model, resulting in high quality geometry. Extensive evaluation supports that our approach achieves superior quality over previous approaches, enabling the generation of detailed and view-consistent 3D scenes.

Generating large-scale 3D scenes cannot simply apply existing 3D object synthesis technique since 3D scenes usually hold complex spatial configurations and consist of a number of objects at varying scales. We thus propose a practical and efficient 3D representation that incorporates an equivariant radiance field with the guidance of a bird's-eye view (BEV) map. Concretely, objects of synthesized 3D scenes could be easily manipulated through steering the corresponding BEV maps. Moreover, by adequately incorporating positional encoding and low-pass filters into the generator, the representation becomes equivariant to the given BEV map. Such equivariance allows us to produce large-scale, even infinite-scale, 3D scenes via synthesizing local scenes and then stitching them with smooth consistency. Extensive experiments on 3D scene datasets demonstrate the effectiveness of our approach. Code and models will be made publicly available.

Face personalization aims to insert specific faces, taken from images, into pretrained text-to-image diffusion models. However, it is still challenging for previous methods to preserve both the identity similarity and editability due to overfitting to training samples. In this paper, we propose Face2Diffusion (F2D) for high-editability face personalization. The core idea behind F2D is that removing identity-irrelevant information from the training pipeline prevents the overfitting problem and improves editability of encoded faces. F2D consists of the following three novel components: 1) Multi-scale identity encoder provides well-disentangled identity features while keeping the benefits of multi-scale information, which improves the diversity of camera poses. 2) Expression guidance disentangles face expressions from identities and improves the controllability of face expressions. 3) Class-guided denoising regularization encourages models to learn how faces should be denoised, which boosts the text-alignment of backgrounds. Extensive experiments on the FaceForensics++ dataset and diverse prompts demonstrate our method greatly improves the trade-off between the identity- and text-fidelity compared to previous state-of-the-art methods. Code is available at https://github.com/mapooon/Face2Diffusion.

To serve the intricate and varied demands of image editing, precise and flexible manipulation in image content is indispensable. Recently, Drag-based editing methods have gained impressive performance.However, these methods predominantly center on point dragging, resulting in two noteworthy drawbacks, namely ''miss tracking", where difficulties arise in accurately tracking the predetermined handle points, and ''ambiguous tracking", where tracked points are potentially positioned in wrong regions that closely resemble the handle points. To address the above issues, we propose FreeDrag, a feature dragging methodology designed to free the burden on point tracking. The FreeDrag incorporates two key designs, i.e., template feature via adaptive updating and line search with backtracking, the former improves the stability against drastic content change by elaborately controls feature updating scale after each dragging, while the latter alleviates the misguidance from similar points by actively restricting the search area in a line. These two technologies together contribute to a more stable semantic dragging with higher efficiency. Comprehensive experimental results substantiate that our approach significantly outperforms pre-existing methodologies, offering reliable point-based editing even in various complex scenarios.

Omnidirectional cameras are extensively used in various applications to provide a wide field of vision. However, they face a challenge in synthesizing novel views due to the inevitable presence of dynamic objects, including the photographer, in their wide field of view. In this paper, we introduce a new approach called Omnidirectional Local Radiance Fields (OmniLocalRF) that can render static-only scene views, removing and inpainting dynamic objects simultaneously. Our approach combines the principles of local radiance fields with the bidirectional optimization of omnidirectional rays. Our input is an omnidirectional video, and we evaluate the mutual observations of the entire angle between the previous and current frames. To reduce ghosting artifacts of dynamic objects and inpaint occlusions, we devise a multi-resolution motion mask prediction module. Unlike existing methods that primarily separate dynamic components through the temporal domain, our method uses multi-resolution neural feature planes for precise segmentation, which is more suitable for long 360-degree videos. Our experiments validate that OmniLocalRF outperforms existing methods in both qualitative and quantitative metrics, especially in scenarios with complex real-world scenes. In particular, our approach eliminates the need for manual interaction, such as drawing motion masks by hand and additional pose estimation, making it a highly effective and efficient solution.

We present DIRECT-3D, a diffusion-based 3D generative model for creating high-quality 3D assets (represented by Neural Radiance Fields) from text prompts. Unlike recent 3D generative models that rely on clean and well-aligned 3D data, limiting them to single or few-class generation, our model is directly trained on extensive noisy and unaligned `in-the-wild' 3D assets, mitigating the key challenge (i.e., data scarcity) in large-scale 3D generation. In particular, DIRECT-3D is a tri-plane diffusion model that integrates two innovations: 1) A novel learning framework where noisy data are filtered and aligned automatically during the training process. Specifically, after an initial warm-up phase using a small set of clean data, an iterative optimization is introduced in the diffusion process to explicitly estimate the 3D pose of objects and select beneficial data based on conditional density. 2) An efficient 3D representation that is achieved by disentangling object geometry and color features with two separate conditional diffusion models that are optimized hierarchically. Given a prompt input, our model generates high-quality, high-resolution, realistic, and complex 3D objects with accurate geometric details in seconds. We achieve state-of-the-art performance in both single-class generation and text-to-3D generation. We also demonstrate that DIRECT-3D can serve as a useful 3D geometric prior of objects, for example, to alleviate the well-known Janus problem in 2D-lifting methods such as DreamFusion. The code and models are available for research purposes at: https://github.com/qihao067/direct3d

Few-shot font generation (FFG) produces stylized font images with a limited number of reference samples, which can significantly reduce labor costs in manual font designs. Most existing FFG methods follow the style-content disentanglement paradigm and employ the Generative Adversarial Network (GAN) to generate target fonts by combining the decoupled content and style representations. The complicated structure and detailed style are simultaneously generated in those methods, which may be the sub-optimal solutions for FFG task. Inspired by most manual font design processes of expert designers, in this paper, we model font generation as a multi-stage generative process. Specifically, as the injected noise and the data distribution in diffusion models can be well-separated into different sub-spaces, we are able to incorporate the font transfer process into these models. Based on this observation, we generalize diffusion methods to model font generative process by separating the reverse diffusion process into three stages with different functions: The structure construction stage first generates the structure information for the target character based on the source image, and the font transfer stage subsequently transforms the source font to the target font. Finally, the font refinement stage enhances the appearances and local details of the target font images. Based on the above multi-stage generative process, we construct our font generation framework, named MSD-Font, with a dual-network approach to generate font images. The superior performance demonstrates the effectiveness of our model. The code is available at: https://github.com/fubinfb/MSD-Font .

The field of autonomous driving increasingly demands high-quality annotated training data. In this paper, we propose Panacea, an innovative approach to generate panoramic and controllable videos in driving scenarios, capable of yielding an unlimited numbers of diverse, annotated samples pivotal for autonomous driving advancements. Panacea addresses two critical challenges: 'Consistency' and 'Controllability.' Consistency ensures temporal and cross-view coherence, while Controllability ensures the alignment of generated content with corresponding annotations. Our approach integrates a novel 4D attention and a two-stage generation pipeline to maintain coherence, supplemented by the ControlNet framework for meticulous control by the Bird's-Eye-View (BEV) layouts. Extensive qualitative and quantitative evaluations of Panacea on the nuScenes dataset prove its effectiveness in generating high-quality multi-view driving-scene videos. This work notably propels the field of autonomous driving by effectively augmenting the training dataset used for advanced BEV perception techniques.

Panorama video recently attracts more interest in both study and application, courtesy of its immersive experience. Due to the expensive cost of capturing 360$^{\circ}$ panoramic videos, generating desirable panorama videos by prompts is urgently required. Lately, the emerging text-to-video (T2V) diffusion methods demonstrate notable effectiveness in standard video generation. However, due to the significant gap in content and motion patterns between panoramic and standard videos, these methods encounter challenges in yielding satisfactory 360$^{\circ}$ panoramic videos. In this paper, we propose a pipeline named 360-Degree Video Diffusion model (360DVD) for generating 360$^{\circ}$ panoramic videos based on the given prompts and motion conditions. Specifically, we introduce a lightweight 360-Adapter accompanied by 360 Enhancement Techniques to transform pre-trained T2V models for panorama video generation. We further propose a new panorama dataset named WEB360 consisting of panoramic video-text pairs for training 360DVD, addressing the absence of captioned panoramic video datasets. Extensive experiments demonstrate the superiority and effectiveness of 360DVD for panorama video generation. Our project page is at https://akaneqwq.github.io/360DVD/.

This paper addresses the challenge of learning a local visual pattern of an object from one image, and generating images depicting objects with that pattern.Learning a localized concept and placing it on an object in a target image is a nontrivial task, as the objects may have different orientations and shapes.Our approach builds upon recent advancements in visual concept learning. It involves acquiring a visual concept (e.g., an ornament) from a source image and subsequently applying it to an object (e.g., a chair) in a target image.Our key idea is to perform in-context concept learning, acquiring the local visual concept within the broader context of the objects they belong to. To localize the concept learning, we employ soft masks that contain both the concept within the mask and the surrounding image area. We demonstrate our approach through object generation within an image, showcasing plausible embedding of in-context learned concepts.We also introduce methods for directing acquired concepts to specific locations within target images, employing cross-attention mechanisms, and establishing correspondences between source and target objects. The effectiveness of our method is demonstrated through quantitative and qualitative experiments, along with comparisons against baseline techniques.

Despite the remarkable progress in image style transfer, formulating style in the context of art is inherently subjective and challenging. In contrast to existing methods, this study shows that vanilla diffusion models can directly extract style information and seamlessly integrate the generative prior into the content image without retraining. Specifically, we adopt dual denoising paths to represent content/style references in latent space and then guide the content image denoising process with style latent codes. We further reveal that the cross-attention mechanism in latent diffusion models tends to blend the content and style images, resulting in stylized outputs that deviate from the original content image. To overcome this limitation, we introduce a cross-attention reweighting strategy.Through theoretical analysis and experiments, we demonstrate the effectiveness and superiority of the diffusion-based $\underline{Z}$ero-shot $\underline{s}$tyle $\underline{t}$ransfer via $\underline{a}$ttention $\underline{r}$eweighting, $\mathcal{Z}$-STAR.

Most diffusion models assume that the reverse process adheres to a Gaussian distribution. However, this approximation has not been rigorously validated, especially at singularities, where $t=0$ and $t=1$. Improperly dealing with such singularities leads to an average brightness issue in applications, and limits the generation of images with extreme brightness or darkness. We primarily focus on tackling singularities from both theoretical and practical perspectives. Initially, we establish the error bounds for the reverse process approximation, and showcase its Gaussian characteristics at singularity time steps. Based on this theoretical insight, we confirm the singularity at $t=1$ is conditionally removable while it at $t=0$ is an inherent property. Upon these significant conclusions, we propose a novel plug-and-play module to address the initial singular time step sampling, which not only effectively resolves the average brightness issue for a wide range of diffusion models without extra training efforts, but also enhances their generation capability in achieving notable lower FID scores.

We present CosmicMan, a text-to-image foundation model specialized for generating high-fidelity human images. Unlike current general-purpose foundation models that are stuck in the dilemma of inferior quality and text-image misalignment for humans, CosmicMan enables generating photo-realistic human images with meticulous appearance, reasonable structure, and precise text-image alignment with detailed dense descriptions.At the heart of CosmicMan's success are the new reflections and perspectives on data and model:$(1)$ We found that data quality and a scalable data production flow are essential for the final results from trained models. Hence, we propose a new data production paradigm \textbf{Annotate Anyone}, which serves as a perpetual data flywheel to produce high-quality data with accurate yet cost-effective annotations over time. Based on this, we constructed a large-scale dataset CosmicMan-HQ 1.0, with $6$ Million high-quality real-world human images in a mean resolution of $1488\times 1255$, and attached with precise text annotations deriving from $115$ Million attributes in diverse granularities.$(2)$ We argue that a text-to-image foundation model specialized for humans must be pragmatic -- easy to integrate into down-streaming tasks while effective in producing high-quality human images. Hence, we propose to model the relationship between dense text descriptions and image pixels in a decomposed manner, and present Decomposed-Attention-Refocusing (Daring). Daring is a training framework that seamlessly decomposes the cross-attention features in the existing text-to-image diffusion model, and enforces attention refocusing without adding extra modules. Through Daring, we show that explicitly discretizing continuous text space into several basic groups that align with human body structure is the key to tackling the misalignment problem in a breeze.

In this paper, we target the adaptive source driven 3D scene editing task by proposing a CustomNeRF model that unifies a text description or a reference image as the editing prompt. However, obtaining desired editing results conformed with the editing prompt is nontrivial since there exist two significant challenges, including accurate editing of only foreground regions and multi-view consistency given a single-view reference image. To tackle the first challenge, we propose a Local-Global Iterative Editing (LGIE) training scheme that alternates between foreground region editing and full-image editing, aimed at foreground-only manipulation while preserving the background. For the second challenge, we also design a class-guided regularization that exploits class priors within the generation model to alleviate the inconsistency problem among different views in image-driven editing. Extensive experiments show that our CustomNeRF produces precise editing results under various real scenes for both text- and image-driven settings.

In this paper, we propose a novel virtual try-on from unconstrained designs (ucVTON) task to enable photorealistic synthesis of personalized composite clothing on input human images. Unlike prior arts constrained by specific input types, our method allows flexible specification of style (text or image) and texture (full garment, cropped sections, or texture patches) conditions. To address the entanglement challenge when using full garment images as conditions, we develop a two-stage pipeline with explicit disentanglement of style and texture. In the first stage, we generate a human parsing map reflecting the desired style conditioned on the input. In the second stage, we composite textures onto the parsing map areas based on the texture input. To represent complex and non-stationary textures that have never been achieved in previous fashion editing works, we first propose extracting hierarchical and balanced CLIP features and applying position encoding in VTON. Experiments demonstrate superior synthesis quality and personalization enabled by our method. The flexible control over style and texture mixing brings virtual try-on to a new level of user experience for online shopping and fashion design.

Recently, diffusion-based methods, like InstructPix2Pix (IP2P), have achieved effective instruction-based image editing, requiring only natural language instructions from the user. However, these methods often inadvertently alter unintended areas and struggle with multi-instruction editing, resulting in compromised outcomes. To address these issues, we introduce the Focus on Your Instruction (FoI), a method designed to ensure precise and harmonious editing across multiple instructions without extra training or test-time optimization. In the FoI, we primarily emphasize two aspects: (1) precisely extracting regions of interest for each instruction and (2) guiding the denoising process to concentrate within these regions of interest. For the first objective, we identify the implicit grounding capability of IP2P from the cross-attention between instruction and image, then develop an effective mask extraction method. For the second objective, we introduce a cross attention modulation module for rough isolation of target editing regions and unrelated regions. Additionally, we introduce a mask-guided disentangle sampling strategy to further ensure clear region isolation. Experimental results demonstrate that FoI surpasses existing methods in both quantitative and qualitative evaluations, especially excelling in multi-instruction editing task.

Despite the remarkable process of talking-head-based avatar-creating solutions, directly generating anchor-style videos with full-body motions remains challenging.In this study, we propose Make-Your-Anchor, a novel system necessitating only a one-minute video clip of an individual for training, subsequently enabling the automatic generation of anchor-style videos with precise torso and hand movements. Specifically, we finetune a proposed structure-guided diffusion model on input video to render 3D mesh conditions into human appearances.We adopt a two-stage training strategy for the diffusion model, effectively binding movements with specific appearances. To produce arbitrary long temporal video, we extend the 2D U-Net in the frame-wise diffusion model to a 3D style without additional training cost, and a simple yet effective batch-overlapped temporal denoising module is proposed to bypass the constraints on video length during inference.Finally, a novel identity-specific face enhancement module is introduced to improve the visual quality of facial regions in the output videos.Comparative experiments demonstrate the effectiveness and superiority of the system in terms of visual quality, temporal coherence, and identity preservation, outperforming SOTA diffusion/non-diffusion methods.

The field of image synthesis is currently flourishing due to the advancements in diffusion models. While diffusion models have been successful, their computational intensity has prompted the pursuit of more efficient alternatives. As a representative work, non-autoregressive Transformers (NATs) have been recognized for their rapid generation. However, a major drawback of these models is their inferior performance compared to diffusion models. In this paper, we aim to re-evaluate the full potential of NATs by revisiting the design of their training and inference strategies. Specifically, we identify the complexities in properly configuring these strategies and indicate the possible sub-optimality in existing heuristic-driven designs. Recognizing this, we propose to go beyond existing methods by directly solving the optimal strategies in an automatic framework. The resulting method, named AutoNAT, advances the performance boundaries of NATs notably, and is able to perform comparably with the latest diffusion models with a significantly reduced inference cost. The effectiveness of AutoNAT is comprehensively validated on four benchmark datasets, i.e., ImageNet-256 & 512, MS-COCO, and CC3M. Code and pre-trained models will be available at https://github.com/LeapLabTHU/ImprovedNAT.

Image-based virtual try-on is an increasingly important task for online shopping. It aims to synthesize images of a specific person wearing a specified garment. Diffusion model-based approaches have recently become popular, as they are excellent at image synthesis tasks. However, these approaches usually employ additional image encoders and rely on the cross-attention mechanism for texture transfer from the garment to the person image, which affects the try-on's efficiency and fidelity. To address these issues, we propose an Texture-Preserving Diffusion (TPD) model for virtual try-on, which enhances the fidelity of the results and introduces no additional image encoders. Accordingly, we make contributions from two aspects. First, we propose to concatenate the masked person and reference garment images along the spatial dimension and utilize the resulting image as the input for the diffusion model's denoising UNet. This enables the original self-attention layers contained in the diffusion model to achieve efficient and accurate texture transfer. Second, we propose a novel diffusion-based method that predicts a precise inpainting mask based on the person and reference garment images, further enhancing the reliability of the try-on results. In addition, we integrate mask prediction and image synthesis into a single compact model. The experimental results show that our approach can be applied to various try-on tasks, e.g., garment-to-person and person-to-person try-ons, and significantly outperforms state-of-the-art methods on popular VITON, VITON-HD databases. Code is available at \href{https://github.com/Gal4way/TPD}{\texttt{https://github.com/Gal4way/TPD}}.

Diffusion-based generative models have exhibited remarkable capability in the production of high-fidelity visual content such as images and videos. However, their performance is significantly contingent upon the quality of textual inputs, commonly referred to as "prompts". The process of traditional prompt engineering, while effective, necessitates empirical expertise and poses challenges for inexperienced users. In this paper, we introduce PromptCoT, an innovative enhancer that autonomously refines prompts for users. PromptCoT is designed based on the observation that prompts, which resemble the textual information of high-quality images in the training set, often lead to superior generation performance. Therefore, we fine-tune the pre-trained Large Language Models (LLM) using a curated text dataset that solely comprises descriptions of high-quality visual content. By doing so, the LLM can capture the distribution of high-quality training texts, enabling it to generate aligned continuations and revisions to boost the original texts. Nonetheless, one drawback of pre-trained LLMs is their tendency to generate extraneous or irrelevant information. We employ the Chain-of-Thought (CoT) mechanism to improve the alignment between the original text prompts and their refined versions. CoT can extract and amalgamate crucial information from the aligned continuation and revision, enabling reasonable inferences based on the contextual cues to produce a more comprehensive and nuanced final output. Considering computational efficiency, instead of allocating a dedicated LLM for prompt enhancement to each individual model or dataset, we integrate adapters that facilitate dataset-specific adaptation, leveraging a shared pre-trained LLM as the foundation for this process. With independent fine-tuning of these adapters, we can adapt PromptCoT to new datasets while minimally increasing training costs and memory usage. We evaluate the effectiveness of PromptCoT by assessing its performance on widely-used latent diffusion models for image and video generation. The results demonstrate significant improvements in key performance metrics.

Contemporary models for generating images show remarkable quality and versatility. Swayed by these advantages, the research community repurposes them to generate videos. Since video content is highly redundant, we argue that naively bringing advances of image models to the video generation domain reduces motion fidelity, visual quality and impairs scalability. In this work, we build Snap Video, a video-first model that systematically addresses these challenges. To do that, we first extend the EDM framework to take into account spatially and temporally redundant pixels and naturally support video generation. Second, we show that a U-Net—a workhorse behind image generation—scales poorly when generating videos, requiring significant computational overhead. Hence, we propose a new transformer-based architecture that trains 3.31 times faster than U-Nets (and is ~4.5 faster at inference). This allows us to efficiently train a text-to-video model with billions of parameters for the first time, reach state-of-the-art results on a number of benchmarks, and generate videos with substantially higher quality, temporal consistency, and motion complexity. The user studies showed that our model was favored by a large margin over the most recent methods.

In the current era of generative AI breakthroughs, generating panoramic scenes from a single input image remains a key challenge. Most existing methods use diffusion-based iterative or simultaneous multi-view inpainting. However, the lack of global scene layout priors leads to subpar outputs with duplicated objects (e.g., multiple beds in a bedroom) or requires time-consuming human text inputs for each view. We propose L-MAGIC, a novel method leveraging large language models for guidance while diffusing multiple coherent views of $360^\circ$ panoramic scenes. L-MAGIC harnesses pre-trained diffusion and language models without fine-tuning, ensuring zero-shot performance. The output quality is further enhanced by super-resolution and multi-view fusion techniques. Extensive experiments demonstrate that the resulting panoramic scenes feature better scene layouts and perspective view rendering quality compared to related works, with $>$$70\%$ preference in human evaluations.Combined with conditional diffusion models, L-MAGIC can accept various input modalities, including but not limited to text, depth maps, sketches, and colored scripts. Applying depth estimation further enables 3D point cloud generation and dynamic scene exploration with fluid camera motion. We will release the code upon acceptance.

Language has emerged as a natural interface for image editing. In this paper, we introduce a method for region-based image editing driven by textual prompts, without the need for user-provided masks or sketches. Specifically, our approach leverages an existing pretrained text-to-image model and introduces a bounding box generator to find the edit regions that are aligned with the textual prompts. We show that this simple approach enables flexible editing that is compatible with current image generation models, and is able to handle complex prompts featuring multiple objects, complex sentences or long paragraphs. We conduct an extensive user study to compare our method against state-of-the-art baseline methods. Experiments demonstrate our method's competitive performance in manipulating images with high fidelity and realism that align with the language descriptions provided.

Diffusion probabilistic models (DPMs) are a key component in modern generative models. DPM-solvers have achieved reduced latency and enhanced quality significantly, but have posed challenges to find the exact inverse (i.e., finding the initial noise from the given image). Here we investigate the exact inversions for DPM-solvers and propose algorithms to perform them when samples are generated by the first-order as well as higher-order DPM-solvers. For each explicit denoising step in DPM-solvers, we formulated the inversions using implicit methods such as gradient descent or forward step method to ensure the robustness to large classifier-free guidance unlike the prior approach using fixed-point iteration. Experimental results demonstrated that our proposed exact inversion methods significantly reduced the error of both image and noise reconstructions, greatly enhanced the ability to distinguish invisible watermarks and well prevented unintended background changes consistently during image editing.

This paper presents Instruct-Imagen, a model that tackles heterogeneous image generation tasks and generalizes across unseen tasks.We introduce multi-modal instruction for image generation, a task representation articulating a range of generation intents with precision.It uses natural language to amalgamate disparate modalities (e.g., text, edge, style, subject, \etc), such that abundant generation intents can be standardized in a uniform format.We then build Instruct-Imagen by fine-tuning a pre-trained text-to-image diffusion model with two stages. First, we adapt the model using the retrieval-augmented training, to enhance model's capabilities to ground its generation on external multi-modal context.Subsequently, we fine-tune the adapted model on diverse image generation tasks that requires vision-language understanding (e.g., subject-driven generation, etc.), each paired with a multi-modal instruction encapsulating the task's essence. Human evaluation on various image generation datasets reveals that Instruct-Imagen matches or surpasses prior task-specific models in-domain and demonstrates promising generalization to unseen and more complex tasks. Our evaluation suite will be made publicly available.

Given a single image of a 3D object, this paper proposes a novel method (named ConsistNet) that is able to generate multiple images of the same object, as if seen they are captured from different viewpoints, while the 3D (multi-view) consistencies among those multiple generated images are effectively exploited. Central to our method is a light-weight multi-view consistency block which enables information exchange across multiple single-view diffusion processes based on the underlying multi-view geometry principles. ConsistNet is an extension to the standard latent diffusion model, and consists of two sub-modules: (a) a view aggregation module that unprojects multi-view features into global 3D volumes and infer consistency, and (b) a ray aggregation module that samples and aggregate 3D consistent features back to each view to enforce consistency. Our approach departs from previous methods in multi-view image generation, in that it can be easily dropped-in pre-trained LDMs without requiring explicit pixel correspondences or depth prediction. Experiments show that our method effectively learns 3D consistency over a frozen Zero123-XL backbone and can generate 16 surrounding views of the object within 11 seconds on a single A100 GPU.

In this paper, we present a few-shot text-to-video framework, **LAMP**, which enables a text-to-image diffusion model to **L**earn **A** specific **M**otion **P**attern with 8 $\sim$16 videos on a single GPU. Unlike existing methods, which require a large number of training resources or learn motions that are precisely aligned with template videos, it achieves a trade-off between the degree of generation freedom and the resource costs for model training. Specifically, we design a motion-content decoupled pipeline that uses an off-the-shelf text-to-image model for content generation so that our tuned video diffusion model mainly focuses on motion learning. The well-developed text-to-image techniques can provide visually pleasing and diverse content as generation conditions, which highly improves video quality and generation freedom. To capture the features of temporal dimension, we expand the pre-trained 2D convolution layers of the T2I model to our novel temporal-spatial motion learning layers and modify the attention blocks to the temporal level. Additionally, we develop an effective inference trick, shared-noise sampling, which can improve the stability of videos without computational costs. Our method can also be flexibly applied to other tasks, e.g. real-world image animation and video editing. Extensive experiments demonstrate that LAMP can effectively learn the motion pattern on limited data and generate high-quality videos. The code and models are available at https://rq-wu.github.io/projects/LAMP.

General image fusion aims at integrating important information from multi-source images. However, due to the significant cross-task gap, the respective fusion mechanism varies considerably in practice, resulting in limited performance across subtasks. To handle this problem, we propose a novel task-customized mixture of adapters (TC-MoA) for general image fusion, adaptively prompting various fusion tasks in a unified model. We borrow the insight from the mixture of experts (MoE), taking the experts as efficient tuning adapters to prompt a pre-trained foundation model. The task-specific routing networks customize these adapters to extract task-specific information from different sources with dynamic dominant intensity, performing adaptive visual feature prompt fusion. These adapters are shared across different tasks and constrained by mutual information regularization, ensuring compatibility with different tasks while complementarity for multi-source images. Notably, our TC-MoA controls the dominant intensity bias for different fusion tasks, successfully unifying multiple fusion tasks in a single model. Extensive experiments show that TC-MoA outperforms the competing approaches in learning commonalities while retaining compatibility for general image fusion (multi-modal, multi-exposure, and multi-focus), and also demonstrating striking controllability on more generalization experiments. The code is available at https://github.com/YangSun22/TC-MoA.

In this paper, we delve into a novel aspect of learning novel diffusion conditions with datasets an order of magnitude smaller. The rationale behind our approach is the elimination of textual constraints during the few-shot learning process. To that end, we implement two optimization strategies. The first, prompt-free conditional learning, utilizes a prompt-free encoder derived from a pre-trained Stable Diffusion model. This strategy is designed to adapt new conditions to the diffusion process by minimizing the textual-visual correlation, thereby ensuring a more precise alignment between the generated content and the specified conditions. The second strategy entails condition-specific negative rectification, which addresses the inconsistencies typically brought about by Classifier-free guidance in few-shot training contexts. Our extensive experiments across a variety of condition modalities demonstrate the effectiveness and efficiency of our framework, yielding results comparable to those obtained with datasets a thousand times larger. Our codes are available at https://github.com/Yuyan9Yu/BeyondTextConstraint.

Existing one-shot 4D head synthesis methods usually learn from monocular videos with the aid of 3DMM reconstruction, yet the latter is evenly challenging which restricts them from reasonable 4D head synthesis. We present a method to learn one-shot 4D head synthesis via large-scale synthetic data. The key is to first learn a part-wise 4D generative model from monocular images via adversarial learning, to synthesize multi-view images of diverse identities and full motions as training data; then leverage a transformer-based animatable triplane reconstructor to learn 4D head reconstruction using the synthetic data. A novel learning strategy is enforced to enhance the generalizability to real images by disentangling the learning process of 3D reconstruction and reenactment. Experiments demonstrate our superiority over the prior art.

We present the pioneering Large Visual Motion Model~(LVMM), meticulously engineered to analyze the intrinsic dynamics encapsulated within real-world imagery. Our model, fortified with a wealth of prior knowledge extracted from billions of image pairs, demonstrates promising results in predicting a diverse spectrum of scene dynamics. As a result, it can infuse any generic image with authentic dynamic effects, enhancing its visual allure. For a more comprehensive view of our results, please visit our project page: \url{https://anonymous.4open.science/r/AGI-LVMM-7826}.

Directly generating scenes from satellite imagery offers exciting possibilities for integration into applications like games and map services. However, challenges arise from significant view changes and scene scale. Previous efforts mainly focused on image or video generation, lacking exploration into the adaptability of scene generation for arbitrary views. Existing 3D generation works either operate at the object level or are difficult to utilize the geometry obtained from satellite imagery. To overcome these limitations, we propose a novel architecture for direct 3D scene generation by introducing diffusion models into 3D sparse representations and combining them with neural rendering techniques. Specifically, our approach generates texture colors at the point level for a given geometry using a 3D diffusion model first, which is then transformed into a scene representation in a feed-forward manner. The representation can be utilized to render arbitrary views which would excel in both single-frame quality and inter-frame consistency. Experiments in two city-scale datasets show that our model demonstrates proficiency in generating photo-realistic street-view image sequences and cross-view urban scenes from satellite imagery.

Video and audio content creation serves as the core technique for the movie industry and professional users. Recently, existing diffusion-based methods tackle video and audio generation separately, which hinders users the technique transfer from academia to industry. In this work, we aim at filling the gap, with a carefully designed joint video and audio generation framework. We observe the powerful generation ability of off-the-shelf video or audio generation models. Thus, instead of training the giant models from scratch, we propose to bridge the existing strong models with a shared latent representation space. Specifically, we propose a multimodality latent aligner with the pre-trained ImageBind model. Our latent aligner shares a similar core as the classifier guidance that guides the diffusion denoising process during inference time. Through carefully designed optimization strategy and loss functions, we show the superior performance of our method on joint video-audio generation, visual-steered audio generation, and audio-steered visual generation tasks.

Recent advances in diffusion models have successfully enabled text-guided image inpainting.While it seems straightforward to extend such editing capability into the video domain, there have been fewer works regarding text-guided video inpainting.Given a video, a masked region at its initial frame, and an editing prompt, it requires a model to do infilling at each frame following the editing guidance while keeping the out-of-mask region intact.There are three main challenges in text-guided video inpainting: ($i$) temporal consistency of the edited video, ($ii$) supporting different inpainting types at different structural fidelity levels, and ($iii$) dealing with variable video length.To address these challenges, we introduce Any-Length Video Inpainting with Diffusion Model, dubbed as AVID.At its core, our model is equipped with effective motion modules and adjustable structure guidance, for fixed-length video inpainting. Building on top of that, we propose a novel Temporal MultiDiffusion sampling pipeline with a middle-frame attention guidance mechanism, facilitating the generation of videos with any desired duration. Our comprehensive experiments show our model can robustly deal with various inpainting types at different video duration ranges, with high quality.

We present a method that uses a text-to-image model to generate consistent content across multiple image scales, enabling extreme semantic zooms into a scene, e.g., ranging from a wide-angle landscape view of a forest to a macro shot of an insect sitting on one of the tree branches. This representation allows us to render continuously zooming videos, or explore different scales of the scene interactively. We achieve this through a joint multi-scale diffusion sampling approach that encourages consistency across different scales while preserving the integrity of each individual sampling process. Since each generated scale is guided by a different text prompt, our method enables deeper levels of zoom than traditional super-resolution methods that may struggle to create new contextual structure at vastly different scales. We compare our method qualitatively with alternative techniques in image super-resolution and outpainting, and show that our method is most effective at generating consistent multi-scale content.

Diffusion models have achieved great success in synthesizing high-quality images. However, generating high-resolution images with diffusion models is still challenging due to the enormous computational costs, resulting in a prohibitive latency for interactive applications. In this paper, we propose DistriFusion to tackle this problem by leveraging parallelism across multiple GPUs. Our method splits the model input into multiple patches and assigns each patch to a GPU. However, naively implementing such an algorithm breaks the interaction between patches and loses fidelity, while incorporating such an interaction will incur tremendous communication overhead. To overcome this dilemma, we observe the high similarity between the input from adjacent diffusion steps and propose Displaced Patch Parallelism, which takes advantage of the sequential nature of the diffusion process by reusing the pre-computed feature maps from the previous timestep to provide context for the current step. Therefore, our method supports asynchronous communication, which can be pipelined by computation. Extensive experiments show that our method can be applied to recent Stable Diffusion XL with no quality degradation and achieve up to a 6.1$\times$ speedup on eight NVIDIA A100s compared to one. Our code is publicly available at https://github.com/mit-han-lab/distrifuser.

We present Condition-Aware Neural Network (CAN), a new method for adding control to image generative models. In parallel to prior conditional control methods, CAN controls the image generation process by dynamically manipulating the weight of the neural network. This is achieved by introducing a condition-aware weight generation module that generates conditional weight for convolution/linear layers based on the input condition. We test CAN on class-conditional image generation on ImageNet and text-to-image generation on COCO. CAN consistently delivers significant improvements for diffusion transformer models, including DiT and UViT. In particular, CAN combined with EfficientViT (CaT) achieves 2.78 FID on ImageNet 512x512, surpassing DiT-XL/2 while requiring 52x fewer MACs per sampling step.

This paper unravels the potential of sketches for diffusion models, addressing the deceptive promise of direct sketch control in generative AI. We importantly democratise the process, enabling amateur sketches to generate precise images, living up to the commitment of "what you sketch is what you get". A pilot study underscores the necessity, revealing that deformities in existing models stem from spatial-conditioning. To rectify this, we propose an abstraction-aware framework, utilising a sketch adapter, adaptive time-step sampling, and discriminative guidance from a pre-trained fine-grained sketch-based image retrieval model, working synergistically to reinforce fine-grained sketch-photo association. Our approach operates seamlessly during inference without the need for textual prompts; a simple, rough sketch akin to what you and I can create suffices! We welcome everyone to examine results presented in the paper and its supplementary. Contributions include democratising sketch control, introducing an abstraction-aware framework, and leveraging discriminative guidance, validated through extensive experiments.

Recently, subject-driven generation has garnered significant interest due to its ability to personalize text-to-image generation. Typical works focus on learning the new subject's private attributes. However, an important fact has not been taken seriously that a subject is not an isolated new concept but should be a specialization of a certain category in the pre-trained model. This results in the subject failing to comprehensively inherit the attributes in its category, causing poor attribute-related generations. In this paper, motivated by object-oriented programming, we model the subject as a derived class whose base class is its semantic category. This modeling enables the subject to inherit public attributes from its category while learning its private attributes from the user-provided example. Specifically, we propose a plug-and-play method, Subject-Derived regularization (SuDe). It constructs the base-derived class modeling by constraining the subject-driven generated images to semantically belong to the subject's category. Extensive experiments under three baselines and two backbones on various subjects show that our SuDe enables imaginative attribute-related generations while maintaining subject fidelity. For the codes, please refer to \href{https://github.com/modelscope/facechain}{FaceChain}.

3D-aware GANs offer new capabilities for view synthesis while preserving the editing functionalities of their 2D counterparts. GAN inversion is a crucial step that seeks the latent code to reconstruct input images or videos, subsequently enabling diverse editing tasks through manipulation of this latent code. However, a model pre-trained on a particular dataset (e.g., FFHQ) often has difficulty reconstructing images with out-of-distribution (OOD) objects such as faces with heavy make-up or occluding objects. We address this issue by explicitly modeling OOD objects from the input in 3D-aware GANs. Our core idea is to represent the image using two individual neural radiance fields: one for the in-distribution content and the other for the out-of-distribution object. The final reconstruction is achieved by optimizing the composition of these two radiance fields with carefully designed regularization. We demonstrate that our explicit decomposition alleviates the inherent trade-off between reconstruction fidelity and editability. We evaluate reconstruction accuracy and editability of our method on challenging real face images and videos and showcase favorable results against other baselines.

Diffusion models have made significant strides in image generation, mastering tasks such as unconditional image synthesis, text-image translation, and image-to-image conversions. However, their capability falls short in the realm of video prediction, mainly because they treat videos as a collection of independent images, relying on external constraints such as temporal attention mechanism to enforce temporal coherence. In our paper, we introduce a novel model class, that treats video as a continuous multi-dimensional process rather than a series of discrete frames. Through extensive experimentation, we establish state-of-the-art performance in video prediction, validated on benchmark datasets including KTH, BAIR, Human3.6M and UCF101.

Current perceptive models heavily depend on resource-intensive datasets, prompting the need for innovative solutions. Leveraging recent advances in diffusion models, synthetic data, by constructing image inputs from various annotations, proves beneficial for downstream tasks. While prior methods have separately addressed generative and perceptive models, DetDiffusion, for the first time, harmonizes both, tackling the challenges in generating effective data for perceptive models. To enhance image generation with perceptive models, we introduce perception-aware loss (P.A. loss) through segmentation, improving both quality and controllability. To boost the performance of specific perceptive models, our method customizes data augmentation by extracting and utilizing perception-aware attribute (P.A. Attr) during generation. Experimental results from the object detection task highlight DetDiffusion's superior performance, establishing a new state-of-the-art in layout-guided generation. Furthermore, image syntheses from DetDiffusion can effectively augment training data, significantly enhancing downstream detection performance.

Diffusion models have demonstrated exceptional efficacy in various generative applications. While existing models focus on minimizing a weighted sum of denoising score matching losses for data distribution modeling, their training primarily emphasizes instance-level optimization, overlooking valuable structural information within each mini-batch, indicative of pair-wise relationships among samples. To address this limitation, we introduce Structure-guided Adversarial training of Diffusion Models (SADM). In this pioneering approach, we compel the model to learn manifold structures between samples in each training batch. To ensure the model captures authentic manifold structures in the data distribution, we advocate adversarial training of the diffusion generator against a novel structure discriminator in a minimax game, distinguishing real manifold structures from the generated ones. SADM substantially outperforms existing methods in image generation and cross-domain fine-tuning tasks across 12 datasets, establishing a new state-of-the-art FID of 1.58 and 2.11 on ImageNet for class-conditional image generation at resolutions of 256x256 and 512x512, respectively.

Speech-preserving facial expression manipulation (SPFEM) aims to modify facial emotions while meticulously maintaining the mouth animation associated with spoken content. Current works depend on inaccessible paired training samples for the person, where two aligned frames exhibit the same speech content yet differ in emotional expression, limiting the SPFEM applications in real-world scenarios. In this work, we discover that speakers who convey the same content with different emotions exhibit highly correlated local facial animations, providing valuable supervision for SPFEM. To capitalize on this insight, we propose a novel adaptive spatial coherent correlation learning (ASCCL) algorithm, which models the aforementioned correlation as an explicit metric and integrates the metric to supervise manipulating facial expression and meanwhile better preserving the facial animation of spoken contents. To this end, it first learns a spatial coherent correlation metric, ensuring the visual disparities of adjacent local regions of the image belonging to one emotion are similar to those of the corresponding counterpart of the image belonging to another emotion. Recognizing that visual disparities are not uniform across all regions, we have also crafted a disparity-aware adaptive strategy that prioritizes regions that present greater challenges. During SPFEM model training, we construct the adaptive spatial coherent correlation metric between corresponding local regions of the input and output images as addition loss to supervise the generation process. We conduct extensive experiments on variant datasets, and the results demonstrate the effectiveness of the proposed ASCCL algorithm.

Fréchet Video Distance (FVD), a prominent metric for evaluating video generation models, is known to conflict with human perception occasionally. In this paper, we aim to explore the extent of FVD's bias toward per-frame quality over temporal realism and identify its sources. We first quantify the FVD's sensitivity to the temporal axis by decoupling the frame and motion quality and find that the FVD increases only slightly with large temporal corruption. We then analyze the generated videos and show that via careful sampling from a large set of generated videos that do not contain motions, one can drastically decrease FVD without improving the temporal quality. Both studies suggest FVD's bias towards the quality of individual frames. We further observe that the bias can be attributed to the features extracted from a supervised video classifier trained on the content-biased dataset. We show that FVD with features extracted from the recent large-scale self-supervised video models is less biased toward image quality. Finally, we revisit a few real-world examples to validate our hypothesis.

The past few years have witnessed great success in the use of diffusion models (DMs) to generate high-fidelity images with the help of stochastic differential equations (SDEs).Nevertheless, a gap emerges in the model sampling trajectory constructed by reverse-SDE due to the accumulation of score estimation and discretization errors. This gap results in a residual in the generated images, adversely impacting the image quality.To remedy this, we propose a novel residual learning framework built upon a correction function.The optimized function enables to improve image quality via rectifying the sampling trajectory effectively.Importantly, our framework exhibits transferable residual correction ability, i.e., a correction function optimized for one pre-trained DM can also enhance the sampling trajectory constructed by other different DMs on the same dataset.Experimental results on four widely-used datasets demonstrate the effectiveness and transferable capability of our framework.

Large-scale diffusion generative models are greatly simplifying image, video and 3D asset creation from user-provided text prompts and images. However, the challenging problem of text-to-4D dynamic 3D scene generation with diffusion guidance remains largely unexplored. We propose Dream-in-4D, which features a novel two-stage approach for text-to-4D synthesis, leveraging (1) 3D and 2D diffusion guidance to effectively learn a high-quality static 3D asset in the first stage; (2) a deformable neural radiance field that explicitly disentangles the learned static asset from its deformation, preserving quality during motion learning; and (3) a multi-resolution feature grid for the deformation field with a displacement total variation loss to effectively learn motion with video diffusion guidance in the second stage. Through a user preference study, we demonstrate that our approach significantly advances image and motion quality, 3D consistency and text fidelity for text-to-4D generation compared to baseline approaches. Thanks to its motion-disentangled representation, Dream-in-4D can also be easily adapted for controllable generation where appearance is defined by one or multiple images, without the need to modify the motion learning stage. Thus, our method offers, for the first time, a unified approach for text-to-4D, image-to-4D and personalized 4D generation tasks.

Text-to-video generation aims to produce a video based on a given prompt. Recently, several commercial video models have been able to generate plausible videos with minimal noise, excellent details, and high aesthetic scores. However, these models rely on large-scale, well-filtered, high-quality videos that are not accessible to the community. Many existing research works, which train models using the low-quality WebVid-10M dataset, struggle to generate high-quality videos because the models are optimized to fit WebVid-10M. In this work, we explore the training scheme of video models extended from Stable Diffusion and investigate the feasibility of leveraging low-quality videos and synthesized high-quality images to obtain a high-quality video model. We first analyze the connection between the spatial and temporal modules of video models and the distribution shift to low-quality videos. We observe that full training of all modules results in a stronger coupling between spatial and temporal modules than only training temporal modules. Based on this stronger coupling, we shift the distribution to higher quality without motion degradation by finetuning spatial modules with high-quality images, resulting in a generic high-quality video model. Evaluations are conducted to demonstrate the superiority of the proposed method, particularly in picture quality, motion, and concept composition.

Face morphing is a problem in computer graphics with numerous artistic and forensic applications. It is challenging due to variations in pose, lighting, gender, and ethnicity. This task consists of a warping for feature alignment and a blending for a seamless transition between the warped images.We propose to leverage coord-based neural networks to represent such warpings and blendings of face images. During training, we exploit the smoothness and flexibility of such networks by combining energy functionals employed in classical approaches without discretizations. Additionally, our method is time-dependent, allowing a continuous warping/blending of the images.During morphing inference, we need both direct and inverse transformations of the time-dependent warping. The first (second) is responsible for warping the target (source) image into the source (target) image. Our neural warping stores those maps in a single network dismissing the need for inverting them.The results of our experiments indicate that our method is competitive with both classical and generative models under the lens of image quality and face-morphing detectors. Aesthetically, the resulting images present a seamless blending of diverse faces not yet usual in the literature.

It is well known that many open-released foundational diffusion models have difficulty in generating images that substantially depart from average brightness, despite such images being present in the training data.This is due to an inconsistency: while denoising starts from pure Gaussian noise during inference, the training noise schedule retains residual data even in the final timestep distribution, due to difficulties in numerical conditioning in mainstream formulation, leading to unintended bias during inference.To mitigate this issue, certain epsilon-prediction models are combined with an ad-hoc offset-noise methodology. In parallel, some contemporary models have adopted zero-terminal SNR noise schedules together with v-prediction, which necessitate major alterations to pre-trained models. However, such changes risk destabilizing a large multitude of community-driven applications anchored on these pre-trained models. In light of this, our investigation revisits the fundamental causes, leading to our proposal of an innovative and principled remedy, called One More Step (OMS). By integrating a compact network and incorporating an additional simple yet effective step during inference, OMS elevates image fidelity and harmonizes the dichotomy between training and inference, while preserving original model parameters. Once trained, various pre-trained diffusion models with the same latent domain can share the same OMS module.

We present VIDIM, a generative model for video interpolation, which creates short videos given a start and end frame. In order to achieve high fidelity and generate motions unseen in the input data, VIDIM uses cascaded diffusion models to first generate the target video at low resolution, and then generate the high-resolution video conditioned on the low-resolution generated video. We compare VIDIM to previous state-of-the-art methods on video interpolation, and demonstrate how such works fail in most settings where the underlying motion is complex, nonlinear, or ambiguous while VIDIM can easily handle such cases. We additionally demonstrate how classifier-free guidance on the start and end frame and conditioning the super-resolution model on the original high-resolution frames without additional parameters unlocks high-fidelity results. VIDIM is fast to sample from as it jointly denoises all the frames to be generated, requires less than a billion parameters per diffusion model to produce compelling results, and still enjoys scalability and improved quality at larger parameter counts. Please see our project page at vidim-interpolation.github.io.

We propose DiffSHEG, a Diffusion-based approach for Speech-driven Holistic 3D Expression and Gesture generation with arbitrary length. While previous works focused on co-speech gesture or expression generation individually, the joint generation of synchronized expressions and gestures remains barely explored. To address this, our diffusion-based co-speech motion generation transformer enables uni-directional information flow from expression to gesture, facilitating improved matching of joint expression-gesture distributions. Furthermore, we introduce an outpainting-based sampling strategy for arbitrary long sequence generation in diffusion models, offering flexibility and computational efficiency. Our method provides a practical solution that produces high-quality synchronized expression and gesture generation driven by speech. Evaluated on two public datasets, our approach achieves state-of-the-art performance both quantitatively and qualitatively. Additionally, a user study confirms the superiority of DiffSHEG over prior approaches. By enabling the real-time generation of expressive and synchronized motions, DiffSHEG showcases its potential for various applications in the development of digital humans and embodied agents.

The Diffusion model, a prevalent framework for image generation, encounters significant challenges in terms of broad applicability due to its extended inference times and substantial memory requirements. Efficient Post-training Quantization (PTQ) is pivotal for addressing these issues in traditional models. Different from traditional models, diffusion models heavily depend on the time-step $t$ to achieve satisfactory multi-round denoising. Usually, $t$ from the finite set $\\{1, \ldots, T\\}$ is encoded to a temporal feature by a few modules totally irrespective of the sampling data. However, existing PTQ methods do not optimize these modules separately. They adopt inappropriate reconstruction targets and complex calibration methods, resulting in a severe disturbance of the temporal feature and denoising trajectory, as well as a low compression efficiency. To solve these, we propose a Temporal Feature Maintenance Quantization (TFMQ) framework building upon a Temporal Information Block which is just related to the time-step $t$ and unrelated to the sampling data. Powered by the pioneering block design, we devise temporal information aware reconstruction (TIAR) and finite set calibration (FSC) to align the full-precision temporal features in a limited time. Equipped with the framework, we can maintain the most temporal information and ensure the end-to-end generation quality. Extensive experiments on various datasets and diffusion models prove our state-of-the-art results. Remarkably, our quantization approach, for the first time, achieves model performance nearly on par with the full-precision model under 4-bit weight quantization. Additionally, our method incurs almost no extra computational cost and accelerates quantization time by $2.0 \times$ on LSUN-Bedrooms $256 \times 256$ compared to previous works. Our code is publicly available at https://github.com/ModelTC/TFMQ-DM.

Diffusion models, emerging as powerful deep generative tools, excel in various applications. They operate through a two-steps process: introducing noise into training samples and then employing a model to convert random noise into new samples (e.g., images). However, their remarkable generative performance is hindered by slow training and sampling. This is due to the necessity of tracking extensive forward and reverse diffusion trajectories, and employing a large model with numerous parameters across multiple timesteps (i.e., noise levels).To tackle these challenges, we present a multi-stage framework inspired by our empirical findings. These observations indicate the advantages of employing distinct parameters tailored to each timestep while retaining universal parameters shared across all time steps. Our approach involves segmenting the time interval into multiple stages where we employ custom multi-decoder U-net architecture that blends time-dependent models with a universally shared encoder. Our framework enables the efficient distribution of computational resources and mitigates inter-stage interference, which substantially improves training efficiency. Extensive numerical experiments affirm the effectiveness of our framework, showcasing significant training and sampling efficiency enhancements on three state-of-the-art diffusion models, including large-scale latent diffusion models. Furthermore, our ablation studies illustrate the impact of two important components in our framework: (i) a novel time-step clustering algorithm for stage division, and (ii) an innovative multi-decoder U-net architecture, seamlessly integrating universal and customized hyperparameters.

Recent significant advances in text-to-image models unlock the possibility of training vision systems using synthetic images, potentially overcoming the difficulty of collecting curated data at scale. It is unclear, however, how these models behave at scale, as more synthetic data is added to the training set. In this paper we study the scaling laws of synthetic images generated by state of the art text-to-image models, for the training of supervised models: image classifiers with label supervision, and CLIP with language supervision. We identify several factors, including text prompts, classifier-free guidance scale, and types of text-to-image models, that significantly affect scaling behavior. After tuning these factors, we observe that synthetic images demonstrate a scaling trend similar to, but slightly less effective than, real images in CLIP training, while they significantly underperform in scaling when training supervised image classifiers. Our analysis indicates that the main reason for this underperformance is the inability of off-the-shelf text-to-image models to generate certain concepts, a limitation that significantly impairs the training of image classifiers. Our findings also suggest that scaling synthetic data can be particularly effective in scenarios such as: (1) when there is a limited supply of real images for a supervised problem (e.g., fewer than 0.5 million images in ImageNet), (2) when the evaluation dataset diverges significantly from the training data, indicating the out-of-distribution scenario, or (3) when synthetic data is used in conjunction with real images, as demonstrated in the training of CLIP models.

Diffusion models have made tremendous progress in text-driven image and video generation. Now text-to-image foundation models are widely applied to various downstream image synthesis tasks, such as controllable image generation and image editing, while downstream video synthesis tasks are less explored for several reasons. First, it requires huge memory and computation overhead to train a video generation foundation model. Even with video foundation models, additional costly training is still required for downstream video synthesis tasks. Second, although some works extend image diffusion models into videos in a training-free manner, temporal consistency cannot be well preserved. Finally, these adaption methods are specifically designed for one task and fail to generalize to different tasks. To mitigate these issues, we propose a training-free general-purpose video synthesis framework, coined as BIVDiff, via bridging specific image diffusion models and general text-to-video foundation diffusion models. Specifically, we first use a specific image diffusion model (e.g., ControlNet and Instruct Pix2Pix) for frame-wise video generation, then perform Mixed Inversion on the generated video, and finally input the inverted latents into the video diffusion models (e.g., VidRD and ZeroScope) for temporal smoothing. This decoupled framework enables flexible image model selection for different purposes with strong task generalization and high efficiency. To validate the effectiveness and general use of BIVDiff, we perform a wide range of video synthesis tasks, including controllable video generation, video editing, video inpainting, and outpainting.

Recent advances in generative AI have significantly enhanced image and video editing, particularly in the context of text prompt control. State-of-the-art approaches predominantly rely on diffusion models to accomplish these tasks. However, the computational demands of diffusion-based methods are substantial, often necessitating large-scale paired datasets for training, and therefore challenging the deployment in real applications. To address these issues, this paper breaks down the text-based video editing task into two stages. First, we leverage an pre-trained text-to-image diffusion model to simultaneously edit few keyframes in an zero-shot way. Second, we introduce an efficient model called MaskINT, which is built on non-autoregressive masked generative transformers and specializes in frame interpolation between the edited keyframes, using the structural guidance from intermediate frames. Experimental results suggest that our MaskINT achieves comparable performance with diffusion-based methodologies, while significantly improve the inference time. This research offers a practical solution for text-based video editing and showcases the potential of non-autoregressive masked generative transformers in this domain.

We propose the Pose Adapted Shape Learning (PASL) for large-pose face reenactment. The PASL framework consists of three modules, namely the Pose-Adapted face Encoder (PAE), the Cycle-consistent Shape Generator (CSG), and the Attention-Embedded Generator (AEG). Different from previous approaches that use a single face encoder for identity preservation, we propose multiple Pose-Adapted face Encodes (PAEs) to better preserve facial identity across large poses. Given a source face and a reference face, the CSG generates a recomposed shape that fuses the source identity and reference action in the shape space and meets the cycle consistency requirement. Taking the shape code and the source as inputs, the AEG learns the attention within the shape code and between the shape code and source style to enhance the generation of the desired target face. As existing benchmark datasets are inappropriate for evaluating large-pose face reenactment, we propose a scheme to compose large-pose face pairs and introduce the MPIE-LP (Large Pose) and VoxCeleb2-LP datasets as the new large-pose benchmarks. We compared our approach with state-of-the-art methods on MPIE-LP and VoxCeleb2-LP for large-pose performance and on VoxCeleb1 for the common scope of pose variation.

Reward finetuning has emerged as a promising approach to aligning foundation models with downstream objectives. Remarkable success has been achieved in the language domain by using reinforcement learning (RL) to maximize rewards that reflect human preference. However, in the vision domain, existing RL-based reward finetuning methods are limited by their instability in large-scale training, rendering them incapable of generalizing to complex, unseen prompts. In this paper, we propose Proximal Reward Difference Prediction (PRDP), enabling stable black-box reward finetuning for diffusion models for the first time on large-scale prompt datasets with over 100K prompts. Our key innovation is the Reward Difference Prediction (RDP) objective that has the same optimal solution as the RL objective while enjoying better training stability. Specifically, the RDP objective is a supervised regression objective that tasks the diffusion model with predicting the reward difference of generated image pairs from their denoising trajectories. We theoretically prove that the diffusion model that obtains perfect reward difference prediction is exactly the maximizer of the RL objective. We further develop an online algorithm with proximal updates to stably optimize the RDP objective. In experiments, we demonstrate that PRDP can match the reward maximization ability of well-established RL-based methods in small-scale training. Furthermore, through large-scale training on text prompts from the Human Preference Dataset v2 and the Pick-a-Pic v1 dataset, PRDP achieves superior generation quality on a diverse set of complex, unseen prompts whereas RL-based methods completely fail.

Despite advancements in text-to-image generation (T2I), prior methods often face text-image misalignment problems such as relation confusion in generated images. Existing solutions involve cross-attention manipulation for better compositional understanding or integrating large language models for improved layout planning. However, the inherent alignment capabilities of T2I models are still inadequate. By reviewing the link between generative and discriminative modeling, we posit that T2I models' discriminative abilities may reflect their text-image alignment proficiency during generation. In this light, we advocate bolstering the discriminative abilities of T2I models to achieve more precise text-to-image alignment for generation. We present a discriminative adapter built on T2I models to probe their discriminative abilities on two representative tasks and leverage discriminative fine-tuning to improve their text-image alignment. As a bonus of the discriminative adapter, a self-correction mechanism can leverage discriminative gradients to better align generated images to text prompts during inference. Comprehensive evaluations across three benchmark datasets, including both in-distribution and out-of-distribution scenarios, demonstrate our method's superior generation performance. Meanwhile, it achieves state-of-the-art discriminative performance on the two discriminative tasks compared to other generative models. The code is available at https://anonymous.4open.science/r/DPT-5535.

Movie trailers are an essential tool for promoting films and attracting audiences. However, the process of creating trailers can be time-consuming and expensive. To streamline this process, we propose an automatic trailer generation framework that generates plausible trailers from a full movie by automating shot selection and composition. Our approach draws inspiration from machine translation techniques and models the movies and trailers as sequences of shots, thus formulating the trailer generation problem as a sequence-to-sequence task. We introduce Trailer Generation Transformer (TGT), a deep-learning framework utilizing an encoder-decoder architecture. TGT movie encoder is tasked with contextualizing each movie shot representation via self-attention, while the autoregressive trailer decoder predicts the feature representation of the next trailer shot, accounting for the relevance of shots' temporal order in trailers. Our TGT significantly outperforms previous methods on a comprehensive suite of metrics.

Existing Blind image Super-Resolution (BSR) methods focus on estimating either kernel or degradation information, but have long overlooked the essential content details. In this paper, we propose a novel BSR approach, Content-aware Degradation-driven Transformer (CDFormer), to capture both degradation and content representations. However, low-resolution images cannot provide enough content details, and thus we introduce a diffusion-based module $CDFormer_{diff}$ to first learn Content Degradation Prior (CDP) in both low- and high-resolution images, and then approximate the real distribution given only low-resolution information. Moreover, we apply an adaptive SR network $CDFormer_{SR}$ that effectively utilizes CDP to refine features. Compared to previous diffusion-based SR methods, we treat the diffusion model as an estimator that can overcome the limitations of expensive sampling time and excessive diversity. Experiments show that CDFormer can outperform existing methods, establishing a new state-of-the-art performance on various benchmarks under blind settings. Codes and models will be available at https://github.com/I2-Multimedia-Lab/CDFormer.

Recent approaches such as ControlNet offer users fine-grained spatial control over text-to-image (T2I) diffusion models. However, auxiliary modules have to be trained for each type of spatial condition, model architecture, and checkpoint, putting them at odds with the diverse intents and preferences a human designer would like to convey to the AI models during the content creation process. In this work, we present FreeControl, a training-free approach for controllable T2I generation that supports multiple conditions, architectures, and checkpoints simultaneously. FreeControl designs structure guidance to facilitate the structure alignment with a guidance image, and appearance guidance to enable the appearance sharing between images generated using the same seed. Extensive qualitative and quantitative experiments demonstrate the superior performance of FreeControl across a variety of pre-trained T2I models. In particular, FreeControl facilitates convenient training-free control over many different architectures and checkpoints, allows the challenging input conditions on which most of the existing training-free methods fail, and achieves competitive synthesis quality with training-based approaches.

Text-to-image customization, which aims to synthesize text-driven images for the given subjects, has recently revolutionized content creation. Existing works follow the pseudo-word paradigm, i.e., represent the given subjects as pseudo-words and then compose them with the given text. However, the inherent entangled influence scope of pseudo-words with the given text results in a dual-optimum paradox, i.e., the similarity of the given subjects and the controllability of the given text could not be optimal simultaneously. We present $\textbf{RealCustom}$ that, for the first time, disentangles similarity from controllability by precisely limiting subject influence to relevant parts only, achieved by gradually narrowing $\textbf{real}$ text word from its general connotation to the specific subject and using its cross-attention to distinguish relevance. Specifically, RealCustom introduces a novel decoupled "train-inference" framework: (1) during training, RealCustom learns general alignment between visual conditions to original textual conditions by a novel adaptive scoring module to adaptively modulate influence quantity; (2) during inference, a novel adaptive mask guidance strategy is proposed to iteratively update the influence scope and influence quantity of the given subjects to gradually narrow the generation of the real text word. Comprehensive experiments demonstrate the superior real-time customization ability of RealCustom in the open domain, achieving both unprecedented similarity of the given subjects and controllability of the given text for the first time.

Diffusion models have made significant advances in generating high-quality images, but their application to video generation has remained challenging due to the complexity of temporal motion. Zero-shot video editing offers a solution by utilizing pre-trained image diffusion models to translate source videos into new ones. Nevertheless, existing methods struggle to maintain strict temporal consistency and efficient memory consumption. In this work, we propose a novel approach to enhance temporal consistency in generated videos by merging self-attention tokens across frames. By aligning and compressing temporally redundant tokens across frames, our method improves temporal coherence and reduces memory consumption in self-attention computations. The merging strategy matches and aligns tokens according to the temporal correspondence between frames, facilitating natural temporal consistency in generated video frames. To manage the complexity of video processing, we divide videos into chunks and develop intra-chunk local token merging and inter-chunk global token merging, ensuring both short-term video continuity and long-term content consistency. Our video editing approach seamlessly extends the advancements in image editing to video editing, rendering favorable results in temporal consistency over state-of-the-art methods.

While diffusion models have significantly advanced the quality of image generation, their capability to accurately and coherently render text within these images remains a substantial challenge. Conventional diffusion-based methods for scene text generation are typically limited by their reliance on an intermediate layout output. This dependency often results in a constrained diversity of text styles and fonts, an inherent limitation stemming from the deterministic nature of the layout generation phase. To address these challenges, this paper introduces SceneTextGen, a novel diffusion-based model specifically designed to circumvent the need for a predefined layout stage. By doing so, SceneTextGen facilitates a more natural and varied representation of text. The novelty of SceneTextGen lies in its integration of three key components: a character-level encoder for capturing detailed typographic properties, coupled with a character-level instance segmentation model and a word-level spotting model to address the issues of unwanted text generation and minor character inaccuracies. We validate the performance of our method by demonstrating improved character recognition rates on generated images across different public visual text datasets in comparison to both standard diffusion based methods and text specific methods.

In this paper, we present RStab, a novel framework for video stabilization that integrates 3D multi-frame fusion through volume rendering. Departing from conventional methods, we introduce a 3D multi-frame perspective to generate stabilized images, addressing the challenge of full-frame generation while preserving structure. The core of our RStab framework lies in Stabilized Rendering (SR), a volume rendering module, fusing multi-frame information in 3D space. Specifically, SR involves warping features and colors from multiple frames by projection, fusing them into descriptors to render the stabilized image. However, the precision of warped information depends on the projection accuracy, a factor significantly influenced by dynamic regions. In response, we introduce the Adaptive Ray Range (ARR) module to integrate depth priors, adaptively defining the sampling range for the projection process. Additionally, we propose Color Correction (CC) assisting geometric constraints with optical flow for accurate color aggregation. Thanks to the three modules, our RStab demonstrates superior performance compared with previous stabilizers in the field of view (FOV), image quality, and video stability across various datasets.

Recent advancements in dynamic neural radiance field methods have yielded remarkable outcomes. However, these approaches rely on the assumption of sharp input images. When faced with motion blur, existing dynamic NeRF methods often struggle to generate high-quality novel views. In this paper, we propose DyBluRF, a dynamic radiance field approach that synthesizes sharp novel views from a monocular video affected by motion blur. To account for motion blur in input images, we simultaneously capture the camera trajectory and object Discrete Cosine Transform (DCT) trajectories within the scene. Additionally, we employ a global cross-time rendering approach to ensure consistent temporal coherence across the entire scene. We curate a dataset comprising diverse dynamic scenes that are specifically tailored for our task. Experimental results on our dataset demonstrate that our method outperforms existing approaches in generating sharp novel views from motion-blurred inputs while maintaining spatial-temporal consistency of the scene.

This paper presents a video inversion approach for zero-shot video editing, which aims to model the input video with low-rank representation during the inversion process. The existing video editing methods usually apply the typical 2D DDIM inversion or na\"ive spatial-temporal DDIM inversion before editing, which leverages time-varying representation for each frame to derive noisy latent. Unlike most existing approaches, we propose a Spatial-Temporal Expectation-Maximization (STEM) inversion, which formulates the dense video feature under an expectation-maximization manner and iteratively estimates a more compact basis set to represent the whole video. Each frame applies the fixed and global representation for inversion, which is more friendly for temporal consistency during reconstruction and editing. Extensive qualitative and quantitative experiments demonstrate that our STEM inversion can achieve consistent improvement on two state-of-the-art video editing methods.

Current 3D-aware facial NeRF generation approaches control the facial appearance by text, lighting conditions or reference images, limiting precise manipulation of local facial regions and interactivity. Color stroke, a user-friendly and effective tool to depict appearance, is challenging to edit 3D faces because of the lack of texture, coarse geometry representation and detailed editing operations. To solve the above problems, we introduce StrokeFaceNeRF, a novel stroke-based method for editing facial NeRF appearance. In order to infer the missing texture and 3D geometry information, 2D edited stroke maps are firstly encoded into the EG3D's latent space, followed by a transformer-based editing module to achieve effective appearance changes while preserving the original geometry in editing regions. Notably, we design a novel geometry loss function to ensure surface density remains consistent during training.To further enhance the local manipulation accuracy, we propose a stereo fusion approach which lifts the 2D mask (inferred from strokes or drawn by users) into 3D mask volume,allowing explicit blending of the original and edited faces. Extensive experiments validate that the proposed method outperforms existing 2D and 3D methods in both editing reality and geometry retention.

Recently, diffusion models have made remarkable progress in text-to-image (T2I) generation, synthesizing images with high fidelity and diverse contents. Despite this advancement, latent space smoothness within diffusion models remains largely unexplored. Smooth latent spaces ensure that a perturbation on an input latent corresponds to a steady change in the output image. This property proves beneficial in downstream tasks, including image interpolation, inversion, and editing. In this work, we expose the non-smoothness of diffusion latent spaces by observing noticeable visual fluctuations resulting from minor latent variations. To tackle this issue, we propose Smooth Diffusion, a new category of diffusion models that can be simultaneously high-performing and smooth. Specifically, we introduce Step-wise Variation Regularization to enforce the proportion between the variations of an arbitrary input latent and that of the output image is a constant at any diffusion training step. In addition, we devise an interpolation standard deviation (ISTD) metric to effectively assess the latent space smoothness of a diffusion model. Extensive quantitative and qualitative experiments demonstrate that Smooth Diffusion stands out as a more desirable solution not only in T2I generation but also across various downstream tasks. Smooth Diffusion is implemented as a plug-and-play Smooth-LoRA to work with various community models. Code is available at https://github.com/SHI-Labs/Smooth-Diffusion.

The prevalent use of commercial and open-source diffusion models (DMs) for text-to-image generation prompts the risk mitigation to prevent undesired behaviors. Existing concept erasing methods in academia are all based on full parameter or specification-based fine-tuning, from which we observe following issues: 1) Generation alternation towards erosion: Parameter drift during target elimination causes alternations and potential deformations across all generations, even eroding other concepts at varying degrees, which is more evident with multi-concept erasing; 2) Transfer inability \& deployment inefficiency: Previous model-specific erasure impedes the flexible combination of concepts and the training-free transfer towards other models, resulting in linear cost growth as the deployment scenarios increase.To achieve non-invasive, precise, customizable and transferable elimination, we ground our erasing framework on one-dimensional adapters to erase multiple concepts from most of DMs at once across versatile erasing applications. The concept-SemiPermeable structure is injected as a Membrane (SPM) into any DM to learn targeted erasing, and meantime the alteration and erosion phenomenon is effectively minimized via a novel Latent Anchoring fine-tuning strategy. Once obtained, SPMs can be flexibly combined and plug-and-play for other DMs without specific re-tuning, enabling timely and efficient adaptation to diverse scenarios. During generation, our Facilitated Transport mechanism dynamically regulates the permeability of each SPM to respond to different input prompts, further minimizing the impact on other concepts. Quantitative and qualitative results across $\sim$40 concepts, 7 DMs and 4 erasing applications have demonstrated the superior erasing of SPM. Our code and pre-tuned SPMs will be available on the project https://***.

Diffusion models have demonstrated remarkable performance in image and video synthesis. However, scaling them to high-resolution inputs is challenging and requires restructuring the diffusion pipeline into multiple independent components, limiting scalability and complicating downstream applications. In this work, we study patch diffusion models (PDMs) --- a diffusion paradigm which models the distribution of patches, rather than whole inputs, keeping up to ${\approx}$0.7\% of the original pixels. This makes it very efficient during training and unlocks end-to-end optimization on high-resolution videos. We improve PDMs in two principled ways. First, to enforce consistency between patches, we develop \emph{deep context fusion} --- an architectural technique that propagates the context information from low-scale to high-scale patches in a hierarchical manner. Second, to accelerate training and inference, we propose \emph{adaptive computation}, which allocates more network capacity and computation towards coarse image details. The resulting model sets a new state-of-the-art FVD score of 66.32 and Inception Score of 87.68 in class-conditional video generation on UCF-101 $256^2$, surpassing recent methods by more than 100\%. Then, we show that it can be rapidly fine-tuned from a base $36\times 64$ low-resolution generator for high-resolution $64 \times 288 \times 512$ text-to-video synthesis. To the best of our knowledge, our model is the first diffusion-based architecture which is trained on such high resolutions entirely end-to-end. Project webpage: https://snap-research.github.io/hpdm.

Despite the extensive research on training generative adversarial networks (GANs) with limited training data, learning to generate images from long-tailed training distributions remains fairly unexplored. In the presence of imbalanced multi-class training data, GANs tend to favor classes with more samples, leading to the generation of low-quality and less diverse samples in tail classes. In this study, we aim to improve the training of class-conditional GANs with long-tailed data. We propose a straightforward yet effective method for knowledge sharing, allowing tail classes to borrow from the rich information from classes with more abundant training data. More concretely, we propose modifications to existing class-conditional GAN architectures to ensure that the lower-resolution layers of the generator are trained entirely unconditionally, while reserving class-conditional generation for the higher-resolution layers.Experiments on several long-tail benchmarks and GAN architectures demonstrate a significant improvement over existing methods in both the diversity and fidelity of the generated images. The code will be publicly released.

We present a method for large-mask pluralistic image inpainting based on the generative framework of discrete latent codes. Our method learns latent priors, discretized as tokens, by only performing computations at the visible locations of the image. This is realized by a restrictive partial encoder that predicts the token label for each visible block, a bidirectional transformer that infers the missing labels by only looking at these tokens, and a dedicated synthesis network that couples the tokens with the partial image priors to generate coherent and pluralistic complete image even under extreme mask settings. Experiments on public benchmarks validate our design choices as the proposed method outperforms strong baselines in both visual quality and diversity metrics.

The absence of real targets to guide the model training is one of the main problems with the makeup transfer task. Most existing methods tackle this problem by synthesizing pseudo ground truths (PGTs). However, the generated PGTs are often sub-optimal and their imprecision will eventually lead to performance degradation. To alleviate this issue, in this paper, we propose a novel Content-Style Decoupled Makeup Transfer (CSD-MT) method, which works in a purely unsupervised manner and thus eliminates the negative effects of generating PGTs. Specifically, based on the frequency characteristics analysis, we assume that the low-frequency (LF) component of a face image is more associated with its makeup style information, while the high-frequency (HF) component is more related to its content details. This assumption allows CSD-MT to decouple the content and makeup style information in each face image through the frequency decomposition. After that, CSD-MT realizes makeup transfer by maximizing the consistency of these two types of information between the transferred result and input images, respectively. Two newly designed loss functions are also introduced to further improve the transfer performance. Extensive quantitative and qualitative analyses show the effectiveness of our CSD-MT method.

Traditional 3D content creation tools empower users to bring their imagination to life by giving them direct control over a scene's geometry, appearance, motion, and camera path. Creating computer-generated videos, however, is a tedious manual process, which can be automated by emerging text-to-video diffusion models. Despite great promise, video diffusion models are difficult to control, hindering a user to apply their own creativity rather than amplifying it. To address this challenge, we present a novel approach that combines the controllability of dynamic 3D meshes with the expressivity and editability of emerging diffusion models. For this purpose, our approach takes an animated, low-fidelity rendered mesh as input and injects the ground truth correspondence information obtained from the dynamic mesh into various stages of a pre-trained text-to-image generation model to output high-quality and temporally consistent frames. We demonstrate our approach on various examples where motion can be obtained by animating rigged assets or changing the camera path.

Current diffusion-based video editing primarily focuses on structure-preserved editing by utilizing various dense correspondences to ensure temporal consistency and motion alignment. However, these approaches are often ineffective when the target edit involves a shape change. To embark on video editing with shape change, we explore customized video subject swapping in this work, where we aim to replace the main subject in a source video with a target subject having a distinct identity and potentially different shape. In contrast to previous methods that rely on dense correspondences, we introduce the VideoSwap framework that exploits semantic point correspondences, inspired by our observation that only a small number of semantic points are necessary to align the subject's motion trajectory and modify its shape. We also introduce various user-point interactions (\eg, removing points and dragging points) to address various semantic point correspondence. Extensive experiments demonstrate state-of-the-art video subject swapping results across a variety of real-world videos.

Vector-Quantized (VQ-based) generative models usually consist of two basic components, \textit{i.e.}, VQ tokenizers and generative transformers. Prior research focuses on improving the reconstruction fidelity of VQ tokenizers but rarely examines how the improvement in reconstruction affects the generation ability of generative transformers. In this paper, we find that improving the reconstruction fidelity of VQ tokenizers does not necessarily improve the generation. Instead, learning to compress semantic features within VQ tokenizers significantly improves generative transformers' ability to capture textures and structures. We thus highlight two competing objectives of VQ tokenizers for image synthesis: \textbf{semantic compression} and \textbf{details preservation}. Different from previous work that prioritizes better details preservation, we propose \textbf{Se}mantic-\textbf{Q}uantized GAN (SeQ-GAN) with two learning phases to balance the two objectives. In the first phase, we propose a semantic-enhanced perceptual loss for better semantic compression. In the second phase, we fix the encoder and codebook, but finetune the decoder to achieve better details preservation. Our proposed SeQ-GAN significantly improves VQ-based generative models for both unconditional and conditional image generation. Specifically, SeQ-GAN achieves a Fréchet Inception Distance (FID) of 6.25 and Inception Score (IS) of 140.9 on 256×256 ImageNet generation, a remarkable improvement over VIT-VQGAN, which obtains 11.2 FID and 97.2 IS.

Text-to-video (T2V) synthesis has gained increasing attention in the community, in which the recently emerged diffusion models (DMs) have promisingly shown stronger performance than the past approaches. While existing state-of-the-art DMs are competent to achieve high-resolution video generation, they may largely suffer from key limitations (e.g., action occurrence disorders, crude video motions) with respect to the intricate temporal dynamics modeling, one of the crux of video synthesis. In this work, we investigate strengthening the awareness of video dynamics for DMs, for high-quality T2V generation. Inspired by human intuition, we design an innovative dynamic scene manager (dubbed as Dysen) module, which includes (step-1) extracting from input text the key actions with proper time-order arrangement, (step-2) transforming the action schedules into the dynamic scene graph (DSG) representations, and (step-3) enriching the scenes in the DSG with sufficient and reasonable details. Taking advantage of the existing powerful LLMs (e.g., ChatGPT) via in-context learning, Dysen realizes (nearly) human-level temporal dynamics understanding. Finally, the resulting video DSG with rich action scene details is encoded as fine-grained spatio-temporal features, integrated into the backbone T2V DM for video generating. Experiments on popular T2V datasets suggest that our Dysen-VDM consistently outperforms prior arts with significant margins, especially in scenarios with complex actions.

Generalizable NeRF aims to synthesize novel views for unseen scenes. Common practices involve constructing variance-based cost volumes for geometry reconstruction and encoding 3D descriptors for decoding novel views. However, existing methods show limited generalization ability in challenging conditions due to inaccurate geometry, sub-optimal descriptors, and decoding strategies. We address these issues point by point. First, we find the variance-based cost volume exhibits failure patterns as the features of pixels corresponding to the same point can be inconsistent across different views due to occlusions or reflections. We introduce an Adaptive Cost Aggregation (ACA) approach to amplify the contribution of consistent pixel pairs and suppress inconsistent ones. Unlike previous methods that solely fuse 2D features into descriptors, our approach introduces a Spatial-View Aggregator (SVA) to incorporate 3D context into descriptors through spatial and inter-view interaction. When decoding the descriptors, we observe the two existing decoding strategies excel in different areas, which are complementary. A Consistency-Aware Fusion (CAF) strategy is introduced to leverage the advantages of both. We incorporate the above ACA, SVA, and CAF into a coarse-to-fine framework, termed Geometry-aware Reconstruction and Fusion-refined Rendering (GeFu). GeFu attains state-of-the-art performance across multiple datasets. Code will be released.

Despite recent progress in diffusion-based video editing, existing methods are limited to short-length videos due to the contradiction between long-range consistency and frame-wise editing. Prior attempts to address this challenge by introducing video-2D representations encounter significant difficulties with large-scale motion- and view-change videos, especially in human-centric scenarios. To overcome this, we propose to introduce the dynamic Neural Radiance Fields (NeRF) as the innovative video representation, where the editing can be performed in the 3D spaces and propagated to the entire video via the deformation field. To provide consistent and controllable editing, we propose the image-based video-NeRF editing pipeline with a set of innovative designs, including multi-view multi-pose Score Distillation Sampling (SDS) from both the 2D personalized diffusion prior and 3D diffusion prior, reconstruction losses, text-guided local parts super-resolution, and style transfer. Extensive experiments demonstrate that our method, dubbed as DynVideo-E, significantly outperforms SOTA approaches on two challenging datasets by a large margin of 50%~95% for human preference. Our code, data, all video comparisons, and compelling examples of free-viewpoint renderings of edited dynamic scenes are provided in supplementary materials and will be released.

Current subject-driven image generation methods encounter significant challenges in person-centric image generation. The reason is that they learn the semantic scene and person generation by fine-tuning a common pre-trained diffusion, which involves an irreconcilable training imbalance. Precisely, to generate realistic persons, they need to sufficiently tune the pre-trained model, which inevitably causes the model to forget the rich semantic scene prior and makes scene generation over-fit to the training data. Moreover, even with sufficient fine-tuning, these methods can still not generate high-fidelity persons since joint learning of the scene and person generation also lead to quality compromise. In this paper, we propose Face-diffuser, an effective collaborative generation pipeline to eliminate the above training imbalance and quality compromise. Specifically, we first develop two specialized pre-trained diffusion models, i.e., Text-driven Diffusion Model (TDM) and Subject-augmented Diffusion Model (SDM), for scene and person generation, respectively. The sampling process is divided into three sequential stages, i.e., semantic scene construction, subject-scene fusion, and subject enhancement. The first and last stages are performed by TDM and SDM respectively. The subject-scene fusion stage, that is the collaboration achieved through a novel and highly effective mechanism, Saliency-adaptive Noise Fusion (SNF). Specifically, it is based on our key observation that there exists a robust link between classifier-free guidance responses and the saliency of generated images. In each time step, SNF leverages the unique strengths of each model and allows for the spatial blending of predicted noises from both models automatically in a saliency-aware manner, all of which can be seamlessly integrated into the DDIM sampling process. Extensive experiments confirm the impressive effectiveness and robustness of the Face-diffuser in generating high-fidelity person images depicting multiple unseen persons with varying contexts.

Despite their exceptional generative abilities, large text-to-image diffusion models, much like skilled but careless artists, often struggle with accurately depicting visual relationships between objects. This issue, as we uncover through careful analysis, arises from a misaligned text encoder that struggles to interpret specific relationships and differentiate the logical order of associated objects. To resolve this, we introduce a novel task termed \textbf{Relation Rectification}, aiming to refine the model to accurately represent a given relationship it initially fails to generate. To address this, we propose an innovative solution utilizing a Heterogeneous Graph Convolutional Network (HGCN). It models the directional relationships between relation terms and corresponding objects within the input prompts. Specifically, we optimize the HGCN on a pair of prompts with identical relational words but reversed object orders, supplemented by a few reference images. The lightweight HGCN adjusts the text embeddings generated by the text encoder, ensuring the accurate reflection of the textual relation in the embedding space. Crucially, our method retains the parameters of the text encoder and diffusion model, preserving the model's robust performance on unrelated descriptions. We validated our approach on a newly curated dataset of diverse relational data, demonstrating both quantitative and qualitative enhancements in generating images with precise visual relations.

We present a new training-free method for 3D-aware object edits on images using pretrained text-to-image diffusion models. 3D edits, like translation, rotation and scale, are implemented by lifting the activations of the diffusion model to 3D using depth information. In this paper, we present our method, followed by results on both real and generated images, and a comparative user study to position our method with respect to relevant work. We further illustrate compelling 3D applications of our method such as object editing in scenes and camera movement.

This paper introduces LeftRefill, an innovative approach to efficiently harness large Text-to-Image (T2I) diffusion models for reference-guided image synthesis. As the name implies, LeftRefill horizontally stitches reference and target views together as a whole input. The reference image occupies the left side, while the target canvas is positioned on the right. Then, LeftRefill paints the right-side target canvas based on the left-side reference and specific task instructions. Such a task formulation shares some similarities with contextual inpainting, akin to the actions of a human painter. This novel formulation efficiently learns both structural and textured correspondence between reference and target without other image encoders or adapters. We inject task and view information through cross-attention modules in T2I models, and further exhibit multi-view reference ability via the re-arranged self-attention modules. These enable LeftRefill to perform consistent generation as a generalized model without requiring test-time fine-tuning or model modifications. Thus, LeftRefill can be seen as a simple yet unified framework to address reference-guided synthesis. As an exemplar, we leverage LeftRefill to address two different challenges: reference-guided inpainting and novel view synthesis, based on the pre-trained StableDiffusion. Codes and models are released at https://github.com/ewrfcas/LeftRefill.

The task of face reenactment is to transfer the head motion and facial expressions from a driving video to the appearance of a source image, which may be of a different person (cross-reenactment). Most existing methods are CNN-based and estimate optical flow from the source image to the current driving frame, which is then inpainted and refined to produce the output animation. We propose a transformer-based encoder for computing a set-latent representation of the source image(s). We then predict the output color of a query pixel using a transformer-based decoder, which is conditioned with keypoints and a facial expression vector extracted from the driving frame. Latent representations of the source person are learned in a self-supervised manner that factorize their appearance, head pose, and facial expressions. Thus, they are perfectly suited for cross-reenactment. In contrast to most related work, our method naturally extends to multiple source images and can thus adapt to person-specific facial dynamics. We also propose data augmentation and regularization schemes that are necessary to prevent overfitting and support generalizability of the learned representations. We evaluated our approach in a randomized user study. The results indicate superior performance compared to the state-of-the-art in terms of motion transfer quality and temporal consistency.

Despite significant progress in the field, it is still challenging to create personalized visual representations that align closely with the desires and preferences of individual users. This process requires users to articulate their ideas in words that are both comprehensible to the models and accurately capture their vision, posing difficulties for many users. In this paper, we tackle this challenge by leveraging historical user interactions with the system to enhance user prompts. We propose a novel approach that involves rewriting user prompts based on a newly collected large-scale text-to-image dataset with over 300k prompts from 3115 users. Our rewriting model enhances the expressiveness and alignment of user prompts with their intended visual outputs. Experimental results demonstrate the superiority of our methods over baseline approaches, as evidenced in our new offline evaluation method and online tests. Our code and dataset are available at https://github.com/zzjchen/Tailored-Visions.

In recent years, Text-to-Image (T2I) models have seen remarkable advancements, gaining widespread adoption. However, this progress has inadvertently opened avenues for potential misuse, particularly in generating inappropriate or Not-Safe-For-Work (NSFW) content. Our work introduces MMA-Diffusion, a framework that presents a significant and realistic threat to the security of T2I models by effectively circumventing current defensive measures in both open-source models and commercial online services. Unlike previous approaches, MMA-Diffusion leverages both textual and visual modalities to bypass safeguards like prompt filters and post-hoc safety checkers, thus exposing and highlighting the vulnerabilities in existing defense mechanisms. Our codes are available at https://github.com/cure-lab/MMA-Diffusion.

Recent advancements in personalized text-to-image (T2I) models have revolutionized content creation, empowering non-experts to generate stunning images with unique styles. While promising, adding realistic motions into these personalized images by text poses significant challenges in preserving distinct styles, high-fidelity details, and achieving motion controllability by text. In this paper, we present PIA, a Personalized Image Animator that excels in aligning with condition images, achieving motion controllability by text, and compatibility with various personalized T2I models without specific tuning. To achieve these goals, PIA builds upon a base T2I model with well-trained temporal alignment layers, allowing for the seamless transformation of any personalized T2I model into an image animation model. A key component of PIA is the introduction of the condition module, which utilizes the condition frame and inter-frame affinity as input to transfer appearance information guided by the affinity hint for individual frame synthesis in the latent space. This design mitigates the challenges of appearance-related image alignment within PIA and allows for a stronger focus on aligning with motion-related guidance. To address the lack of a benchmark for this field, we introduce AnimateBench, a comprehensive benchmark comprising diverse personalized T2I models, curated images, and motion-related prompts. We show extensive experiments on AnimateBench to verify the superiority of PIA. We will make our codes and models publicly available.

Vector-Quantized Image Modeling (VQIM) is a fundamental research problem in image synthesis, which aims to represent an image with a discrete token sequence. Existing studies effectively address this problem by learning a discrete codebook from scratch and in a code-independent manner to quantize continuous representations into discrete tokens. However, learning a codebook from scratch and in a code-independent manner is highly challenging, which may be a key reason causing codebook collapse, i.e., some code vectors can rarely be optimized without regard to the relationship between codes and good codebook priors such that die off finally. In this paper, inspired by pretrained language models, we find that these language models have actually pretrained a superior codebook via a large number of text corpus, but such information is rarely exploited in VQIM. To this end, we propose a novel codebook transfer framework with part-of-speech, called VQCT, which aims to transfer a well-trained codebook from pretrained language models to VQIM for robust codebook learning. Specifically, we first introduce a pretrained codebook from language models and part-of-speech knowledge as priors. Then, we construct a vision-related codebook with these priors for achieving codebook transfer. Finally, a novel codebook transfer network is designed to exploit abundant semantic relationships between codes contained in pretrained codebooks for robust VQIM codebook learning. Experimental results on four datasets show that our VQCT method achieves superior VQIM performance over previous state-of-the-art methods.

This paper addresses the challenge of example-based non-stationary texture synthesis. We introduce a novel two-step approach wherein users first modify a reference texture using standard image editing tools, yielding an initial rough target for the synthesis. Subsequently, our proposed method, termed "self-rectification", automatically refines this target into a coherent, seamless texture, while faithfully preserving the distinct visual characteristics of the reference exemplar. Our method leverages a pre-trained diffusion network, and uses self-attention mechanisms, to gradually align the synthesized texture with the reference, ensuring the retention of the structures in the provided target. Through experimental validation, our approach exhibits exceptional proficiency in handling non-stationary textures, demonstrating significant advancements in texture synthesis when compared to existing state-of-the-art techniques.

Sampling from diffusion models can be treated as solving the corresponding ordinary differential equations (ODEs), with the aim of obtaining an accurate solution with as few number of function evaluations (NFE) as possible. Recently, various fast samplers utilizing higher-order ODE solvers have emerged and achieved better performance than the initial first-order one.However, these numerical methods inherently result in certain approximation errors, which significantly degrades sample quality with extremely small NFE (e.g., around 5).In contrast, based on the geometric observation that each sampling trajectory almost lies in a two-dimensional subspace embedded in the ambient space, we propose **A**pproximate **ME**an-**D**irection Solver (AMED-Solver) that eliminates truncation errors by directly learning the mean direction for fast diffusion sampling. Besides, our method can be easily used as a plugin to further improve existing ODE-based samplers. Extensive experiments on image synthesis with the resolution ranging from 32 to 512 demonstrate the effectiveness of our method. With only 5 NFE, we achieve 6.61 FID on CIFAR-10, 10.74 FID on ImageNet 64$\times$64, and 13.20 FID on LSUN Bedroom. Our code is available at https://github.com/zju-pi/diff-sampler.

This paper addresses the complex issue of one-shot face stylization, focusing on the simultaneous consideration of appearance and structure, where previous methods have fallen short. We explore deformation-aware face stylization that diverges from traditional single-image style reference, opting for a real-style image pair instead. The cornerstone of our method is the utilization of a self-supervised vision transformer, specifically DINO-ViT, to establish a robust and consistent facial structure representation across both real and style domains. Our stylization process begins by adapting the StyleGAN generator to be deformation-aware through the integration of spatial transformers (STN). We then introduce two innovative constraints for generator fine-tuning under the guidance of DINO semantics: i) a directional deformation loss that regulates directional vectors in DINO space, and ii) a relative structural consistency constraint based on DINO token self-similarities, ensuring diverse generation. Additionally, style-mixing is employed to align the color generation with the reference, minimizing inconsistent correspondences. This framework delivers enhanced deformability for general one-shot face stylization, achieving notable efficiency with a fine-tuning duration of approximately 10 minutes. Extensive qualitative and quantitative comparisons demonstrate the superiority of our approach over existing state-of-the-art one-shot face stylization methods.

This study focuses on a novel task in text-to-image (T2I) generation, namely action customization. The objective of this task is to learn the co-existing action from limited data and generalize it to unseen humans or even animals. Experimental results show that existing subject-driven customization methods fail to learn the representative characteristics of actions and struggle in decoupling actions from context features, including appearance. To overcome the preference for low-level features and the entanglement of high-level features, we propose an inversion-based method Action-Disentangled Identifier (ADI) to learn action-specific identifiers from the exemplar images. ADI first expands the semantic conditioning space by introducing layer-wise identifier tokens, thereby increasing the representational richness while distributing the inversion across different features. Then, to block the inversion of action-agnostic features, ADI extracts the gradient invariance from the constructed sample triples and masks the updates of irrelevant channels. To comprehensively evaluate the task, we present an ActionBench that includes a variety of actions, each accompanied by meticulously selected samples. Both quantitative and qualitative results show that our ADI outperforms existing baselines in action-customized T2I generation.

Despite their ability to generate high-resolution and diverse images from text prompts, text-to-image diffusion models often suffer from slow iterative sampling processes. Model distillation is one of the most effective directions to accelerate these models. However, previous distillation methods fail to retain the generation quality while requiring a significant amount of images for training, either from real data or synthetically generated by the teacher model. In response to this limitation, we present a novel image-free distillation scheme named SwiftBrush. Drawing inspiration from text-to-3D synthesis, in which a 3D neural radiance field that aligns with the input prompt can be obtained from a 2D text-to-image diffusion prior via a specialized loss without the use of any 3D data ground-truth, our approach re-purposes that same loss for distilling a pretrained multi-step text-to-image model to a student network that can generate high-fidelity images with just a single inference step. In spite of its simplicity, our model stands as one of the first one-step text-to-image generators that can produce images of comparable quality to Stable Diffusion without reliance on any training image data. Remarkably, SwiftBrush achieves an FID score of 16.67 and a CLIP score of 0.29 on the COCO-30K benchmark, achieving competitive results or even substantially surpassing existing state-of-the-art distillation techniques.

Deep Text-to-Image Synthesis (TIS) models such as Stable Diffusion have recently gained significant popularity for creative text-to-image generation. However, for domain-specific scenarios, tuning-free Text-guided Image Editing (TIE) is of greater importance for application developers. This approach modifies objects or object properties in images by manipulating feature components in attention layers during the generation process. Nevertheless, little is known about the semantic meanings that these attention layers have learned and which parts of the attention maps contribute to the success of image editing. In this paper, we conduct an in-depth probing analysis and demonstrate that cross-attention maps in Stable Diffusion often contain object attribution information, which can result in editing failures. In contrast, self-attention maps play a crucial role in preserving the geometric and shape details of the source image during the transformation to the target image. Our analysis offers valuable insights into understanding cross and self-attention mechanisms in diffusion models. Furthermore, based on our findings, we propose a simplified, yet more stable and efficient, tuning-free procedure that modifies only the self-attention maps of specified attention layers during the denoising process. Experimental results show that our simplified method consistently surpasses the performance of popular approaches on multiple datasets.

The recent wave of AI-generated content has witnessed the great development and success of Text-to-Image (T2I) technologies.By contrast, Text-to-Video (T2V) still falls short of expectations though attracting increasing interests. Existing works either train from scratch or adapt large T2I model to videos, both of which are computation and resource expensive. In this work, we propose a Simple Diffusion Adapter (SimDA) that fine-tunes only 24M out of 1.1B parameters of a strong T2I model, adapting it to video generation in a parameter-efficient way. In particular, we turn the T2I model for T2V by designing light-weight spatial and temporal adapters for transfer learning. Besides, we change the original spatial attention to the proposed Latent-Shift Attention (LSA) for temporal consistency. With a similar model architecture, we further train a video super-resolution model to generate high-definition (1024 x 1024) videos. In addition to T2V generation in the wild, SimDA could also be utilized in one-shot video editing with only 2 minutes tuning. Doing so, our method could minimize the training effort with extremely few tunable parameters for model adaptation.

Semantic image synthesis, i.e., generating images from user-provided semantic label maps, is an important conditional image generation task as it allows to control both the content as well as the spatial layout of generated images. Although diffusion models have pushed the state of the art in generative image modeling, the iterative nature of their inference process makes them computationally demanding. Other approaches such as GANs are more efficient as they only need a single feed-forward pass for generation, but the image quality tends to suffer on large and diverse datasets. In this work, we propose a new class of GAN discriminators for semantic image synthesis that generates highly realistic images by exploiting feature backbone networks pre-trained for tasks such as image classification. We also introduce a new generator architecture with better context modeling and using cross-attention to inject noise into latent variables, leading to more diverse generated images. Our model, which we dub DP-SIMS, achieves state-of-the-art results in terms of image quality and consistency with the input label maps on ADE-20K, COCO-Stuff, and Cityscapes, surpassing recent diffusion models while requiring two orders of magnitude less compute for inference.

Conventional image outpainting methods usually treat unobserved areas as unknown and extend the scene only in terms of semantic consistency, thus overlooking the hidden information in shadows cast by unobserved areas, such as the invisible shapes and semantics.In this paper, we propose to extract and utilize the hidden information of unobserved areas from their shadows to enhance image outpainting.To this end, we propose an end-to-end deep approach that explicitly looks into the shadows within the image.Specifically, we extract shadows from the input image and identify instance-level shadow regions cast by the unobserved areas.Then, the instance-level shadow representations are concatenated to predict the scene layout of each unobserved instance and outpaint the unobserved areas.Finally, two discriminators are implemented to enhance alignment between the extended semantics and their shadows. In the experiments, we show that our proposed approach provides complementary cues for outpainting and achieves considerable improvement on all datasets by adopting our approach as a plug-in module.

Contents generated by recent advanced Text-to-Image (T2I) diffusion models are sometimes too imaginative for existing off-the-shelf dense predictors to estimate due to the immitigable domain gap. We introduce DMP, a pipeline utilizing pre-trained T2I models as a prior for dense prediction tasks. To address the misalignment between deterministic prediction tasks and stochastic T2I models, we reformulate the diffusion process through a sequence of interpolations, establishing a deterministic mapping between input RGB images and output prediction distributions. To preserve generalizability, we use low-rank adaptation to fine-tune pre-trained models. Extensive experiments across five tasks, including 3D property estimation, semantic segmentation, and intrinsic image decomposition, showcase the efficacy of the proposed method. Despite limited-domain training data, the approach yields faithful estimations for arbitrary images, surpassing existing state-of-the-art algorithms.

We propose a method that can generate cinemagraphs automatically from a still landscape image using a pre-trained StyleGAN. Inspired by the success of recent unconditional video generation, we leverage a powerful pre-trained image generator to synthesize high-quality cinemagraphs. Unlike previous approaches that mainly utilize the latent space of a pre-trained StyleGAN, our approach utilizes its deep feature space for both GAN inversion and cinemagraph generation. Specifically, we propose multi-scale deep feature warping (MSDFW), which warps the intermediate features of a pre-trained StyleGAN at different resolutions. By using MSDFW, the generated cinemagraphs are of high resolution and exhibit plausible looping animation. We demonstrate the superiority of our method through user studies and quantitative comparisons with state-of-the-art cinemagraph generation methods and a video generation method that uses a pre-trained StyleGAN.

Existing diffusion-based video editing models have made gorgeous advances for editing attributes of a source video over time but struggle to manipulate the motion information while preserving the original protagonist's appearance and background. To address this, we propose MotionEditor, the first diffusion model for video motion editing. MotionEditor incorporates a novel content-aware motion adapter into ControlNet to capture temporal motion correspondence. While ControlNet enables direct generation based on skeleton poses, it encounters challenges when modifying the source motion in the inverted noise due to contradictory signals between the noise (source) and the condition (reference). Our adapter complements ControlNet by involving source content to transfer adapted control signals seamlessly. Further, we build up a two-branch architecture (a reconstruction branch and an editing branch) with a high-fidelity attention injection mechanism facilitating branch interaction. This mechanism enables the editing branch to query the key and value from the reconstruction branch in a decoupled manner, making the editing branch retain the original background and protagonist appearance. We also propose a skeleton alignment algorithm to address the discrepancies in pose size and position. Experiments demonstrate the promising motion editing ability of MotionEditor, both qualitatively and quantitatively. To the best of our knowledge, MotionEditor is the first diffusion-based model capable of video motion editing. More examples can be found at the anonymous website https://motioneditor.github.io.

Choreographers determine what the dances look like, while cameramen determine the final presentation of dances. Recently, various methods and datasets have showcased the feasibility of dance synthesis. However, camera movement synthesis with music and dance remains an unsolved challenging problem due to the scarcity of paired data. Thus, we present \textbf{DCM}, a new multi-modal 3D dataset, which for the first time combines camera movement with dance motion and music audio. This dataset encompasses 108 dance sequences (3.2 hours) of paired dance-camera-music data from the anime community, covering 4 music genres. With this dataset, we uncover that dance camera movement is multifaceted and human-centric, and possesses multiple influencing factors, making dance camera synthesis a more challenging task compared to camera or dance synthesis alone. To overcome these difficulties, we propose \textbf{DanceCamera3D}, a transformer-based diffusion model that incorporates a novel bones attention loss and a condition separation strategy. For evaluation, we devise new metrics measuring camera movement quality, diversity, and dancer fidelity. Utilizing these metrics, we conduct extensive experiments on our DCM dataset, providing both quantitative and qualitative evidence showcasing the effectiveness of our DanceCamera3D model.

StyleGAN has shown remarkable performance in unconditional image generation. However, its high computational cost poses a significant challenge for practical applications. Although recent efforts have been made to compress StyleGAN while preserving its performance, existing compressed models still lag behind the original model, particularly in terms of sample diversity. To overcome this, we propose a novel channel pruning method that leverages varying sensitivities of channels to latent vectors, which is a key factor in sample diversity. Specifically, by assessing channel importance based on their sensitivities to latent vector perturbations, our method enhances the diversity of samples in the compressed model. Since our method solely focuses on the channel pruning stage, it has complementary benefits with prior training schemes without additional training cost. Extensive experiments demonstrate that our method significantly enhances sample diversity across various datasets. Moreover, in terms of FID scores, our method not only surpasses state-of-the-art by a large margin but also achieves comparable scores with only half training iterations.

Diffusion models have achieved remarkable image generation quality surpassing previous generative models. However, a notable limitation of diffusion models, in comparison to GANs, is their difficulty in smoothly interpolating between two image samples, due to their highly unstructured latent space. Such a smooth interpolation is intriguing as it naturally serves as a solution for the image morphing task with many applications. In this work, we present DiffMorpher, the first approach enabling smooth and natural image interpolation using diffusion models. Our key idea is to capture the semantics of the two images by fitting two LoRAs to them respectively, and interpolate between both the LoRA parameters and the latent noises to ensure a smooth semantic transition, where correspondence automatically emerges without the need for annotation. In addition, we propose an attention interpolation and injection technique and a new sampling schedule to further enhance the smoothness between consecutive images. Extensive experiments demonstrate that DiffMorpher achieves starkly better image morphing effects than previous methods across a variety of object categories, bridging a critical functional gap that distinguished diffusion models from GANs.

Most image-to-image translation models postulate that a unique correspondence exists between the semantic classes of the source and target domains. However, this assumption does not always hold in real-world scenarios due to divergent distributions, different class sets, and asymmetrical information representation. As conventional GANs attempt to generate images that match the distribution of the target domain, they may hallucinate spurious instances of classes absent from the source domain, thereby diminishing the usefulness and reliability of translated images. CycleGAN-based methods are also known to hide the mismatched information in the generated images to bypass cycle consistency objectives, a process known as steganography.In response to the challenge of non-bijective image translation, we introduce StegoGAN, a novel model that leverages steganography to prevent spurious features in generated images. Our approach enhances the semantic consistency of the translated images without requiring additional postprocessing or supervision. Our experimental evaluations demonstrate that StegoGAN outperforms existing GAN-based models across various non-bijective image-to-image translation tasks, both qualitatively and quantitatively. Our code and pretrained models are accessible at https://github.com/sian-wusidi/StegoGAN.

Driven by the scalable diffusion models trained on large-scale datasets, text-to-image synthesis methods have shown compelling results. However, these models still fail to precisely follow the text prompt involving multiple objects, attributes, or spatial compositions.In this paper, we reveal the potential causes of the diffusion model's cross-attention and self-attention layers. We propose two novel losses to refocus attention maps according to a given spatial layout during sampling. Creating the layouts manually requires additional effort and can be tedious. Therefore, we explore using large language models (LLM) to produce these layouts for our method. We conduct extensive experiments on the DrawBench, HRS, and TIFA benchmarks to evaluate our proposed method. We show that our proposed attention refocusing effectively improves the controllability of existing approaches.

We present VecFusion, a new neural architecture that can generate vector fonts with varying topological structures and precise control point positions. Our approach is a cascaded diffusion model which consists of a raster diffusion model followed by a vector diffusion model. The raster model generates low-resolution, rasterized fonts with auxiliary control point information, capturing the global style and shape of the font, while the vector model synthesizes vector fonts conditioned on the low-resolution raster fonts from the first stage. To synthesize long and complex curves, our vector diffusion model uses a transformer architecture and a novel vector representation that enables the modeling of diverse vector geometry and the precise prediction of control points. Our experiments show that, in contrast to previous generative models for vector graphics, our new cascaded vector diffusion model generates higher quality vector fonts, with complex structures and diverse styles.

We introduce a framework for intrinsic latent diffusion models operating directly on the surfaces of 3D shapes, with the goal of synthesizing high-quality textures. Our approach is underpinned by two contributions: field latents, a latent representation encoding textures as discrete vector fields on the mesh vertices, and field latent diffusion models, which learn to denoise a diffusion process in the learned latent space on the surface. We consider a single-textured-mesh paradigm, where our models are trained to generate variations of a given texture on a mesh. We show the synthesized textures are of superior fidelity compared those from existing single-textured-mesh generative models. Our models can also be adapted for user-controlled editing tasks such as inpainting and label-guided generation. The efficacy of our approach is due in part to the equivariance of our proposed framework under isometries, allowing our models to seamlessly reproduce details across locally similar regions and opening the door to a notion of generative texture transfer.

Customization techniques for text-to-image models have paved the way for a wide range of previously unattainable applications, enabling the generation of specific concepts across diverse contexts and styles. While existing methods facilitate high-fidelity customization for individual concepts or a limited, pre-defined set of them, they fall short of achieving scalability, where a single model can seamlessly render countless concepts. In this paper, we address a new problem called Modular Customization, with the goal of efficiently merging customized models that were fine-tuned independently for individual concepts. This allows the merged model to jointly synthesize concepts in one image without compromising fidelity or incurring any additional computational costs.To address this problem, we introduce Orthogonal Adaptation, a method designed to encourage the customized models, which do not have access to each other during fine-tuning, to have orthogonal residual weights. This ensures that during inference time, the customized models can be summed with minimal interference. Our proposed method is both simple and versatile, applicable to nearly all optimizable weights in the model architecture. Through an extensive set of quantitative and qualitative evaluations, our method consistently outperforms relevant baselines in terms of efficiency and identity preservation, demonstrating a significant leap toward scalable customization of diffusion models.

The rise of new video modalities like virtual reality or autonomous driving has increased the demand for efficient multi-view video compression methods, both in terms of rate-distortion (R-D) performance and in terms of delay and runtime. While most recent stereo video compression approaches have shown promising performance, they compress left and right views sequentially, leading to poor parallelization and runtime performance. This work presents Low-Latency neural codec for Stereo video Streaming (LLSS), a novel parallel stereo video coding method designed for fast and efficient low-latency stereo video streaming. Instead of using a sequential cross-view motion compensation like existing methods, LLSS introduces a bidirectional feature shifting module to directly exploit mutual information among views and encode them effectively with a joint cross-view prior model for entropy coding. Thanks to this design, LLSS processes left and right views in parallel, minimizing latency; all while substantially improving R-D performance compared to both existing neural and conventional codecs.

Diffusion-based text-to-image generative models, e.g., Stable Diffusion, have revolutionized the field of content generation, enabling significant advancements in areas like image editing and video synthesis.Despite their formidable capabilities, these models are not without their limitations. It is still challenging to synthesize an image that aligns well with the input text, and multiple runs with carefully crafted prompts are required to achieve satisfactory results.To mitigate these limitations, numerous studies have endeavored to fine-tune the pre-trained diffusion models, i.e., UNet, utilizing various technologies. Yet, amidst these efforts, a pivotal question of text-to-image diffusion model training has remained largely unexplored: Is it possible and feasible to fine-tune the text encoder to improve the performance of text-to-image diffusion models? Our findings reveal that, instead of replacing the CLIP text encoder used in Stable Diffusion with other large language models, we can enhance it through our proposed fine-tuning approach, TextCraftor, leading to substantial improvements in quantitative benchmarks and human assessments. Interestingly, our technique also empowers controllable image generation through the interpolation of different text encoders fine-tuned with various rewards.We also demonstrate that TextCraftor is orthogonal to UNet finetuning, and can be combined to further improve generative quality.

Recent breakthroughs in text-to-4D generation rely on pre-trained text-to-image and text-to-video models to generate dynamic 3D scenes. However, current text-to-4D methods face a three-way tradeoff between the quality of scene appearance, 3D structure, and motion. For example, text-to-image models and their 3D-aware variants are trained on internet-scale image datasets and can be used to produce scenes with realistic appearance and 3D structure---but no motion. Text-to-video models are trained on relatively smaller video datasets and can produce scenes with motion, but poorer appearance and 3D structure. While these models have complementary strengths, they also have opposing weaknesses, making it difficult to combine them in a way that alleviates this three-way tradeoff. Here, we introduce hybrid score distillation sampling, an alternating optimization procedure that blends supervision signals from multiple pre-trained diffusion models and incorporates benefits of each for high-fidelity text-to-4D generation.Using hybrid SDS, we demonstrate synthesis of 4D scenes with compelling appearance, 3D structure, and motion.

Diffusion models have shown an impressive ability to model complex data distributions, with several key advantages over GANs, such as stable training, better coverage of the training distribution's modes, and the ability to solve inverse problems without extra training. However, most diffusion models learn the distribution of fixed-resolution images. We propose to learn the distribution of continuous images by training diffusion models on image neural fields, which can be rendered at any resolution, and show its advantages over fixed-resolution models. To achieve this, a key challenge is to obtain a latent space that represents photorealistic image neural fields. We propose a simple and effective method, inspired by several recent techniques but with key changes to make the image neural fields photorealistic. Our method can be used to convert existing latent diffusion autoencoders into image neural field autoencoders. We show that image neural field diffusion models can be trained using mixed-resolution image datasets, outperform fixed-resolution diffusion models followed by super-resolution models, and can solve inverse problems with conditions applied at different scales efficiently.

Current metrics for text-to-image models typically relies on statistical metrics which inadequately represent the real preference of humans. Although recent works attempt to learn these preferences via human annotated images, they reduce the rich tapestry of human preference to a single overall score. However, the preference results vary when humans evaluate images with different aspects. Therefore, to learn the multi-dimensional human preferences, we propose Multi-dimensional Preference Score (MPS), the first multi-dimensional preference scoring model for the evaluation of text-to-image models. The MPS introduces the preference condition module upon CLIP model to learn these diverse preferences. It is trained based on our Multi-dimensional Human Preference (MHP) Dataset, which comprises 918,315 human preference choices across 4 dimensions (i.e., aesthetics, semantic alignment, detail quality and overall assessment) on 607,541 images. The images are generated by a wide range of latest text-to-image models. The MPS outperforms existing scoring methods across 3 datasets in 4 dimensions, enabling it a promising metric for evaluating and improving text-to-image generation. The model and dataset will be made publicly available to facilitate future research.

Multi-Instance Learning (MIL) has shown impressive performance for histopathology whole slide image (WSI) analysis using bags or pseudo-bags. It involves instance sampling, feature representation, and decision-making. However, existing MIL-based technologies at least suffer from one or more of the following problems: 1) requiring high storage and intensive pre-processing for numerous instances (sampling); 2) potential over-fitting with limited knowledge to predict bag labels (feature representation); 3) pseudo-bag counts and prior biases affect model robustness and generalizability (decision-making). Inspired by clinical diagnostics, using the past sampling instances can facilitate the final WSI analysis, but it is barely explored in prior technologies. To break free these limitations, we integrate the dynamic instance sampling and reinforcement learning into a unified framework to improve the instance selection and feature aggregation, forming a novel Dynamic Policy Instance Selection (DPIS) scheme for better and more credible decision-making. Specifically, the measurement of feature distance and reward function are employed to boost continuous instance sampling. To alleviate the over-fitting, we explore the latent global relations among instances for more robust and discriminative feature representation while establishing reward and punishment mechanisms to correct biases in pseudo-bags using contrastive learning. These strategies form the final Dynamic Policy-Driven Adaptive Multi-Instance Learning (PAMIL) method for WSI tasks. Extensive experiments reveal that our PAMIL method outperforms the state-of-the-art by 3.8\% on CAMELYON16 and 4.4\% on TCGA lung cancer datasets.

Denoising diffusion probabilistic models (DDPMs) for image inpainting aim to add the noise to the texture of the image during the diffusion process and recover the masked regions with the unmasked ones of the texture via the reverse denoising process. Despite the meaningful semantics generation, the existing arts suffer from the semantic discrepancy between the masked and unmasked regions, since the semantically dense unmasked texture fails to be completely degraded while the masked regions turn to the pure noise in diffusion process, leading to the large discrepancy between them. In this paper, we aim to answer how the unmasked semantics guide the texture denoising process; together with how to tackle the semantic discrepancy, to enable the consistent and meaningful semantics generation. To this end, we propose a novel structure-guided diffusion model for image inpainting (namely StrDiffusion), which reformulates the conventional texture denoising process under the guidance of the structure to derive a simplified denoising objective for inpainting, while revealing: 1) unlike the texture, the semantically sparse structure is beneficial to tackle the semantic discrepancy; 2) the semantics from the unmasked regions essentially offer the time-dependent guidance for the texture denoising process, benefiting from the time-dependent sparsity of the structure semantics. For the denoising process, a structure-guided neural network is trained to estimate the simplified denoising objective by exploiting the consistency of the denoised structure between masked and unmasked regions. Besides, we devise an adaptive resampling strategy as a formal criterion on whether the structure is competent to guide the texture denoising process, while regulate their semantic correlations. Extensive experiments validate the merits of StrDiffusion over the state-of-the-arts. Our code is available in the supplementary material.

Generative object compositing emerges as a promising new avenue for compositional image editing. However, the requirement of object identity preservation poses a significant challenge, limiting practical usage of most existing methods. In response, this paper introduces IMPRINT, a novel diffusion-based generative model trained with a two-stage learning framework that decouples learning of identity preservation from that of compositing. The first stage is targeted for context-agnostic, identity-preserving pretraining of the object encoder, enabling the encoder to learn an embedding that is both view-invariant and conducive to enhanced detail preservation. The subsequent stage leverages this representation to learn seamless harmonization of the object composited to the background. In addition, IMPRINT incorporates a shape-guidance mechanism offering user-directed control over the compositing process. Extensive experiments demonstrate that IMPRINT significantly outperforms existing methods and various baselines on identity preservation and composition quality.

Style Transfer belongs to the task of image generation. It aims to render an image with the artistic features of a style image, while maintaining the original structure. Many methods have been put forward for this task, but some challenges still exist. It is difficult for CNN-based methods to handle global information and long-range dependencies between the input images. Afterwards, some researchers have proposed transformer-based methods. Although transformer can better model the relationship between the content image and style image, these methods require high-cost hardware and time-consuming inference. To address these issues, we design a novel transformer model that includes only encoders, thus significantly reducing the computational cost. In addition, we also find that the result images generated by existing style transfer methods may lead to the image under-stylied or missing content. In order to achieve better stylization, we design a content feature extractor and style feature extractor. Then we can feed pure content and style images into the transformer. Finally, we propose a network model termed Puff-Net, i.e., efficient style transfer with pure content and style feature fusion network. Through qualitative and quantitative experiments, we verify the performance advantages of our model compared to state-of-the-art models in the literature.

Recent advancements in subject-driven image generation have led to zero-shot generation, yet precise selection and focus on crucial subject representations remain challenging. Addressing this, we introduce the SSR-Encoder, a novel architecture designed for selectively capturing any subject from single or multiple reference images. It responds to various query modalities including text and masks, without necessitating test-time fine-tuning. The SSR-Encoder combines a Token-to-Patch Aligner that aligns query inputs with image patches and a Detail-Preserving Subject Encoder for extracting and preserving fine features of the subjects, thereby generating subject embeddings. These embeddings, used in conjunction with original text embeddings, condition the generation process. Characterized by its model generalizability and efficiency, the SSR-Encoder adapts to a range of custom models and control modules. Enhanced by the Embedding Consistency Regularization Loss for improved training, our extensive experiments demonstrate its effectiveness in versatile and high-quality image generation, indicating its broad applicability.

Modern video generation models like Sora have achieved remarkable success in producing high-quality videos.However, a significant limitation is their inability to offer interactive control to users, a feature that promises to open up unprecedented applications and creativity. In this work, we introduce the first solution to equip diffusion-based video generation models with spatio-temporal control. We present Peekaboo, a novel masked attention module, which seamlessly integrates with current video generation models offering control without the need for additional training or inference overhead. To facilitate future research, we also introduce a comprehensive benchmark for interactive video generation. This benchmark offers a standardized framework for the community to assess the efficacy of emerging interactive video generation models.Our extensive qualitative and quantitative assessments reveal that Peekaboo achieves up to a $3.8\times$ improvement in mIoU over baseline models, all while maintaining the same latency. Code and benchmark are available on the webpage.

We present the content deformation field (CoDeF) as a new type of video representation, which consists of a canonical content field aggregating the static contents in the entire video and a temporal deformation field recording the transformations from the canonical image (i.e., rendered from the canonical content field) to each individual frame along the time axis. Given a target video, these two fields are jointly optimized to reconstruct it through a carefully tailored rendering pipeline. We advisedly introduce some regularizations into the optimization process, urging the canonical content field to inherit semantics (e.g., the object shape) from the video. With such a design, CoDeF naturally supports lifting image algorithms for video processing, in the sense that one can apply an image algorithm to the canonical image and effortlessly propagate the outcomes to the entire video with the aid of the temporal deformation field. We experimentally show that CoDeF is able to lift image-to-image translation to video-to-video translation and lift keypoint detection to keypoint tracking without any training. More importantly, thanks to our lifting strategy that deploys the algorithms on only one image, we achieve superior cross-frame consistency in processed videos compared to existing video-to-video translation approaches, and even manage to track non-rigid objects like water and smog. Code is made available at https://qiuyu96. github.io/CoDeF/.

The objective of text-to-image (T2I) personalization is to customize a diffusion model to a user-provided reference concept, generating diverse images of the concept aligned with the target prompts. Conventional methods representing the reference concepts using unique text embeddings often fail to accurately mimic the appearance of the reference. To address this, one solution may be explicitly conditioning the reference images into the target denoising process, known as key-value replacement. However, prior works are constrained to local editing since they disrupt the structure path of the pre-trained T2I model. To overcome this, we propose a novel plug-in method, called DreamMatcher, which reformulates T2I personalization as semantic matching. Specifically, DreamMatcher replaces the target values with reference values aligned by semantic matching, while leaving the structure path unchanged to preserve the versatile capability of pre-trained T2I models for generating diverse structures. We also introduce a semantic-consistent masking strategy to isolate the personalized concept from irrelevant regions introduced by the target prompts. Compatible with existing T2I models, DreamMatcher shows significant improvements in complex scenarios. Intensive analyses demonstrate the effectiveness of our approach.

Utilizing pre-trained 2D large-scale generative models, recent works are capable of generating high-quality novel views from a single in-the-wild image. However, due to the lack of information from multiple views, these works encounter difficulties in generating controllable novel views. In this paper, we present DreamComposer, a flexible and scalable framework that can enhance existing view-aware diffusion models by injecting multi-view conditions. Specifically, DreamComposer first uses a view-aware 3D lifting module to obtain 3D representations of an object from multiple views. Then, it renders the latent features of the target view from 3D representations with the multi-view feature fusion module. Finally the target view features extracted from multi-view inputs are injected into a pre-trained diffusion model. Experiments show that DreamComposer is compatible with state-of-the-art diffusion models for zero-shot novel view synthesis, further enhancing them to generate high-fidelity novel view images with multi-view conditions, ready for controllable 3D object reconstruction and various other applications. Codes will be released upon acceptance.

In the realm of image composition, generating realistic shadow for the inserted foreground remains a formidable challenge. Previous works have developed image-to-image translation models which are trained on paired training data. However, they are struggling to generate shadows with accurate shapes and intensities, hindered by data scarcity and the inherent task complexity. In this paper, we resort to foundational model with rich prior knowledge of natural shadow images. Specifically, we first adapt ControlNet to our task and then propose intensity modulation modules to improve the shadow intensity. Moreover, we extend the small-scale DESOBA dataset to DESOBAv2 using a novel data acquisition pipeline. Experimental results on both DESOBA and DESOBAv2 datasets as well as real composite images demonstrate the superior capability of our model in shadow generation task.

Existing score distillation methods are sensitive to classifier-free guidance (CFG) scale: manifested as over-smoothness or instability at small CFG scales, while over-saturation at large ones. To explain and analyze these issues, we revisit the derivation of Score Distillation Sampling (SDS) and decipher existing score distillation with the Wasserstein Generative Adversarial Network (WGAN) paradigm. With the WGAN paradigm, we find that existing score distillation either employs a fixed sub-optimal discriminator or conducts incomplete discriminator optimization, resulting in the scale-sensitive issue. We propose the Adversarial Score Distillation (ASD), which maintains an optimizable discriminator and updates it using the complete optimization objective. Experiments show that the proposed ASD performs favorably in 2D distillation and text-to-3D tasks against existing methods. Furthermore, to explore the generalization ability of our WGAN paradigm, we extend ASD to the image editing task, which achieves competitive results.

Unsupervised Domain Adaptation (UDA) can effectively address domain gap issues in real-world image Super-Resolution (SR) by accessing both the source and target data. Considering privacy policies or transmission restrictions of source data in practical scenarios, we propose a SOurce-free Domain Adaptation framework for image SR (SODA-SR) to address this issue, i.e., adapt a source-trained model to a target domain with only unlabeled target data. SODA-SR leverages the source-trained model to generate refined pseudo-labels for teacher-student learning. To better utilize pseudo-labels, we propose a novel wavelet-based augmentation method, named Wavelet Augmentation Transformer (WAT), which can be flexibly incorporated with existing networks, to implicitly produce useful augmented data. WAT learns low-frequency information of varying levels across diverse samples, which is aggregated efficiently via deformable attention. Furthermore, an uncertainty-aware self-training mechanism is proposed to improve the accuracy of pseudo-labels, with inaccurate predictions being rectified by uncertainty estimation. To acquire better SR results and avoid overfitting pseudo-labels, several regularization losses are proposed to constrain target LR and SR images in the frequency domain. Experiments show that without accessing source data, SODA-SR outperforms state-of-the-art UDA methods in both synthetic$\rightarrow$real and real$\rightarrow$real adaptation settings, and is not constrained by specific network architectures.

Character Animation aims to generating character videos from still images through driving signals. Currently, diffusion models have become the mainstream in visual generation research, owing to their robust generative capabilities. However, challenges persist in the realm of image-to-video, especially in character animation, where temporally maintaining consistency with detailed information from character remains a formidable problem. In this paper, we leverage the power of diffusion models and propose a novel framework tailored for character animation. To preserve consistency of intricate appearance features from reference image, we design ReferenceNet to merge detail features via spatial attention. To ensure controllability and continuity, we introduce an efficient pose guider to direct character's movements and employ an effective temporal modeling approach to ensure smooth inter-frame transitions between video frames. By expanding the training data, our approach can animate arbitrary characters, yielding superior results in character animation compared to other image-to-video methods. Furthermore, we evaluate our method on image animation benchmarks, achieving state-of-the-art results.

Recently, there were remarkable advances in image editing tasks in various ways. Nevertheless, existing image editing models are not designed for Human-Object Interaction (HOI) image editing. One of these approaches (e.g. ControlNet) employs the skeleton guidance to offer precise representations of human, showing better results in HOI image editing. However, using conventional methods, manually creating HOI skeleton guidance is necessary. This paper proposes the object interactive diffuser with associative attention that considers both the interaction with objects and the joint graph structure, automating the generation of HOI skeleton guidance. Additionally, we propose the HOI loss with novel scaling parameter, demonstrating its effectiveness in generating skeletons that interact better. To evaluate generated object-interactive skeletons, we propose two metrics, top-N accuracy and skeleton probabilistic distance. Our framework integrates object interactive diffuser that generates object-interactive skeletons with previous methods, demonstrating the outstanding results in HOI image editing. Finally, we present potentials of our framework beyond HOI image editing, as applications to human-to-human interaction, skeleton editing, and 3D mesh optimization. The code is available at https://github.com/YangChangHee/CVPR2024Person-In-PlaceRELEASE

Given a clothing image and a person image, an image-based virtual try-on aims to generate a customized image that appears natural and accurately reflects the characteristics of the clothing image. In this work, we aim to expand the applicability of the pre-trained diffusion model so that it can be utilized independently for the virtual try-on task. The main challenge is to preserve the clothing details while effectively utilizing the robust generative capability of the pre-trained model. In order to tackle these issues, we propose StableVITON, learning the semantic correspondence between the clothing and the human body within the latent space of the pre-trained diffusion model in an end-to-end manner. Our proposed zero cross-attention blocks not only preserve the clothing details by learning the semantic correspondence but also generate high-fidelity images by utilizing the inherent knowledge of the pre-trained model in the warping process. Through our proposed novel attention total variation loss and applying augmentation, we achieve the sharp attention map, resulting in a more precise representation of clothing details. StableVITONoutperforms the baselines in qualitative and quantitative evaluation, showing promising quality in arbitrary person images.

Recent thrilling progress in large-scale text-to-image (T2I) models has unlocked unprecedented synthesis quality of AI-generated content (AIGC) including image generation, 3D and video composition. Further, personalized techniques enable appealing customized production of a novel concept given only several images as reference. However, an intriguing problem persists: Is it possible to capture \textbf{multiple, novel concepts} from \textbf{one single reference image}? In this paper, we identify that existing approaches fail to preserve visual consistency with the reference image and eliminate cross-influence from concepts. To alleviate this, we propose an attention calibration mechanism to improve the concept-level understanding of the T2I model. Specifically, we first introduce new learnable modifiers bound with classes to capture attributes of multiple concepts. Then, the classes are separated and strengthened following the activation of the cross-attention operation, ensuring comprehensive and self-contained concepts. Additionally, we suppress the attention activation of different classes to mitigate mutual influence among concepts. Together, our proposed method, dubbed \textbf{DisenDiff}, can learn disentangled multiple concepts from one single image and produce novel customized images with learned concepts. We demonstrate that our method outperforms the current state of the art in both qualitative and quantitative evaluations. More importantly, our proposed techniques are compatible with LoRA and inpainting pipelines, enabling more interactive experiences.

We present personalized residuals and localized attention-guided sampling for efficient concept-driven generation using text-to-image diffusion models. Our method first represents concepts by freezing the weights of a pretrained text-conditioned diffusion model and learning low-rank residuals for a small subset of the model's layers. The residual-based approach then directly enables application of our proposed sampling technique, which applies the learned residuals only in areas where the concept is localized via cross-attention and applies the original diffusion weights in all other regions. Localized sampling therefore combines the learned identity of the concept with the existing generative prior of the underlying diffusion model. We show that personalized residuals effectively capture the identity of a concept in ~3 minutes on a single GPU without the use of regularization images and with fewer parameters than previous models, and localized sampling allows using the original model as strong prior for large parts of the image.

Text-to-image diffusion models have demonstrated remarkable capabilities in transforming text prompts into coherent images, yet the computational cost of the multi-step inference remains a persistent challenge. To address this issue, we present UFOGen, a novel generative model designed for ultra-fast, one-step text-to-image generation. In contrast to conventional approaches that focus on improving samplers or employing distillation techniques for diffusion models, UFOGen adopts a hybrid methodology, integrating diffusion models with a GAN objective. Leveraging a newly introduced diffusion-GAN objective and initialization with pre-trained diffusion models, UFOGen excels in efficiently generating high-quality images conditioned on textual descriptions in a single step. Beyond traditional text-to-image generation, UFOGen showcases versatility in applications. Notably, UFOGen stands among the pioneering models enabling one-step text-to-image generation and diverse downstream tasks, presenting a significant advancement in the landscape of efficient generative models.

Diffusion models have transformed the image-to-image (I2I) synthesis and are now permeating into videos. However, the advancement of video-to-video (V2V) synthesis has been hampered by the challenge of maintaining temporal consistency across video frames. This paper proposes a consistent V2V synthesis framework by jointly leveraging spatial conditions and temporal optical flow clues within the source video. Contrary to prior methods that strictly adhere to optical flow, our approach harnesses its benefits while handling the imperfection in flow estimation. We encode the optical flow via warping from the first frame and serve it as a supplementary reference in the diffusion model. This enables our model for video synthesis by editing the first frame with any prevalent I2I models, and then propagating edits to successive frames. Our V2V model, FlowVid, demonstrates remarkable properties: (1) Flexibility: FlowVid works seamlessly with existing I2I models, facilitating various modifications, including stylization, object swaps, and local edits. (2) Efficiency: Generation of a 4-second video with 30 FPS and 512$\times$512 resolution takes only 1.5 minutes, which is 3.1$\times$, 7.2$\times$, and 10.5$\times$ faster than CoDeF, Rerender, and TokenFlow, respectively. (3) High-quality: In user studies, our FlowVid is preferred 45.7\% of the time, outperforming CoDeF (3.5\%), Rerender (10.2\%), and TokenFlow (40.4\%).

We present Readout Guidance, a method for controlling text-to-image diffusion models with learned signals. Readout Guidance uses readout heads, lightweight networks trained to extract signals from the features of a pre-trained, frozen diffusion model at every timestep. These readouts can encode single-image properties, such as pose, depth, and edges; or higher-order properties that relate multiple images, such as correspondence and appearance similarity. Furthermore, by comparing the readout estimates to a user-defined target, and back-propagating the gradient through the readout head, these estimates can be used to guide the sampling process. Compared to prior methods for conditional generation, Readout Guidance requires significantly fewer added parameters and training samples, and offers a convenient and simple recipe for reproducing different forms of conditional control under a single framework, with a single architecture and sampling procedure. We showcase these benefits in the applications of drag-based manipulation, identity-consistent generation, and spatially aligned control.

Large language models (LLMs) are fine-tuned using human comparison data with Reinforcement Learning from Human Feedback (RLHF) methods to make them better aligned with users' preferences. In contrast to LLMs, human preference learning has not been widely explored in text-to-image diffusion models; the best existing approach is to fine-tune a pretrained model using carefully curated high quality images and captions to improve visual appeal and text alignment. We propose Diffusion-DPO, a method to align diffusion models to human preferences by directly optimizing on human comparison data. Diffusion-DPO is adapted from the recently developed Direct Preference Optimization (DPO), a simpler alternative to RLHF which directly optimizes a policy that best satisfies human preferences under a classification objective. We re-formulate DPO to account for a diffusion model notion of likelihood, utilizing the evidence lower bound to derive a differentiable objective. Using the Pick-a-pic dataset of 851K crowdsourced pairwise preferences, we fine-tune the base model of the state-of-the-art Stable Diffusion XL (SDXL)-1.0 model with Diffusion-DPO. Our fine-tuned base model significantly outperforms both base SDXL-1.0 and the larger SDXL-1.0 model consisting of an additional refinement model in human evaluation, improving visual appeal and prompt alignment. We also develop a variant that uses AI feedback and has comparable performance to training on human preferences, opening the door for scaling of diffusion model alignment methods.

In recent advancements in high-fidelity image generation, Denoising Diffusion Probabilistic Models (DDPMs) have emerged as a key player. However, their application at high resolutions presents significant computational challenges. Current methods, such as patchifying, expedite processes in UNet and Transformer architectures but at the expense of representational capacity. Addressing this, we introduce the Diffusion State Space Model (DiffuSSM), an innovative architecture that supplants attention mechanisms with a more scalable state space model backbone. This approach effectively handles higher resolutions without resorting to global compression, thus preserving detailed image representation throughout the diffusion process. Our focus on FLOP-efficient architectures in diffusion training marks a significant step forward. Comprehensive evaluations on both ImageNet and LSUN datasets at two resolutions demonstrate that DiffuSSMs are on par or even outperform existing diffusion models with attention modules in FID and Inception Score metrics while significantly reducing total FLOP usage.

We assemble a dataset of Creative-Commons-licensed (CC) images, which we use to train a set of open diffusion models that are qualitatively competitive with Stable Diffusion 2 (SD2). This task presents two challenges: (1) high-resolution CC images lack the captions necessary to train text-to-image generative models; (2) CC images are relatively scarce. In turn, to address these challenges, we use an intuitive transfer learning technique to produce a set of high-quality synthetic captions paired with curated CC images. We then develop a data- and compute-efficient training recipe that requires as little as 3% of the LAION-2B data needed to train existing SD2 models, but obtains comparable quality. These results indicate that we have a sufficient number of CC images (~70 million) for training high-quality models. Our training recipe also implements a variety of optimizations that achieve ~3X training speed-ups, enabling rapid model iteration. We leverage this recipe to train several high-quality text-to-image models, which we dub the CommonCanvas family. Our largest model achieves comparable performance to SD2 on a human evaluation, despite being trained on our CC dataset that is significantly smaller than LAION and using synthetic captions for training. We release our models, data, and code at [REDACTED].

In this paper, we introduce Fairy, a minimalist yet robust adaptation of image-editing diffusion models, enhancing them for video editing applications. Our approach centers on the concept of anchor-based cross-frame attention, a mechanism that implicitly propagates diffusion features across frames, ensuring superior temporal coherence and high-fidelity synthesis. Fairy not only addresses limitations of previous models, including memory and processing speed, by optimizing parallel computing but also improves temporal consistency through a unique data augmentation strategy. This strategy renders the model equivariant to affine transformations in both source and target images. Remarkably efficient, Fairy generates 120 frames of high-resolution video (4-second duration at 30 FPS) in just 14 seconds, significantly outpacing prior works by at least 44x. A comprehensive user study, involving 1000 generated samples, confirms that our approach delivers superior quality, decisively outperforming established methods.

In recent years, image editing has advanced remarkably. With increased human control, it is now possible to edit an image in a plethora of ways; from specifying in text what we want to change, to straight up dragging the contents of the image in an interactive point-based manner.However, most of the focus has remained on editing single images at a time.Whether and how we can simultaneously edit large batches of images has remained understudied.With the goal of minimizing human supervision in the editing process, this paper presents a novel method for interactive batch image editing using StyleGAN as the medium.Given an edit specified by users in an example image (e.g., make the face frontal), our method can automatically transfer that edit to other test images, so that regardless of their initial state (pose), they all arrive at the same final state (e.g., all facing front).Extensive experiments demonstrate that edits performed using our method have similar visual quality to existing single-image-editing methods, while having more visual consistency and saving significant time and human effort.

Underwater images are subject to intricate and diverse degradation, exerting an inevitable impact on the efficacy of underwater visual tasks. However, most approaches primarily operate in the raw pixel space of images, showing constrained exploration of underwater image frequency properties, leading to an inadequate utilization of deep models' representational capabilities in producing high-quality images. In this paper, we introduce a novel Underwater Image Enhancement (UIE) framework, named WF-Diff, designed to fully leverage the characteristics of frequency domain information and diffusion models. WF-Diff consists of two detachable networks: Wavelet-based Fourier information interaction network (WFI2-net) and Frequency Residual Diffusion Adjustment Module (FRDAM). With our full exploration of the frequency domain information, WFI2-net aims to achieve preliminary enhancement of frequency information in the wavelet space. Our proposed FRDAM can further refine the high- and low-frequency information of the initial enhanced images, which can be viewed as a plug-and-play universal module to adjustment the detail of the underwater images. With the above techniques, our algorithm can show SOTA performance on real-world underwater image datasets, and achieves competitive performance in visual quality. The code is availableat https://github.com/zhihefang/WF-Diff.

Diffusion probabilistic models (DPMs) have shown remarkable performance in high-resolution image synthesis, but their sampling efficiency is still to be desired due to the typically large number of sampling steps. Recent advancements in high-order numerical ODE solvers for DPMs have enabled the generation of high-quality images with much fewer sampling steps. While this is a significant development, most sampling methods still employ uniform time steps, which is not optimal when using a small number of steps. To address this issue, we propose a general framework for designing an optimization problem that seeks more appropriate time steps for a specific numerical ODE solver for DPMs. This optimization problem aims to minimize the distance between the ground-truth solution to the ODE and an approximate solution corresponding to the numerical solver. It can be efficiently solved using the constrained trust region method, taking less than $15$ seconds. Our extensive experiments on both unconditional and conditional sampling using pixel- and latent-space DPMs demonstrate that, when combined with the state-of-the-art sampling method UniPC, our optimized time steps significantly improve image generation performance in terms of FID scores for datasets such as CIFAR-10 and ImageNet, compared to using uniform time steps.

While GAN-based models have been successful in image stylization tasks, they often struggle with structure preservation while stylizing a wide range of input images. Recently, diffusion models have been adopted for image stylization but still lack the capability to maintain the original quality of input images. Building on this, we propose OSASIS: a novel one-shot stylization method that is robust in structure preservation. We show that OSASIS is able to effectively disentangle the semantics from the structure of an image, allowing it to control the level of content and style implemented to a given input. We apply OSASIS to various experimental settings, including stylization with out-of-domain reference images and stylization with text-driven manipulation. Results show that OSASIS outperforms other stylization methods, especially for input images that were rarely encountered during training, providing a promising solution to stylization via diffusion models.

In text-to-image personalization, a timely and crucial challenge is the tendency of generated images overfitting to the biases present in the reference images. We initiate our study with a comprehensive categorization of the biases into background, nearby-object, tied-object, substance (in style re-contextualization), and pose biases. These biases manifest in the generated images due to their entanglement into the subject embedding. This undesired embedding entanglement not only results in the reflection of biases from the reference images into the generated images but also notably diminishes the alignment of the generated images with the given generation prompt. To address this challenge, we propose SID (Selectively Informative Description), a text description strategy that deviates from the prevalent approach of only characterizing the subject’s class identification. SID is generated utilizing multimodal GPT-4 and can be seamlessly integrated into optimization-based models. We present comprehensive experimental results along with analyses of cross-attention maps, subject-alignment, non-subject-disentanglement, and text-alignment.

We propose a novel diffusion model called observation-guided diffusion probabilistic model (OGDM), which effectively addresses the trade-off between quality control and fast sampling. Our approach reestablishes the training objective by integrating the guidance of the observation process with the Markov chain in a principled way. This is achieved by introducing an additional loss term derived from the observation based on the conditional discriminator on noise level, which employs Bernoulli distribution indicating whether its input lies on the (noisy) real manifold or not. This strategy allows us to optimize the more accurate negative log-likelihood induced in the inference stage especially when the number of function evaluations is limited. The proposed training method is also advantageous even when incorporated only into the fine-tuning process, and it is compatible with various fast inference strategies since our method yields better denoising networks using the exactly same inference procedure without incurring extra computational cost. We demonstrate the effectiveness of the proposed training algorithm using diverse inference methods on strong diffusion model baselines.

Long-form video content constitutes a significant portion of internet traffic, making automated video summarization an essential research problem. However, existing video summarization datasets are notably limited in their size, constraining the effectiveness of state-of-the-art methods for generalization. Our work aims to overcome this limitation by capitalizing on the abundance of long-form videos with dense speech-to-video alignment and the remarkable capabilities of recent large language models (LLMs) in summarizing long text. We introduce an automated and scalable pipeline for generating a large-scale video summarization dataset using LLMs as Oracle summarizers. By leveraging the generated dataset, we analyze the limitations of existing approaches and propose a new video summarization model that effectively addresses them. To facilitate further research in the field, our work also presents a new benchmark dataset that contains 1200 long videos each with high-quality summaries annotated by professionals. Extensive experiments clearly indicate that our proposed approach sets a new state-of-the-art in video summarization across several benchmarks.

We present DREAM, a novel training framework representing Diffusion Rectification and Estimation-Adaptive Models, requiring minimal code changes (just three lines) yet significantly enhancing the alignment of training with sampling in diffusion models. DREAM features two components: diffusion rectification, which adjusts training to reflect the sampling process, and estimation adaptation, which balances perception against distortion. When applied to image super-resolution (SR), DREAM adeptly navigates the tradeoff between minimizing distortion and preserving high image quality. Experiments demonstrate DREAM's superiority over standard diffusion-based SR methods, showing a $2$ to $3\times $ faster training convergence and a $10$ to $20\times$ reduction in necessary sampling steps to achieve comparable or superior results. We hope DREAM will inspire a rethinking of diffusion model training paradigms.

This work aims to improve the efficiency of text-to-image diffusion models. While diffusion models use computationally expensive UNet-based denoising operations in every generation step, we identify that not all operations are equally relevant for the final output quality. In particular, we observe that UNet layers operating on high-res feature maps are relatively sensitive to small perturbations. In contrast, low-res feature maps influence the semantic layout of the final image and can often be perturbed with no noticeable change in the output. Based on this observation, we propose \emph{Clockwork Diffusion}, a method that periodically reuses computation from preceding denoising steps to approximate low-res feature maps at one or more subsequent steps. For multiple baselines, and for both text-to-image generation and image editing, we demonstrate that Clockwork leads to comparable or improved perceptual scores with drastically reduced computational complexity. As an example, for Stable Diffusion v1.5 with 8 DPM++ steps we save $32\%$ of FLOPs with negligible FID and CLIP change. We release code at https://github.com/Qualcomm-AI-research/clockwork-diffusion

Current instruction-based image editing methods, such as InstructPix2Pix, often fail to produce satisfactory results in complex scenarios due to their dependence on the simple CLIP text encoder in diffusion models. To rectify this, this paper introduces SmartEdit, a novel approach of instruction-based image editing that leverages Multimodal Large Language Models (MLLMs) to enhance its understanding and reasoning capabilities. However, direct integration of these elements still faces challenges in situations requiring complex reasoning. To mitigate this, we propose a Bidirectional Interaction Module (BIM) that enables comprehensive bidirectional information interactions between the input image and the MLLM output. During training, we initially incorporate perception data to boost the perception and understanding capabilities of diffusion models. Subsequently, we demonstrate that a small amount of complex instruction editing data can effectively stimulate SmartEdit's editing capabilities for more complex instructions. We further construct a new evaluation dataset, Reason-Edit, specifically tailored for complex instruction-based image editing. Both quantitative and qualitative results on this evaluation dataset indicate that our SmartEdit surpasses previous methods, paving the way for the practical application of complex instruction-based image editing.

Generative Adversarial Networks (GANs) dominate the research field in image-based virtual try-on, but have not resolved problems such as unnatural deformation of garments and the blurry generation quality. While the generative quality of diffusion models is impressive, achieving controllability poses a significant challenge when applying it to virtual try-on and multiple denoising iterations limit its potential for real-time applications. In this paper, we propose Controllable Accelerated virtual Try-on with Diffusion Model (CAT-DM). To enhance the controllability, a basic diffusion-based virtual try-on network is designed, which utilizes ControlNet to introduce additional control conditions and improves the feature extraction of garment images. In terms of acceleration, CAT-DM initiates a reverse denoising process with an implicit distribution generated by a pre-trained GAN-based model. Compared with previous try-on methods based on diffusion models, CAT-DM not only retains the pattern and texture details of the in-shop garment but also reduces the sampling steps without compromising generation quality. Extensive experiments demonstrate the superiority of CAT-DM against both GAN-based and diffusion-based methods in producing more realistic images and accurately reproducing garment patterns.

Few-shot image generation, as an important yet challenging visual task, still suffers from the trade-off between generation quality and diversity. According to the principle of feature-matching learning, existing fusion-based methods usually fuse different features by using similarity measurements or attention mechanisms, which may match features inaccurately and lead to artifacts in the texture and structure of generated images. In this paper, we propose an exact $\textbf{F}$usion via $\textbf{F}$eature $\textbf{D}$istribution matching $\textbf{G}$enerative $\textbf{A}$dversarial $\textbf{N}$etwork ($\textbf{F2DGAN}$) for few-shot image generation. The rationale behind this is that feature distribution matching is much more reliable than feature matching to explore the statistical characters in image feature space for limited real-world data. To model feature distributions from only a few examples for feature fusion, we design a novel variational feature distribution matching fusion module to perform exact fusion by empirical cumulative distribution functions. Specifically, we employ a variational autoencoder to transform deep image features into distributions and fuse different features exactly by applying histogram matching. Additionally, we formulate two effective losses to guide the matching process for better fitting our fusion strategy. Extensive experiments compared with state-of-the-art methods on three public datasets demonstrate the superiority of F2DGAN for few-shot image generation in terms of generation quality and diversity, and the effectiveness of data augmentation in downstream classification tasks.

Recently, there has been a surge in face personalization techniques, benefiting from the advanced capabilities of pretrained text-to-image diffusion models. Among these, a notable method is Textual Inversion, which generates personalized images by inverting given images into textual embeddings. However, methods based on Textual Inversion still struggle with balancing the trade-off between reconstruction quality and editability. In this study, we examine this issue through the lens of initialization. Upon closely examining traditional initialization methods, we identified a significant disparity between the initial and learned embeddings in terms of both scale and orientation. The scale of the learned embedding can be up to 100 times greater than that of the initial embedding. Such a significant change in the embedding could increase the risk of overfitting, thereby compromising the editability. Driven by this observation, we introduce a novel initialization method, termed Cross Initialization, that significantly narrows the gap between the initial and learned embeddings. This method not only improves both reconstruction and editability but also reduces the optimization steps from 5,000 to 320. Furthermore, we apply a regularization term to keep the learned embedding close to the initial embedding. We show that when combined with Cross Initialization, this regularization term can effectively improve editability. We provide comprehensive empirical evidence to demonstrate the superior performance of our method compared to the baseline methods. Notably, in our experiments, Cross Initialization is the only method that successfully edits an individual's facial expression. Additionally, a fast version of our method allows for capturing an input image in roughly 26 seconds, while surpassing the baseline methods in terms of both reconstruction and editability. Code is available at https://github.com/lyuPang/CrossInitialization.

Generative models are gaining increasing popularity, and the demand for precisely generating images is on the rise. However, generating an image that perfectly aligns with users' expectations is extremely challenging. The shapes of objects, the poses of animals, the structures of landscapes, and more may not match the user's desires, and this applies to real images as well. This is where point-based image editing becomes essential. An excellent image editing method needs to meet the following criteria: user-friendly interaction, high performance, and good generalization capability. Due to the limitations of StyleGAN, DragGAN exhibits limited robustness across diverse scenarios, while DragDiffusion lacks user-friendliness due to the necessity of LoRA fine-tuning and masks. In this paper, we introduce a novel interactive point-based image editing framework, called EasyDrag, that leverages pretrained diffusion models to achieve high-quality editing outcomes and user-friendship. Extensive experimentation demonstrates that our approach surpasses DragDiffusion in terms of both image quality and editing precision for point-based image manipulation tasks.

We present MicroCinema, a straightforward yet effective framework for high-quality and coherent text-to-video generation. Unlike existing approaches that align text prompts with video directly, MicroCinema introduces a Divide-and-Conquer strategy which divides the text-to-video into a two-stage process: text-to-image generation and image\&text-to-video generation. This strategy offers two significant advantages. a) It allows us to take full advantage of the recent advances in text-to-image models, such as Stable Diffusion, Midjourney, and DALLE, to generate photorealistic and highly detailed images. b) Leveraging the generated image, the model can allocate less focus to fine-grained appearance details, prioritizing the efficient learning of motion dynamics. To implement this strategy effectively, we introduce two core designs. First, we propose the Appearance Injection Network, enhancing the preservation of the appearance of the given image. Second, we introduce the Appearance Noise Prior, a novel mechanism aimed at maintaining the capabilities of pre-trained 2D diffusion models. These design elements empower MicroCinema to generate high-quality videos with precise motion, guided by the provided text prompts. Extensive experiments demonstrate the superiority of the proposed framework. Concretely, MicroCinema achieves SOTA zero-shot FVD of 342.86 on UCF-101 and 377.40 on MSR-VTT.

Pretrained diffusion models and their outputs are widely accessible due to their exceptional capacity for synthesizing high-quality images and their open-source nature. The users, however, may face litigation risks owing to the models' tendency to memorize and regurgitate training data during inference. To address this, we introduce Anti-Memorization Guidance (AMG), a novel framework employing three targeted guidance strategies for the main causes of memorization: image and caption duplication, and highly specific user prompts. Consequently, AMG ensures memorization-free outputs while maintaining high image quality and text alignment, leveraging the synergy of its guidance methods, each indispensable in its own right. AMG also features an innovative automatic detection system for potential memorization during each step of inference process, allows selective application of guidance strategies, minimally interfering with the original sampling process to preserve output utility. We applied AMG to pretrained Denoising Diffusion Probabilistic Models (DDPM) and Stable Diffusion across various generation tasks. The results demonstrate that AMG is the first approach to successfully eradicates all instances of memorization with no or marginal impacts on image quality and text-alignment, as evidenced by FID and CLIP scores.

Diffusion Transformer (DiT) has emerged as the new trend of generative diffusion models on image generation. In view of extremely slow convergence in typical DiT, recent breakthroughs have been driven by mask strategy that significantly improves the training efficiency of DiT with additional intra-image contextual learning. Despite this progress, mask strategy still suffers from two inherent limitations: (a) training-inference discrepancy and (b) fuzzy relations between mask reconstruction \& generative diffusion process, resulting in sub-optimal training of DiT. In this work, we address these limitations by novelly unleashing the self-supervised discrimination knowledge to boost DiT training. Technically, we frame our DiT in a teacher-student manner. The teacher-student discriminative pairs are built on the diffusion noises along the same Probability Flow Ordinary Differential Equation (PF-ODE). Instead of applying mask reconstruction loss over both DiT encoder and decoder, we decouple DiT encoder and decoder to separately tackle discriminative and generative objectives. In particular, by encoding discriminative pairs with student and teacher DiT encoders, a new discriminative loss is designed to encourage the inter-image alignment in the self-supervised embedding space. After that, student samples are fed into student DiT decoder to perform the typical generative diffusion task. Extensive experiments are conducted on ImageNet dataset, and our method achieves a competitive balance between training cost and generative capacity.

Vanilla text-to-image diffusion models struggle with generating accurate human images, commonly resulting in inaccurate anatomies such as unnatural postures or disproportionate limbs. Existing methods address this issue mostly by fine-tuning the model with extra images or adding additional controls, human-centric priors such as pose or depth maps, during the image generation phase. This paper explores the integration of these human-centric priors directly into the model fine-tuning stage, essentially eliminating the need for extra conditions at the inference stage. We realize this idea by proposing a human-centric alignment loss to strengthen human-related information from the textual prompts within the cross-attention maps. To ensure semantic detail richness and human structural accuracy during fine-tuning, we introduce scale-aware and step-wise constraints within the diffusion process, according to an in-depth analysis of the cross-attention layer. Extensive experiments show that our method largely improves over state-of-the-art text-to-image models to synthesize high-quality human images based on user-written prompts.

In the digital era, QR codes serve as a linchpin connecting virtual and physical realms. Their pervasive integration across various applications highlights the demand for aesthetically pleasing codes without compromised scannability. However, prevailing methods grapple with the intrinsic challenge of balancing customization and scannability. Notably, stable-diffusion models have ushered in an epoch of high-quality, customizable content generation. This paper introduces Text2QR, a pioneering approach leveraging these advancements to address a fundamental QR code generation challenge: concurrently achieving user-defined aesthetics and scanning robustness. To ensure stable generation of aesthetic QR codes, we introduce the QR Aesthetic Blueprint (QAB) module, generating a blueprint image exerting control over the entire generation process. Subsequently, the Scannability Enhancing Latent Refinement (SELR) process refines the output iteratively in the latent space, enhancing scanning robustness. This innovative approach harnesses the potent generation capabilities of stable-diffusion models, autonomously navigating the delicate trade-off between image aesthetics and QR code scannability. Our experiments demonstrate the seamless fusion of visual appeal with the practical utility of aesthetic QR codes, markedly outperforming prior methods. Source code will be released upon publication.

We present a new method for text-driven motion transfer -- synthesizing a video that complies with an input text prompt describing the target objects and scene while maintaining an input video's motion and scene layout. Prior methods are confined to transferring motion across two subjects within the same or closely related object categories and are applicable for limited domains (e.g., humans). In this work, we consider a significantly more challenging setting in which the target and source objects differ drastically in shape and fine-grained motion characteristics (e.g., translating a jumping dog into a dolphin). To this end, we leverage a pre-trained and fixed text-to-video diffusion model, which provides us with generative and motion priors. The pillar of our method is a new space-time feature loss derived directly from the model. This loss guides the generation process to preserve the overall motion of the input video while complying with the target object in terms of shape and fine-grained motion traits.

Video Frame Interpolation (VFI) has witnessed a surge in popularity due to its abundant downstream applications. Event-based VFI (E-VFI) has recently propelled the advancement of VFI. Thanks to the high temporal resolution benefits, event cameras can bridge the informational void present between successive video frames. Most state-of-the-art E-VFI methodologies follow the conventional VFI paradigm, which pivots on motion estimation between consecutive frames to generate intermediate frames through a process of warping and refinement. However, this reliance engenders a heavy dependency on the quality and consistency of keyframes, rendering these methods susceptible to challenges in extreme real-world scenarios, such as missing moving objects and severe occlusion dilemmas. This study proposes a novel E-VFI framework that directly synthesize intermediate frames leveraging event-based reference, obviating the necessity for explicit motion estimation and substantially enhancing the capacity to handle motion occlusion. Given the sparse and inherently noisy nature of event data, we prioritize the reliability of the event-based reference, leading to the development of an innovative event-aware reconstruction strategy for accurate reference generation. Besides, we implement a bi-directional event-guided alignment from keyframes to the reference using the introduced E-PCD module. Finally, a transformer-based decoder is adopted for prediction refinement. Comprehensive experimental evaluations on both synthetic and real-world datasets underscore the superiority of our approach and its potential to execute high-quality VFI tasks.

Large-scale Text-to-Image (T2I) diffusion models have revolutionized image generation over the last few years. Although owning diverse and high-quality generation capabilities, translating these abilities to fine-grained image editing remains challenging. In this paper, we propose \textbf{DiffEditor} to rectify two weaknesses in existing diffusion-based image editing: (1) in complex scenarios, editing results often lack editing accuracy and exhibit unexpected artifacts; (2) lack of flexibility to harmonize editing operations, e.g., imagine new content. In our solution, we introduce image prompts in fine-grained image editing, cooperating with the text prompt to better describe the editing content. To increase the flexibility while maintaining content consistency, we locally combine stochastic differential equation (SDE) into the ordinary differential equation (ODE) sampling. In addition, we incorporate regional score-based gradient guidance and a time travel strategy into the diffusion sampling, further improving the editing quality. Extensive experiments demonstrate that our method can efficiently achieve state-of-the-art performance on various fine-grained image editing tasks, including editing within a single image (e.g., object moving, resizing, and content dragging) and across images (e.g., appearance replacing and object pasting).

Head avatars animated by visual signals have gained popularity, particularly in cross-driving synthesis where the driver differs from the animated character, a challenging but highly practical approach. The recently presented MegaPortraits model has demonstrated state-of-the-art results in this domain. We conduct a deep examination and evaluation of this model, with a particular focus on its latent space for facial expression descriptors, and uncover several limitations with its ability to express intense face motions. To address these limitations, we propose substantial changes in both training pipeline and model architecture, to introduce our EMOPortraits model, where we:Enhance the model's capability to faithfully support intense, asymmetric face expressions, setting a new state-of-the-art result in the emotion transfer task, surpassing previous methods in both metrics and quality.Incorporate speech-driven mode to our model, achieving top-tier performance in audio-driven facial animation, making it possible to drive source identity through diverse modalities, including visual signal, audio, or a blend of both. Furthermore, we propose a novel multi-view video dataset featuring a wide range of intense and asymmetric facial expressions, filling the gap with absence of such data in existing datasets.

Novel view synthesis of dynamic scenes has been an intriguing yet challenging problem. Despite recent advancements, simultaneously achieving high-resolution photorealistic results, real-time rendering, and compact storage remains a formidable task. To address these challenges, we propose Spacetime Gaussian Feature Splatting as a novel dynamic scene representation, composed of three pivotal components. First, we formulate expressive Spacetime Gaussians by enhancing 3D Gaussians with temporal opacity and parametric motion/rotation. This enables Spacetime Gaussians to capture static, dynamic, as well as transient content within a scene. Second, we introduce splatted feature rendering, which replaces spherical harmonics with neural features. These features facilitate the modeling of view- and time-dependent appearance while maintaining small size. Third, we leverage the guidance of training error and coarse depth to sample new Gaussians in areas that are challenging to converge with existing pipelines. Experiments on several established real-world datasets demonstrate that our method achieves state-of-the-art rendering quality and speed, while retaining compact storage. At 8K resolution, our lite-version model can render at 60 FPS on an Nvidia RTX 4090 GPU.

3D hand-object interaction data is scarce due to the hardware constraints in scaling up the data collection process. In this paper, we propose HOIDiffusion for generating realistic and diverse 3D hand-object interaction data. Our model is a conditional diffusion model that takes both the 3D hand-object geometric structure and text description as inputs for image synthesis. This offers a more controllable and realistic synthesis as we can specify the structure and style inputs in a disentangled manner. HOIDiffusion is trained by leveraging a diffusion model pre-trained on large-scale natural images and a few 3D human demonstrations. Beyond controllable image synthesis, we adopt the generated 3D data for learning 6D object pose estimation and show its effectiveness in improving perception systems.

To synthesize high-fidelity samples, diffusion models typically require auxiliary data to guide the generation process. However, it is impractical to procure the painstaking patch-level annotation effort required in specialized domains like histopathology and satellite imagery; it is often performed by domain experts and involves hundreds of millions of patches. Modern-day self-supervised learning (SSL) representations encode rich semantic and visual information. In this paper, we posit that such representations are expressive enough to act as proxies to fine-grained human labels. We introduce a novel approach that trains diffusion models conditioned on embeddings from SSL. Our diffusion models successfully project these features back to high-quality histopathology and remote sensing images. In addition, we construct larger images by assembling spatially consistent patches inferred from SSL embeddings, preserving long-range dependencies. Augmenting real data by generating variations of real images improves downstream classifier accuracy for patch-level and larger, image-scale classification tasks. Our models are effective even on datasets not encountered during training, demonstrating their robustness and generalizability. Generating images from learned embeddings is agnostic to the source of the embeddings. The SSL embeddings used to generate a large image can either be extracted from a reference image, or sampled from an auxiliary model conditioned on any related modality (e.g. class labels, text, genomic data). As proof of concept, we introduce the text-to-large image synthesis paradigm where we successfully synthesize large pathology and satellite images out of text descriptions.

Recent advances in personalized image generation have enabled pre-trained text-to-image models to learn new concepts from specific image sets. However, these methods often necessitate extensive test-time finetuning for each new concept, leading to inefficiencies in both time and scalability. To address this challenge, we introduce InstantBooth, an innovative approach leveraging existing text-to-image models for instantaneous text-guided image personalization, eliminating the need for test-time finetuning. This efficiency is achieved through two primary innovations. Firstly, we utilize an image encoder that transforms input images into a global embedding to grasp the general concept. Secondly, we integrate new adapter layers into the pre-trained model, enhancing its ability to capture intricate identity details while maintaining language coherence. Significantly, our model is trained exclusively on text-image pairs, without reliance on concept-specific paired images. When benchmarked against existing finetuning-based personalization techniques like DreamBooth and Textual-Inversion, InstantBooth not only shows comparable proficiency in aligning language with image, maintaining image quality, and preserving the identity but also boasts a 100-fold increase in generation speed. Project Page: https://jshi31.github.io/InstantBooth/

We present TokenCompose, a Latent Diffusion Model for text-to-image generation that achieves enhanced consistency between user-specified text prompts and model-generated images. Despite its tremendous success, the standard denoising process in the Latent Diffusion Model takes text prompts as conditions only, absent explicit constraint for the consistency between the text prompts and the image contents, leading to unsatisfactory results for composing multiple object categories. Our proposed TokenCompose aims to improve multi-category instance composition by introducing the token-wise consistency terms between the image content and object segmentation maps in the finetuning stage. TokenCompose can be applied directly to the existing training pipeline of text-conditioned diffusion models without extra human labeling information. By finetuning Stable Diffusion with our approach, the model exhibits significant improvements in multi-category instance composition and enhanced photorealism for its generated images.

Shape and geometric patterns are essential in defining stylistic identity. However, current 3D style transfer methods predominantly focus on transferring colors and textures, often overlooking geometric aspects. In this paper, we introduce Geometry Transfer, a novel method that leverages geometric deformation for 3D style transfer. This technique employs depth maps to extract a style guide, subsequently applied to stylize the geometry of radiance fields. Moreover, we propose new techniques that utilize geometric cues from the 3D scene, thereby enhancing aesthetic expressiveness and more accurately reflecting intended styles. Our extensive experiments show that Geometry Transfer enables a broader and more expressive range of stylizations, thereby significantly expanding the scope of 3D style transfer.

Text-guided diffusion models have revolutionized image and video generation and have also been successfully used for optimization-based 3D object synthesis. Here, we instead focus on the underexplored text-to-4D setting and synthesize dynamic, animated 3D objects using score distillation methods with an additional temporal dimension. Compared to previous work, we pursue a novel compositional generation-based approach, and combine text-to-image, text-to-video, and 3D-aware multiview diffusion models to provide feedback during 4D object optimization, thereby simultaneously enforcing temporal consistency, high-quality visual appearance and realistic geometry. Our method, called Align Your Gaussians (AYG), leverages dynamic 3D Gaussian Splatting with deformation fields as 4D representation. Crucial to AYG is a novel method to regularize the distribution of the moving 3D Gaussians and thereby stabilize the optimization and induce motion. We also propose a motion amplification mechanism as well as a new autoregressive synthesis scheme to generate and combine multiple 4D sequences for longer generation. These techniques allow us to synthesize vivid dynamic scenes, outperform previous work qualitatively and quantitatively and achieve state-of-the-art text-to-4D performance. Due to the Gaussian 4D representation, different 4D animations can be seamlessly combined, as we demonstrate. AYG opens up promising avenues for animation, simulation and digital content creation as well as synthetic data generation.

While large-scale pre-trained text-to-image models can synthesize diverse and high-quality human-centered images, novel challenges arise with a nuanced task of "identity fine editing" -- precisely modifying specific features of a subject while maintaining its inherent identity and context. Existing personalization methods either require time-consuming optimization or learning additional encoders, adepting in "identity re-contextualization". However, they often struggle with detailed and sensitive tasks like human face editing.To address these challenges, we introduce DreamSalon, a noise-guided, staged-editing framework, uniquely focusing on detailed image manipulations and identity-context preservation. By discerning editing and boosting stages via the frequency and gradient of predicted noises, DreamSalon first performs detailed manipulations on specific features in the editing stage, guided by high-frequency information, and then employs stochastic denoising in the boosting stage to improve image quality. For more precise editing, DreamSalon semantically mixes source and target textual prompts, guided by differences in their embedding covariances, to direct the model's focus on specific manipulation areas.Our experiments demonstrate DreamSalon's ability to efficiently and faithfully edit fine details on human faces, outperforming existing methods both qualitatively and quantitatively.

Video-P2P is the first framework for real-world video editing with cross-attention control. While attention control has proven effective for image editing with pre-trained image generation models, there are currently no large-scale video generation models publicly available. Video-P2P addresses this limitation by adapting an image generation diffusion model to complete various video editing tasks. Specifically, we propose to first tune a Text-to-Set (T2S) model to complete an approximate inversion and then optimize a shared unconditional embedding to achieve accurate video inversion with a small memory cost. We further prove that it is crucial for consistent video editing. For attention control, we introduce a novel decoupled-guidance strategy, which uses different guidance strategies for the source and target prompts. The optimized unconditional embedding for the source prompt improves reconstruction ability, while an initialized unconditional embedding for the target prompt enhances editability. Incorporating the attention maps of these two branches enables detailed editing. These technical designs enable various text-driven editing applications, including word swap, prompt refinement, and attention re-weighting. Video-P2P works well on real-world videos for generating new characters while optimally preserving their original poses and scenes. It significantly outperforms previous approaches.

Generative image editing has recently witnessed extremely fast-paced growth.Some works use high-level conditioning such as text, while others use low-levelconditioning. Nevertheless, most of them lack fine-grained control over the properties of the different objects present in the image, i.e. object-level image editing. In this work, we tackle the task by perceiving the images as an amalgamation ofvarious objects and aim to control the properties of each object in a fine-grainedmanner. Out of these properties, we identify structure and appearance as the mostintuitive to understand and useful for editing purposes. We propose PAIR Diffusion, a generic framework that can enable a diffusion model to control the structure and appearance properties of each object in the image. We show that having controlover the properties of each object in an image leads to comprehensive editingcapabilities. Our framework allows for various object-level editing operations onreal images such as reference image-based appearance editing, free-form shapeediting, adding objects, and variations. Thanks to our design, we do not requireany inversion step. Additionally, we propose multimodal classifier-free guidancewhich enables editing images using both reference images and text when usingour approach with foundational diffusion models. We validate the above claimsby extensively evaluating our framework on both unconditional and foundationaldiffusion models.

This work introduces ArtAdapter, a transformative text-to-image (T2I) style transfer framework that transcends traditional limitations of color, brushstrokes, and object shape, capturing high-level style elements such as composition and distinctive artistic expression.The integration of a multi-level style encoder with our proposed explicit adaptation mechanism enables ArtAdapter to achieve unprecedented fidelity in style transfer, ensuring close alignment with textual descriptions.Additionally, the incorporation of an Auxiliary Content Adapter (ACA) effectively separates content from style, alleviating the borrowing of content from style references.Moreover, our novel fast finetuning approach could further enhance zero-shot style representation while mitigating the risk of overfitting.Comprehensive evaluations confirm that ArtAdapter surpasses current state-of-the-art methods.

In this paper, we democratise caricature generation, empowering individuals to effortlessly craft personalised caricatures with just a photo and a conceptual sketch. Our objective is to strike a delicate balance between abstraction and identity, while preserving the creativity and subjectivity inherent in a sketch. To achieve this, we present Explicit Rank-1 Model Editing alongside single-image personalisation, selectively applying nuanced edits to cross-attention layers for a seamless merge of identity and style. Additionally, we propose Random Mask Reconstruction to enhance robustness, directing the model to focus on distinctive identity and style features. Crucially, our aim is not to replace artists but to eliminate accessibility barriers, allowing enthusiasts to engage in the artistry.

Recent advances in text-to-image generation have made remarkable progress in synthesizing realistic human photos conditioned on given text prompts. However, existing personalized generation methods cannot simultaneously satisfy the requirements of high efficiency, promising identity (ID) fidelity, and flexible text controllability. In this work, we introduce PhotoMaker, an efficient personalized text-to-image generation method, which mainly encodes an arbitrary number of input ID images into a stack ID embedding for preserving ID information.Such an embedding also empowers our method to be applied in many interesting scenarios, such as when replacing the corresponding class word and when combining the characteristics of different identities. Besides, to better drive the training of our PhotoMaker, we propose an ID-oriented data creation pipeline to assemble the training data. Under the nourishment of the dataset constructed through the proposed pipeline, our PhotoMaker demonstrates comparable performance to test-time fine-tuning-based methods, yet provides significant speed improvements, strong generalization capabilities, and a wide range of applications.

Diffusion models have achieved remarkable success in generating high-quality, diverse, and creative images. However, in text-based image generation, they often struggle to accurately capture the intended meaning of the text. For instance, a specified object might not be generated, or an adjective might incorrectly alter unintended objects. Moreover, we found that relationships indicating possession between objects are frequently overlooked. Despite the diversity of users' intentions in text, existing methods often focus on only some aspects of these intentions. In this paper, we propose Predicated Diffusion, a unified framework designed to more effectively express users' intentions. It represents the intended meaning as propositions using predicate logic and treats the pixels in attention maps as fuzzy predicates. This approach provides a differentiable loss function that offers guidance for the image generation process to better fulfill the propositions. Comparative evaluations with existing methods demonstrated that Predicated Diffusion excels in generating images faithful to various text prompts, while maintaining high image quality, as validated by human evaluators and pretrained image-text models.

While AI-generated content has garnered significant attention, achieving photo-realistic video synthesis remains a formidable challenge. Despite the promising advances in diffusion models for video generation quality, the complex model architecture and substantial computational demands for both training and inference create a significant gap between these models and real-world applications.In this paper, we introduce SNED, the superposition network architecture search for efficient video diffusion model. Our framework employs a supernet training paradigm that targets various model cost and resolution options using a weight-sharing method. Additionally, we introduce a systematic fast training optimization strategy, including methods, such as supernet training sampling warm-up and image-based diffusion model transfer learning. To showcase the flexibility of our method, we conduct experiments involving both pixel-space and latent-space video diffusion models. The results demonstrate that our framework consistently produces comparable results across different model options with high efficiency. According to the experiment for the pixel-space video diffusion model, we can achieve consistent video generation results simultaneously across 64$\times$64 to 256$\times$256 resolutions with a large model size range from 640M to 1.6B number of parameters for pixel-space video diffusion model.

Recent advances in text-to-video generation have demonstrated the utility of powerful diffusion models. Nevertheless, the problem is not trivial when shaping diffusion models to animate static image (i.e., image-to-video generation). The difficulty originates from the aspect that the diffusion process of subsequent animated frames should not only preserve the faithful alignment with the given image but also pursue temporal coherence among adjacent frames. To alleviate this, we present TRIP, a new recipe of image-to-video diffusion paradigm that pivots on image noise prior derived from static image to jointly trigger inter-frame relational reasoning and ease the coherent temporal modeling via temporal residual learning. Technically, the image noise prior is first attained through one-step backward diffusion process based on both static image and noised video latent codes. Next, TRIP executes a residual-like dual-path scheme for noise prediction: 1) a shortcut path that directly takes image noise prior as the reference noise of each frame to amplify the alignment between the first frame and subsequent frames; 2) a residual path that employs 3D-UNet over noised video and static image latent codes to enable inter-frame relational reasoning, thereby easing the learning of the residual noise for each frame. Furthermore, both reference and residual noise of each frame are dynamically merged via attention mechanism for final video generation. Extensive experiments on WebVid-10M, DTDB and MSR-VTT datasets demonstrate the effectiveness of our TRIP for image-to-video generation.

Text-to-image (T2I) research has grown explosively in the past year, owing to the large-scale pre-trained diffusion models and many emerging personalization and editing approaches. Yet, one pain point persists: the text prompt engineering, and searching high-quality text prompts for customized results is more art than science. Moreover, as commonly argued: “an image is worth a thousand words” - the attempt to describe a desired image with texts often ends up being ambiguous and cannot comprehensively cover delicate visual details, hence necessitating more additional controls from the visual domain. In this paper, we take a bold step forward: taking “Text” out of a pretrained T2I diffusion model, to reduce the burdensome prompt engineering efforts for users. Our proposed framework, Prompt-Free Diffusion, relies on only visual inputs to generate new images: it takes a reference image as “context”, an optional image structural conditioning, and an initial noise, with absolutely no text prompt. The core architecture behind the scene is Semantic Context Encoder(SeeCoder), substituting the commonly used CLIP-based or LLM-based text encoder. The reusability of SeeCoder also makes it a convenient drop-in component: one can also pre-train a SeeCoder in one T2I model and reuse it for another. Through extensive experiments, Prompt-Free Diffusion is experimentally found to (i) outperform prior exemplar-based image synthesis approaches; (ii) perform on par with state-of-the-art T2I models using prompts following the best practice; and (iii) be naturally extensible to other downstream applications such as anime figure generation and virtual try-on, with promising quality. Our code and models will be open-sourced.

The diffusion-based text-to-image model harbors immense potential in transferring reference style. However, current encoder-based approaches significantly impair the text controllability of text-to-image models while transferring styles. In this paper, we introduce $\textit{DEADiff}$ to address this issue using the following two strategies: 1) a mechanism to decouple the style and semantics of reference images. The decoupled feature representations are first extracted by Q-Formers which are instructed by different text descriptions. Then they are injected into mutually exclusive subsets of cross-attention layers for better disentanglement. 2) A non-reconstructive learning method. The Q-Formers are trained using paired images rather than the identical target, in which the reference image and the ground-truth image are with the same style or semantics. We show that DEADiff attains the best visual stylization results and optimal balance between the text controllability inherent in the text-to-image model and style similarity to the reference image, as demonstrated both quantitatively and qualitatively.

The remarkable efficacy of text-to-image diffusion models has motivated extensive exploration of their potential application in video domains. Zero-shot methods seek to extend image diffusion models to videos without necessitating model training. Recent methods mainly focus on incorporating inter-frame correspondence into attention mechanisms. However, the soft constraint imposed on determining where to attend to valid features can sometimes be insufficient, resulting in temporal inconsistency. In this paper, we introduce FRESCO, intra-frame correspondence alongside inter-frame correspondence to establish a more robust spatial-temporal constraint. This enhancement ensures a more consistent transformation of semantically similar content across frames. Beyond mere attention guidance, our approach involves an explicit update of features to achieve high spatial-temporal consistency with the input video, significantly improving the visual coherence of the resulting translated videos. Extensive experiments demonstrate the effectiveness of our proposed framework in producing high-quality, coherent videos, marking a notable improvement over existing zero-shot methods.

Despite diffusion models' superior capabilities in modeling complex distributions, there are still non-trivial distributional discrepancies between generated and ground-truth images, which has resulted in several notable problems in image generation, including missing object errors in text-to-image generation and low image quality. Existing methods that attempt to address these problems mostly do not tend to address the fundamental cause behind these problems, which is the distributional discrepancies, and hence achieve sub-optimal results. In this paper, we propose a particle filtering framework that can effectively address both problems by explicitly reducing the distributional discrepancies. Specifically, our method relies on a set of external guidance, including a small set of real images and a pre-trained object detector, to gauge the distribution gap, and then design the resampling weight accordingly to correct the gap. Experiments show that our methods can effectively correct missing object errors and improve image quality in various image generation tasks. Notably, our method outperforms the existing strongest baseline by 5% in object occurrence and 1.0 in FID on MS-COCO.

Recent advancements in text-to-image technology have significantly advanced the field of image customization. Among various applications, the task of customizing diverse scenes for user-specified composited elements holds great application value but has not been extensively explored. Addressing this gap, we propose AnyScene, a specialized framework designed to create varied scenes from composited foreground using textual prompts. AnyScene addresses the primary challenges inherent in existing methods, particularly scene disharmony due to a lack of foreground semantic understanding and distortion of foreground elements. Specifically, we develop a foreground injection module that guides a pre-trained diffusion model to generate cohesive scenes in visual harmony with the provided foreground. To enhance robust generation, we implement a layout control strategy that prevents distortions of foreground elements. Furthermore, an efficient image blending mechanism seamlessly reintegrates foreground details into the generated scenes, producing outputs with overall visual harmony and precise foreground details. In addition, we propose a new benchmark and a series of quantitative metrics to evaluate this proposed image customization task. Extensive experimental results demonstrate the effectiveness of AnyScene, which confirms its potential in various applications.

Recent advances in the diffusion models have significantly improved text-to-image generation. However, generating videos from text is a more challenging task than generating images from text, due to the much larger dataset and higher computational cost required. Most existing video generation methods use either a 3D U-Net architecture that considers the temporal dimension or autoregressive generation. These methods require large datasets and are limited in terms of computational costs compared to text-to-image generation. To tackle these challenges, we propose a simple but effective novel grid diffusion for text-to-video generation without temporal dimension in architecture and a large text-video paired dataset. We can generate a high-quality video using a fixed amount of GPU memory regardless of the number of frames by representing the video as a grid image. Additionally, since our method reduces the dimensions of the video to the dimensions of the image, various image-based methods can be applied to videos, such as text-guided video manipulation from image manipulation. Our proposed method outperforms the existing methods in both quantitative and qualitative evaluations, demonstrating the suitability of our model for real-world video generation.

Recent advances in generative AI have unveiled significant potential for the creation of 3D content. However, current methods either apply a pre-trained 2D diffusion model with the time-consuming score distillation sampling (SDS), or a direct 3D diffusion model trained on limited 3D data losing generation diversity. In this work, we approach the problem by employing a multi-view 2.5D diffusion fine-tuned from a pre-trained 2D diffusion model. The multi-view 2.5D diffusion directly models the structural distribution of 3D data, while still maintaining the strong generalization ability of the original 2D diffusion model, filling the gap between 2D diffusion-based and direct 3D diffusion-based methods for 3D content generation. During inference, multi-view normal maps are generated using the 2.5D diffusion, and a novel differentiable rasterization scheme is introduced to fuse the almost consistent multi-view normal maps into a consistent 3D model. We further design a normal-conditioned multi-view image generation module for fast appearance generation given the 3D geometry. Our method is a one-pass diffusion process and does not require any SDS optimization as post-processing. We demonstrate through extensive experiments that, our direct 2.5D generation with the specially-designed fusion scheme can achieve diverse, mode-seeking-free, and high-fidelity 3D content generation in only 10 seconds.

With the advancement of generative models, the assessment of generated images becomes more and more important. Previous methods measure distances between features of reference and generated images from trained vision models. In this paper, we conduct an extensive investigation into the relationship between the representation space and input space around generated images. We first propose two measures related to the presence of unnatural elements within images: complexity, which indicates how non-linear the representation space is, and vulnerability, which is related to how easily the extracted feature changes by adversarial input changes. Based on these, we introduce a new metric to evaluating image-generative models called anomaly score (AS). Moreover, we propose AS-i (anomaly score for individual images) that can effectively evaluate generated images individually. Experimental results demonstrate the validity of the proposed approach.

Large-scale Text-to-Image (T2I) models have rapidly gained prominence across creative fields, generating visually compelling outputs from textual prompts. However, controlling these models to ensure consistent style remains challenging, with existing methods necessitating fine-tuning and manual intervention to disentangle content and style. In this paper, we introduce StyleAligned, a novel technique designed to establish style alignment among a series of generated images. By employing minimal `attention sharing' during the diffusion process, our method maintains style consistency across images within T2I models. This approach allows for the creation of style-consistent images using a reference style through a straightforward inversion operation. Our method's evaluation across diverse styles and text prompts demonstrates high-quality synthesis and fidelity, underscoring its efficacy in achieving consistent style across various inputs.

We present Zero-Painter, a novel training-free framework for layout-conditional text-to-image synthesis that facilitates the creation of detailed and controlled imagery from textual prompts. Our method utilizes object masks and individual descriptions, coupled with a global text prompt, to generate images with high fidelity. Zero-Painter employs a two-stage process involving our novel Prompt-Adjusted Cross-Attention (PACA) and Region-Grouped Cross-Attention (ReGCA) blocks, ensuring precise alignment of generated objects with textual prompts and mask shapes. Our extensive experiments demonstrate that Zero-Painter surpasses current state-of-the-art methods in preserving textual details and adhering to mask shapes. We will make the codes and the models publicly available.

We introduce X-Adapter, a universal upgrader to enable the pretrained plug-and-play modules (e.g. ControlNet, LoRA) to work directly with the upgraded text-to-image diffusion model (e.g. SDXL) without further retraining. We achieve this goal by training an additional network to control the frozen upgraded model with the new text-image data pairs. In detail, X-Adapter keeps a frozen copy of the old model to preserve the connectors of different plugins. Additionally, X-Adapter adds trainable mapping layers that bridge the decoders from models of different versions for feature remapping. The remapped features will be used as guidance for the upgraded model. To enhance the guidance ability of X-Adapter, we employ a null-text training strategy for the upgraded model. After training, we also introduce a two-stage denoising strategy to align the initial latents of X-Adapter and the upgraded model. Thanks to our strategies, X-Adapter demonstrates universal compatibility with various plugins and also enables plugins of different versions to work together, thereby expanding the functionalities of diffusion community. To verify the effectiveness of the proposed method, we conduct extensive experiments and the results show that X-Adapter may facilitate wider application in the upgraded foundational diffusion model. Project page at: https://showlab.github.io/X-Adapter.

Controllable generation of 3D assets is important for many practical applications for content creation in movies, games and engineering, as well as in AR/VR. Recently, diffusion models have shown remarkable results in generation quality of 3D objects. However, none of the existing models enable disentangled generation to control the shape and appearance separately. For the first time, we present a suitable representation for 3D diffusion models to enable such disentanglement by introducing a hybrid point cloud and neural radiance field approach. We model a diffusion process over point positions jointly with a high-dimensional feature space for a local density and radiance decoder. While the point positions represent the coarse shape of the object, the point features allow modeling the geometry and appearance details. This disentanglement enables us to sample both independently and therefore to control both separately. Our approach sets a new state of the art in generation compared to previous disentanglement-capable methods by reduced FID scores of 30-90% and is on-par with other non disentanglement-capable state-of-the art methods.

Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines.

In this work, we present Vlogger, a generic AI system for generating a minute-level video blog (i.e., vlog) of user descriptions. Different from short videos with a few seconds, vlog often contains a complex storyline with diversified scenes, which is challenging for most existing video generation approaches. To break through this bottleneck, our Vlogger smartly leverages Large Language Model (LLM) as Director and decomposes a long video generation task of vlog into four key stages, where we invoke various foundation models to play the critical roles of vlog professionals, including (1) Script, (2) Actor, (3) ShowMaker, and (4) Voicer. With such a design of mimicking human beings, our Vlogger can generate vlogs through explainable cooperation of top-down planning and bottom-up shooting. More over, we introduce a novel video diffusion model, ShowMaker, which serves as a videographer in our Vlogger for generating the video snippet of each shooting scene. By incorporating Script and Actor attentively as textual and visual prompts, it can effectively enhance spatial-temporal coherence in the snippet. Besides, we design a concise mixed training paradigm for ShowMaker, boosting its capacity for both T2V generation and prediction. Finally, the extensive experiments show that our method achieves state-of-the-art performance on zero-shot T2V generation and prediction tasks. More importantly, Vlogger can generate over 5-minute vlogs from open-world descriptions, without loss of video coherence on script and actor.

The goal of this work is to simultaneously generate natural talking faces and speech outputs from text. We achieve this by integrating Talking Face Generation (TFG) and Text-to-Speech (TTS) systems into a unified framework.We address the main challenges of each task: (1) generating a range of head poses representative of real-world scenarios, and (2) ensuring voice consistency despite variations in facial motion for the same identity.To tackle these issues, we introduce a motion sampler based on conditional flow matching, which is capable of high-quality motion code generation in an efficient way.Moreover, we introduce a novel conditioning method for the TTS system, which utilises motion-removed features from the TFG model to yield uniform speech outputs.Our extensive experiments demonstrate that our method effectively creates natural-looking talking faces and speech that accurately match the input text. To our knowledge, this is the first effort to build a multimodal synthesis system that can generalise to unseen identities.

Text-guided image editing finds applications in various creative and practical fields. While recent strides in image generation have advanced the field, they often grapple with the dual challenges of coherent image transformation and context preservation. In response, our work introduces prompt augmentation, a method amplifying a single input prompt into a palette of target prompts, strengthening textual context and enabling localised image editing. Specifically, we utilise the augmented prompts to delineate the intended manipulation area. We propose a Contrastive Loss tailored to driving effective image editing by displacing edited areas and drawing preserved regions closer. Acknowledging the continuous nature of image manipulations, we further refine our approach by incorporating the similarity concept, creating a Soft Contrastive Loss. The new losses are incorporated to the diffusion model in an end-to-end manner, demonstrating improved image editing results on public datasets and generated images over the baseline and competitive results against state-of-the-art approaches.

Accurate and controllable image editing is a challenging task that has attracted significant attention recently. Notably, DragGAN developed by Pan et al (2023) is an interactive point-based image editing framework that achieves impressive editing results with pixel-level precision. However, due to its reliance on generative adversarial networks (GANs), its generality is limited by the capacity of pretrained GAN models. In this work, we extend this editing framework to diffusion models and propose a novel approach DragDiffusion. By harnessing large-scale pretrained diffusion models, we greatly enhance the applicability of interactive point-based editing on both real and diffusion-generated images. Unlike other diffusion-based editing methods that provide guidance on diffusion latents of multiple time steps, our approach achieves efficient yet accurate spatial control by optimizing the latent of only one time step. This novel design is motivated by our observations that UNet features at a specific time step provides sufficient semantic and geometric information to support the drag-based editing. Moreover, we introduce two additional techniques, namely identity-preserving fine-tuning and reference-latent-control, to further preserve the identity of the original image. Lastly, we present a challenging benchmark dataset called DragBench---the first benchmark to evaluate the performance of interactive point-based image editing methods. Experiments across a wide range of challenging cases (e.g., images with multiple objects, diverse object categories, various styles, etc.) demonstrate the versatility and generality of DragDiffusion. The code and the DragBench dataset will be released.

Creating high-dynamic videos such as motion-rich actions and sophisticated visual effects poses a significant challenge in the field of artificial intelligence. Unfortunately, current state-of-the-art video generation methods, primarily focusing on text-to-video generation, tend to produce video clips with minimal motions despite maintaining high fidelity. We argue that relying solely on text instructions is insufficient and suboptimal for video generation. In this paper, we introduce PixelDance, a novel approach based on diffusion models that incorporates image instructions for both the first and last frames in conjunction with text instructions for video generation. Comprehensive experimental results demonstrate that PixelDance trained with public data exhibits significantly better proficiency in synthesizing videos with complex scenes and intricate motions, setting a new standard for video generation.

Text-to-image diffusion models have recently received increasing interest for their astonishing ability to produce high-fidelity images from solely text inputs. Subsequent research efforts aim to exploit and apply their capabilities to real image editing. However, existing image-to-image methods are often inefficient, imprecise, and of limited versatility. They either require time-consuming fine-tuning, deviate unnecessarily strongly from the input image, and/or lack support for multiple, simultaneous edits. To address these issues, we introduce LEDITS++, an efficient yet versatile and precise textual image manipulation technique. LEDITS++’s novel inversion approach requires no tuning nor optimization and produces high-fidelity results with a few diffusion steps. Second, our methodology supports multiple simultaneous edits and is architecture-agnostic. Third, we use a novel implicit masking technique that limits changes to relevant image regions. We propose the novel TEdBench++ benchmark as part of our exhaustive evaluation. Our results demonstrate the capabilities of LEDITS++ and its improvements over previous methods.

Instruction-based image editing holds immense potential for a variety of applications, as it enables users to perform any editing operation using a natural language instruction. However, current models in this domain often struggle with accurately executing user instructions. We present IEdit, a multi-task image editing model which sets state-of-the-art results in instruction-based image editing. To develop IEdit we train it to multi-task across an unprecedented range of tasks, such as region-based editing, free-form editing, and Computer Vision tasks, all of which are formulated as generative tasks. Additionally, to enhance IEdit's multi-task learning abilities, we provide it with learned task embeddings which guide the generation process towards the correct edit type. Both these elements are essential for IEdit's outstanding performance. Furthermore, we show that IEdit can generalize to new tasks, such as image inpainting, super-resolution, and compositions of editing tasks, with just a few labeled examples. This capability offers a significant advantage in scenarios where high-quality samples are scarce. Lastly, to facilitate a more rigorous and informed assessment of instructable image editing models, we release a new challenging and versatile benchmark that includes seven different image editing tasks.

While there has been significant progress in customizing text-to-image generation models, generating images that combine multiple personalized concepts remains challenging. In this work, we introduce Concept Weaver, a method for composing customized text-to-image diffusion models at inference time. Specifically, the method breaks the process into two steps: creating a template image aligned with the semantics of input prompts, and then personalizing the template using a concept fusion strategy. The fusion strategy incorporates the appearance of the target concepts into the template image while retaining its structural details. The results indicate that our method can generate multiple custom concepts with higher identity fidelity compared to alternative approaches. Furthermore, the method is shown to seamlessly handle more than two concepts and closely follow the semantic meaning of the input prompt without blending appearances across different subjects.

Though diffusion models excel in image generation, their step-by-step denoising leads to slow generation speeds. Consistency training addresses this issue with single-step sampling but often produces lower-quality generations and requires high training costs. In this paper, we show that optimizing consistency training loss minimizes the Wasserstein distance between target and generated distributions. As timestep increases, the upper bound accumulates previous consistency training losses. Therefore, larger batch sizes are needed to reduce both current and accumulated losses. We propose Adversarial Consistency Training (ACT), which directly minimizes the Jensen-Shannon (JS) divergence between distributions at each timestep using a discriminator. Theoretically, ACT enhances generation quality, and convergence. By incorporating a discriminator into the consistency training framework, our method achieves improved FID scores on CIFAR10 and ImageNet 64$\times$64 and LSUN Cat 256$\times$256 datasets, retains zero-shot image inpainting capabilities, and uses less than $1/6$ of the original batch size and fewer than $1/2$ of the model parameters and training steps compared to the baseline method, this leads to a substantial reduction in resource consumption. Our code is available: https://github.com/kong13661/ACT

In this paper, we propose a 3D geometry-aware deformable Gaussian Splatting method for dynamic view synthesis. Existing neural radiance fields (NeRF) based solutions learn the deformation in an implicit manner, which cannot incorporate 3D scene geometry. Therefore, the learned deformation is not necessarily geometrically coherent, which results in unsatisfactory dynamic view synthesis and 3D dynamic reconstruction. Recently, 3D Gaussian Splatting provides a new representation of the 3D scene, building upon which the 3D geometry could be exploited in learning the complex 3D deformation. Specifically, the scenes are represented as a collection of 3D Gaussian, where each 3D Gaussian is optimized to move and rotate over time to model the deformation. To enforce the 3D scene geometry constraint during deformation, we explicitly extract 3D geometry features and integrate them in learning the 3D deformation. In this way, our solution achieves 3D geometry-aware deformation modeling, which enables improved dynamic view synthesis and 3D dynamic reconstruction. Extensive experimental results on both synthetic and real datasets prove the superiority of our solution, which achieves new state-of-the-art performance. The project is available at https://npucvr.github.io/GaGS/.

Diffusion models have recently brought a powerful revolution in image generation. Despite showing impressive generative capabilities, most of these models rely on the current sample to denoise the next one, possibly resulting in denoising instability. In this paper, we reinterpret the iterative denoising process as model optimization and leverage a moving average mechanism to ensemble all the prior samples. Instead of simply applying moving average to the denoised samples at different timesteps, we first map the denoised samples to data space and then perform moving average to avoid distribution shift across timesteps. In view that diffusion models evolve the recovery from low-frequency components to high-frequency details, we further decompose the samples into different frequency components and execute moving average separately on each component. We name the complete approach ``Moving Average Sampling in Frequency domain (MASF)''. MASF could be seamlessly integrated into mainstream pre-trained diffusion models and sampling schedules. Extensive experiments on both unconditional and conditional diffusion models demonstrate that our MASF leads to superior performances than its baselines, with almost neglectable additional complexity cost.

Layout-aware text-to-image generation is a task to generate multi-object images that reflect the layout condition as well as the text condition. The current layout-aware text-to-image diffusion models still have several issues, including mismatches between the text and layout conditions and quality degradation of generated images. This paper proposes a novel layout-aware text-to-image diffusion model called NoiseCollage to tackle these issues. During the denoising process, NoiseCollage independently estimates noises for individual objects and then crops and merges them into a single noise. This operation helps avoid condition mismatches; in other words, it can put the right objects in the right places. Qualitative and quantitative evaluations show that NoiseCollage outperforms several state-of-the-art models. These successful results indicate that the crop-and-merge operation of noises is a reasonable strategy to control image generation. We also show that NoiseCollage can be integrated with ControlNet to use edges, sketches, and pose skeletons as additional conditions. Experimental results show that this integration boosts the layout accuracy of ControlNet.The code is available at https://github.com/*.

Neural Radiance Fields (NeRFs) have shown remarkable success in synthesizing photorealistic views from multi-view images of static scenes, but face challenges in dynamic, real-world environments with distractors like moving objects, shadows, and lighting changes. Existing methods manage controlled environments and low occlusion ratios but fall short in render quality, especially under high occlusion scenarios. In this paper, we introduce NeRF On-the-go, a simple yet effective approach that enables the robust synthesis of novel views in complex, in-the-wild scenes from only casually captured image sequences. Delving into uncertainty, our method not only efficiently eliminates distractors, even when they are predominant in captures, but also achieves a notably faster convergence speed. Through comprehensive experiments on various scenes, our method demonstrates a significant improvement over state-of-the-art techniques. This advancement opens new avenues for NeRF in diverse and dynamic real-world applications.

Using reinforcement learning with human feedback (RLHF) has shown significant promise in fine-tuning diffusion models. Previous methods start by training a reward model that aligns with human preferences, then leverage RL techniques to fine-tune the underlying models. However, crafting an efficient reward model demands extensive datasets, optimal architecture, and manual hyperparameter tuning, making the process both time and cost-intensive. The direct preference optimization (DPO) method, effective in fine-tuning large language models, eliminates the necessity for a reward model. However, the extensive GPU memory requirement of the diffusion model's denoising process hinders the direct application of the DPO method. To address this issue, we introduce the Direct Preference for Denoising Diffusion Policy Optimization (D3PO) method to directly fine-tune diffusion models. The theoretical analysis demonstrates that although D3PO omits training a reward model, it effectively functions as the optimal reward model trained using human feedback data to guide the learning process. This approach requires no training of a reward model, proving to be more direct, cost-effective, and minimizing computational overhead. In experiments, our method uses the relative scale of objectives as a proxy for human preference, delivering comparable results to methods using ground-truth rewards. Moreover, D3PO demonstrates the ability to reduce image distortion rates and generate safer images, overcoming challenges lacking robust reward models. Our code is publicly available at https://github.com/yk7333/D3PO.

Recently, we have witnessed the explosive growth of various volumetric representations in modeling animatable head avatars.However, due to the diversity of frameworks, there is no practical method to support high-level applications like 3D head avatar editing across different representations.In this paper, we propose a generic avatar editing approach that can be universally applied to various 3DMM-driving volumetric head avatars.To achieve this goal, we design a novel expression-aware modification generative model, which enableslift 2D editing from a single image to a consistent 3D modification field.To ensure the effectiveness of the generative modification process,we develop several techniques, including an expression-dependent modification distillation scheme to draw knowledge from the large-scale head avatar model and 2D facial texture editing tools, implicit latent space guidance to enhance the convergence of training, and a segmentation-based loss reweight strategy for fine-grained texture inversion.Extensive experiments demonstrate that our method delivers high-quality and consistent editing results across multiple expressions and viewpoints.

Layout design in floorplan traditionally involves a labor-intensive iterative process for human designers. Recent advancements in generative modeling present a transformative potential to automate layout creation. However, prevalent models typically neglect crucial guidance from users, particularly their incomplete design ideas at the early stage. To address these limitations, we propose MaskPLAN, a novel user-guided generative design model formulated with Graph-structured Masked AutoEncoders (GMAE). Throughout its training phase, the layout attributes undergo stochastic masking to mimic partial input from users. During inference, layout designs are procedurally reconstructed via generative transformers. MaskPLAN incorporates the partial input as a global conditional prior, enabling users to turn incomplete design ideas into full layouts, which is a key part of real-world floorplan design. Notably, our proposed model offers an extensive range of adaptable user engagements, while also demonstrating superior performance to state-of-the-art methods in both quantitative and qualitative evaluations.

The rapid advancement of generative models, facilitating the creation of hyper-realistic images from textual descriptions, has concurrently escalated critical societal concerns such as misinformation. Although providing some mitigation, traditional fingerprinting mechanisms fall short in attributing responsibility for the malicious use of synthetic images. This paper introduces a novel approach to model fingerprinting that assigns responsibility for the generated images, thereby serving as a potential countermeasure to model misuse. Our method modifies generative models based on each user's unique digital fingerprint, imprinting a unique identifier onto the resultant content that can be traced back to the user. This approach, incorporating fine-tuning into Text-to-Image (T2I) tasks using the Stable Diffusion Model, demonstrates near-perfect attribution accuracy with a minimal impact on output quality. Through extensive evaluation, we show that our method outperforms baseline methods with an average improvement of 11\% in handling image post-processes. Our method presents a promising and novel avenue for accountable model distribution and responsible use. Our code is available in the Appendix.

Deepfake detection faces a critical generalization hurdle, with performance deteriorating when there is a mismatch between the distributions of training and testing data. A broadly received explanation is the tendency of these detectors to be overfitted to forgery-specific artifacts, rather than learning features that are widely applicable across various forgeries. To address this issue, we propose a simple yet effective detector called LSDA (\underline{L}atent \underline{S}pace \underline{D}ata \underline{A}ugmentation), which is based on a heuristic idea: representations with a wider variety of forgeries should be able to learn a more generalizable decision boundary, thereby mitigating the overfitting of method-specific features (see Fig. 1). Following this idea, we propose to enlarge the forgery space by constructing and simulating variations within and across forgery features in the latent space. This approach encompasses the acquisition of enriched, domain-specific features and the facilitation of smoother transitions between different forgery types, effectively bridging domain gaps. Our approach culminates in refining a binary classifier that leverages the distilled knowledge from the enhanced features, striving for a generalizable deepfake detector. Comprehensive experiments show that our proposed method is surprisingly effective and transcends state-of-the-art detectors across several widely used benchmarks.

Image diffusion models have been utilized in various tasks, such as text-to-image generation and controllable image synthesis. Recent research has introduced tuning methods that make subtle adjustments to the original models, yielding promising results in specific adaptations of foundational generative diffusion models. Rather than modifying the main backbone of the diffusion model, we delve into the role of skip connection in U-Net and reveal that hierarchical features aggregating long-distance information across encoder and decoder make a significant impact on the content and quality of image generation. Based on the observation, we propose an efficient generative tuning framework, dubbed SCEdit, which integrates and edits Skip Connection using a lightweight tuning module named SC-Tuner. Furthermore, the proposed framework allows for straightforward extension to controllable image synthesis by injecting different conditions with Controllable SC-Tuner, simplifying and unifying the network design for multi-condition inputs. Our SCEdit substantially reduces training parameters, memory usage, and computational expense due to its lightweight tuners, with backward propagation only passing to the decoder blocks. Extensive experiments conducted on text-to-image generation and controllable image synthesis tasks demonstrate the superiority of our method in terms of efficiency and performance. Project page: https://scedit.github.io/.

Images produced by text-to-image diffusion models might not always faithfully represent the semantic intent of the provided text prompt where the model might overlook or entirely fail to produce certain objects. While recent studies propose various solutions, they often require customly tailored functions for each of these problems, leading to sub-optimal results, especially for complex prompts. Our work introduces a novel perspective by tackling this challenge in a contrastive context. Our approach intuitively promotes the segregation of objects in attention maps, while also maintaining that pairs of related attributes are kept close to each other. We conducted extensive experiments across a wide variety of scenarios, each involving unique combinations of objects, attributes, and scenes. These experiments effectively showcase the versatility, efficiency, and flexibility of our method in working with both latent and pixel-based diffusion models, including Stable Diffusion and Imagen. Moreover, we publicly share our source code to facilitate further research.

Text-conditioned image-to-video generation (TI2V) aims to synthesize a realistic video starting from a given image (e.g., a woman's photo) and a text description (e.g., "a woman is drinking water."). Existing TI2V frameworks often require costly training on video-text datasets and specific model designs for text and image conditioning. In this paper, we propose TI2V-Zero, a zero-shot, tuning-free method that empowers a pretrained text-to-video (T2V) diffusion model to be conditioned on a provided image, enabling TI2V generation without any optimization, fine-tuning, or introducing external modules. Our approach leverages a pretrained T2V diffusion foundation model as the generative prior. To guide video generation with the additional image input, we propose a "repeat-and-slide" strategy that modulates the reverse denoising process, allowing the frozen diffusion model to synthesize a video frame-by-frame starting from the provided image. To ensure temporal continuity, we employ a DDPM inversion strategy to initialize Gaussian noise for each newly synthesized frame and a resampling technique to help preserve visual details. We conduct comprehensive experiments on both domain-specific and open-domain datasets, where TI2V-Zero consistently outperforms a recent open-domain TI2V model. Furthermore, we show that TI2V-Zero can seamlessly extend to other tasks such as video infilling and prediction when provided with more images. Its autoregressive design also supports long video generation.

Incorporating human feedback has been shown to be crucial to align text generated by large language models to human preferences.We hypothesize that state-of-the-art instructional image editing models, where outputs are generated based on an input image and an editing instruction, could similarly benefit from human feedback, as their outputs may not adhere to the correct instructions and preferences of users. In this paper, we present a novel framework to harness human feedback for instructional visual editing (HIVE). Specifically, we collect human feedback on the edited images and learn a reward function to capture the underlying user preferences. We then introduce scalable diffusion model fine-tuning methods that can incorporate human preferences based on the estimated reward. Besides, to mitigate the bias brought by the limitation of data, we contribute a new 1.1M training dataset, a 3.6K reward dataset for rewards learning, and a 1K evaluation dataset to boost the performance of instructional image editing. We conduct extensive empirical experiments quantitatively and qualitatively, showing that HIVE is favored over previous state-of-the-art instructional image editing approaches by a large margin.

Despite the remarkable performance of score distillation in text-to-3D generation, such techniques notoriously suffer from view inconsistency issues, also known as "Janus" artifact, where the generated objects fake each view with multiple front faces. Although empirically effective methods have approached this problem via score debiasing or prompt engineering, a more rigorous perspective to explain and tackle this problem remains elusive. In this paper, we reveal that the existing score distillation-based text-to-3D generation frameworks degenerate to maximal likelihood seeking on each view independently and thus suffer from the mode collapse problem, manifesting as the Janus artifact in practice. To tame mode collapse, we improve score distillation by re-establishing in entropy term in the corresponding variational objective, which is applied to the distribution of rendered images. Maximizing the entropy encourages diversity among different views in generated 3D assets, thereby mitigating the Janus problem. Based on this new objective, we derive a new update rule for 3D score distillation, dubbed Entropic Score Distillation (ESD). We theoretically reveal that ESD can be simplified and implemented by just adopting the classifier-free guidance trick upon variational score distillation. Although embarrassingly straightforward, our extensive experiments successfully demonstrate that ESD can be an effective treatment for Janus artifacts in score distillation.

Large generative diffusion models have revolutionized text-to-image generation and offer immense potential for conditional generation tasks such as image enhancement, restoration, editing, and compositing. However, their widespread adoption is hindered by the high computational cost, which limits their real-time application. To address this challenge, we introduce a novel method dubbed CoDi, that adapts a pre-trained latent diffusion model to accept additional image conditioning inputs while significantly reducing the sampling steps required to achieve high-quality results. Our method can leverage architectures such as ControlNet to incorporate conditioning inputs without compromising the model's prior knowledge gained during large scale pre-training. Additionally, a conditional consistency loss enforces consistent predictions across diffusion steps, effectively compelling the model to generate high-quality images with conditions in a few steps. Our conditional-task learning and distillation approach outperforms previous distillation methods, achieving a new state-of-the-art in producing high-quality images with very few steps (e.g., 1-4) across multiple tasks, including super-resolution, text-guided image editing, and depth-to-image generation.

Most of the recent literature on image Super-Resolution (SR) can be classified into two main approaches. The first one involves learning a corruption model tailored to a specific dataset, aiming to mimic the noise and corruption in low-resolution images, such as sensor noise. However, this approach is data-specific, tends to lack adaptability, and its accuracy diminishes when faced with unseen types of image corruptions. A second and more recent approach, referred to as Robust Super-Resolution (RSR), proposes to improve real-world SR by harnessing the generalization capabilities of a model by making it robust to adversarial attacks. To delve further into this second approach, our paper explores the universality of various methods for enhancing the robustness of deep learning SR models. In other words, we inquire: \enquote{Which robustness method exhibits the highest degree of adaptability when dealing with a wide range of adversarial attacks ?}. Our extensive experimentation on both synthetic and real-world images empirically demonstrates that median randomized smoothing (MRS) is more general in terms of robustness compared to adversarial learning techniques, which tend to focus on specific types of attacks. Furthermore, as expected, we also illustrate that the proposed universal robust method enables the SR model to handle standard corruptions more effectively, such as blur and Gaussian noise, and notably, corruptions naturally present in real-world images. These results support the significance of shifting the paradigm in the development of real-world SR methods towards RSR, especially via MRS.

Text-to-image (T2I) diffusion models, notably the unCLIP models (e.g., DALL-E-2), achieve state-of-the-art (SOTA) performance on various compositional T2I benchmarks, at the cost of significant computational resources. The unCLIP stack comprises T2I prior and diffusion image decoder. The T2I prior model alone adds a billion parameters compared to the Latent Diffusion Models, which increases the computational and high-quality data requirements. We introduce ECLIPSE, a novel contrastive learning method that is both parameter and data-efficient. ECLIPSE leverages pre-trained vision-language models (e.g., CLIP) to distill the knowledge into the prior model. We demonstrate that the ECLIPSE trained prior, with only 3.3% of the parameters and trained on a mere 2.8% of the data, surpasses the baseline T2I priors with an average of 71.6% preference score under resource-limited setting. It also attains performance on par with SOTA larger models, achieving an average of 63.36% preference score in terms of the ability to follow the text compositions. Extensive experiments on two unCLIP diffusion image decoders, Karlo and Kandinsky, affirm that ECLIPSE consistently delivers high performance while significantly reducing resource dependency.

Text-driven video editing poses significant challenges in exhibiting flicker-free visual continuity while preserving the inherent motion patterns of original videos. Existing methods operate under a paradigm where motion and appearance are intricately intertwined. This coupling leads to the network either overfitting appearance content – failing to capture motion patterns – or focusing on motion patterns at the expense of content generalization to diverse textual scenarios. Inspired by the pivotal role of wavelet transform in dissecting video sequences, we propose CAusal Motion Enhancement tailored for Lifting text-driven video editing (CAMEL), a novel technique with two core designs. First, we introduce motion prompts, designed to summarize motion concepts from video templates through direct optimization. The optimized prompts are purposefully integrated into latent representations of diffusion models to enhance the motion fidelity of generated results. Second, to enhance motion coherence and extend the generalization of appearance content to creative textual prompts, we propose the causal motion-enhanced attention mechanism. This mechanism is implemented in tandem with a novel causal motion filter, synergistically enhancing the motion coherence of disentangled high-frequency components, and concurrently preserving the generalization of appearance content across various textual scenarios. Extensive experimental results show the superior performance of CAMEL.

Benefiting from large-scale pre-trained text-to-image (T2I) generative models, impressive progress has been achieved in customized image generation, which aims to generate user-specified concepts.Existing approaches have extensively focused on single-concept customization and still encounter challenges when it comes to complex scenarios that involve combining multiple concepts. These approaches often require retraining/fine-tuning using a few images, leading to time-consuming training processes and impeding their swift implementation. Furthermore, the reliance on multiple images to represent a singular concept increases the difficulty of customization.To this end, we propose FreeCustom, a novel tuning-free method to generate customized images of multi-concept composition based on reference concepts, using only one image per concept as input. Specifically, we introduce a new multi-reference self-attention (MRSA) mechanism and a weighted mask strategy that enables the generated image to access and focus more on the reference concepts. In addition, MRSA leverages our key finding that input concepts are better preserved when providing images with context interactions.Experiments show that our method's produced images are consistent with the given concepts and better aligned with the input text. Our method outperforms or performs on par with other training-based methods in terms of multi-concept composition and single-concept customization, but is simpler. Codes can be found here.

Our brain can effortlessly recognize objects even when partially hidden from view. Seeing the visible of the hidden is called amodal completion; however, this task remains a challenge for generative AI despite rapid progress. We propose to sidestep many of the difficulties of existing approaches, which typically involve a two-step process of predicting amodal masks and then generating pixels. Our method involves thinking outside the box, literally! We go outside the object bounding box to use its context to guide a pre-trained diffusion inpainting model, and then progressively grow the occluded object and trim the extra background. We overcome two technical challenges: 1) how to be free of unwanted co-occurrence bias, which tends to regenerate similar occluders, and 2) how to judge if an amodal completion has succeeded. Our amodal completion method exhibits improved photorealistic completion results compared to existing approaches in numerous successful completion cases. And the best part? It doesn't require any special training or fine-tuning of models.

We introduced StyleEntity, a zero-shot image manipulation model that utilizes named entities as proxies during its training phase. This strategy enables our model to manipulate images using unseen textual descriptions during inference, all within a single training phase. Additionally, we proposed an inference technique termed Prompt Ensemble Latent Averaging (PELA). PELA averages the manipulation directions derived from various named entities during inference, effectively eliminating the noise directions, thus achieving stable manipulation. In our experiments, StyleEntity exhibited superior performance in a zero-shot setting compared to other methods. The code, model weights, and datasets are available at https://github.com/feng-zhida/StyleEntity.

Blind face restoration focuses on restoring high-fidelity details from images subjected to complex and unknown degradations, while preserving identity information. In this paper, we present a Prior-based Latent Transformation approach (PLTrans), which is specifically designed to learn a degradation-unaware representation, thereby allowing the restoration network to effectively generalize to real-world degradation. Toward this end, PLTrans learns a degradation-unaware query via a latent diffusion-based regularization module. Furthermore, conditioned on the features of a degraded face image, a latent dictionary that captures the priors of HQ face images is leveraged to refine the features by mapping the top-$d$ nearest elements. The refined version will be used to build key and value for the cross-attention computation, which is tailored to each degraded image and exhibits reduced sensitivity to different degradation factors. Conditioned on the resulting representation, we train a decoding network that synthesizes face images with authentic details and identity preservation. Through extensive experiments, we verify the effectiveness of the design elements and demonstrate the generalization ability of our proposed approach for both synthetic and unknown degradations. We finally demonstrate the applicability of PLTrans in other vision tasks.

With recent text-to-image models, anyone can generate deceptively realistic images with arbitrary contents, fueling the growing threat of visual disinformation. A key enabler for generating high-resolution images with low computational cost has been the development of latent diffusion models (LDMs). In contrast to conventional diffusion models, LDMs perform the denoising process in the low-dimensional latent space of a pre-trained autoencoder (AE) instead of the high-dimensional image space. Despite their relevance, the forensic analysis of LDMs is still in its infancy. In this work we propose AEROBLADE, a novel detection method which exploits an inherent component of LDMs: the AE used to transform images between image and latent space. We find that generated images can be more accurately reconstructed by the AE than real images, allowing for a simple detection approach based on the reconstruction error. Most importantly, our method is easy to implement and does not require any training, yet nearly matches the performance of detectors that rely on extensive training. We empirically demonstrate that AEROBLADE is effective against state-of-the-art LDMs, including Stable Diffusion and Midjourney. Beyond detection, our approach allows for the qualitative analysis of images, which can be leveraged for identifying inpainted regions. We release our code and data at https://github.com/jonasricker/aeroblade.

Face Video Retouching is a complex task that often requires labor-intensive manual editing. Conventional image retouching methods perform less satisfactorily in terms of generalization performance and stability when applied to videos without exploiting the correlation among frames. To address this issue, we propose a Video Retouching transformEr to remove facial imperfections in videos, which is referred to as VRetouchEr. Specifically, we estimate the apparent motion of imperfections between two consecutive frames, and the resulting displacement vectors are used to refine the imperfection map, which is synthesized from the current frame together with the corresponding encoder features. The flow-based imperfection refinement is critical for precise and stable retouching across frames. To leverage the temporal contextual information, we inject the refined imperfection map into each transformer block for multi-frame masked attention computation, such that we can capture the interdependence between the current frame and multiple reference frames. As a result, the imperfection regions can be replaced with normal skin with high fidelity, while at the same time keeping the other regions unchanged. Extensive experiments are performed to verify the superiority of VRetouchEr over state-of-the-art image retouching methods in terms of fidelity and stability.

Recent advances in generative models trained on large-scale datasets have made it possible to synthesize high-quality samples across various domains. Moreover, the emergence of strong inversion networks enables not only a reconstruction of real-world images but also the modification of attributes through various editing methods. However, in certain domains related to privacy issues, e.g., human faces, advanced generative models along with strong inversion methods can lead to potential misuses. In this paper, we propose an essential yet under-explored task called generative identity unlearning, which steers the model not to generate an image of a specific identity. In the generative identity unlearning, we target the following objectives: (i) preventing the generation of images with a certain identity, and (ii) preserving the overall quality of the generative model. To satisfy these goals, we propose a novel framework, $\textbf{G}$enerative $\textbf{U}$nlearning for Any $\textbf{IDE}$ntity ($\textbf{GUIDE}$), which prevents the reconstruction of a specific identity by unlearning the generator with only a single image. GUIDE consists of two parts: (i) finding a target point for optimization that un-identifies the source latent code and (ii) novel loss functions that facilitate the unlearning procedure while less affecting the learned distribution. Our extensive experiments demonstrate that our proposed method achieves state-of-the-art performance in the generative machine unlearning task. The code will be released after the review.

We study text-based image editing (TBIE) of a single image by counterfactual inference because it is an elegant formulation to precisely address the requirement: the edited image should retain the fidelity of the original one. Through the lens of the formulation, we find that the crux of TBIE is that existing techniques hardly achieve a good trade-off between editability and fidelity, mainly due to the overfitting of the single-image fine-tuning. To this end, we propose a Doubly Abductive Counterfactual inference framework (DAC). We first parameterize an exogenous variable as a UNet LoRA, whose abduction can encode all the image details. Second, we abduct another exogenous variable parameterized by a text encoder LoRA, which recovers the lost editability caused by the overfitted first abduction. Thanks to the second abduction, which exclusively encodes the visual transition from post-edit to pre-edit, its inversion---subtracting the LoRA---effectively reverts pre-edit back to post-edit, thereby accomplishing the edit. Through extensive experiments, our DAC achieves a good trade-off between editability and fidelity. Thus, we can support a wide spectrum of user editing intents, including addition, removal, manipulation, replacement, style transfer, and facial change, which are extensively validated in both qualitative and quantitative evaluations.

With the advent of generative models and vision-language pretraining, significant improvement has been made in text-driven face manipulation. The text embedding can be used as target supervision for expression control. However, it is non-trivial to associate with its 3D attributes, \ie, pose and illumination. To address these issues, we propose a Text-conditional Attribute aLignment approach for 3D controllable face image synthesis, and our model is referred to as TcALign. Specifically, since the 3D rendered image can be precisely controlled with the 3D face representation, we first propose a Text-conditional 3D Editor to produce the target face representation to realize text-driven manipulation in the 3D space. An attribute embedding space spanned by the target-related attributes embeddings is also introduced to infer the disentangled task-specific direction.Next, we train a cross-modal latent mapping network conditioned on the derived difference of 3D representation to infer a correct vector in the latent space of StyleGAN. This correction vector learning design can accurately transfer the attribute manipulation on 3D images to 2D images. We show that the proposed method delivers more precise text-driven multi-attribute manipulation for 3D controllable face image synthesis. Extensive qualitative and quantitative experiments verify the effectiveness and superiority of our method over the other competing methods.

Customizing pre-trained text-to-image generation model has attracted massive research interest recently, due to its huge potential in real-world applications. Although existing methods are able to generate creative content for a novel concept contained in single user-input image, their capability are still far from perfection. Specifically, most existing methods require fine-tuning the generative model on testing images. Some existing methods does not require fine-tuning, while their performance are unsatisfactory. Furthermore, the interaction between users and models are still limited to directive and descriptive prompts such as instructions and captions. In this work, we built a customization assistant based on pre-trained large language model and diffusion model, which can not only perform customized generation in a tuning-free manner within few seconds, but also enable more user-friendly interactions: users can chat with the assistant and input either ambiguous text or clear instruction. Specifically, we propose a new framework consists of a new model design and a novel training strategy with self-distillation. The resulting assistant can perform customized generation in2-5 seconds without any test time fine-tuning. Extensive experiments are conducted, competitive results have been obtained across different domains, illustrating the effectiveness of the proposed method.

With the remarkable advent of text-to-image diffusion models, image editing methods have become more diverse and continue to evolve. A promising recent approach in this realm is Delta Denoising Score (DDS) - an image editing technique based on Score Distillation Sampling (SDS) framework that leverages the rich generative prior of text-to-image diffusion models. However, relying solely on the difference between scoring functions is insufficient for preserving specific structural elements from the original image, a crucial aspect of image editing. To address this, here we present an embarrassingly simple yet very powerful modification of DDS, called Contrastive Denoising Score (CDS), for latent diffusion models (LDM). Inspired by the similarities and differences between DDS and the contrastive learning for unpaired image-to-image translation(CUT), we introduce a straightforward approach using CUT loss within the DDS framework. Rather than employing auxiliary networks as in the original CUT approach, we leverage the intermediate features of LDM, specifically those from the self-attention layers, which possesses rich spatial information. Our approach enables zero-shot image-to-image translation and neural radiance field (NeRF) editing, achieving structural correspondence between the input and output while maintaining content controllability. Qualitative results and comparisons demonstrates the effectiveness of our proposed method.

Super-resolution (SR) and image generation are important tasks in computer vision and are widely adopted in real-world applications.Most existing methods, however, generate images only at fixed-scale magnification and suffer from over-smoothing and artifacts.Additionally, they do not offer enough diversity of output images nor image consistency at different scales.Most relevant work applied Implicit Neural Representation (INR) to the denoising diffusion model to obtain continuous-resolution yet diverse and high-quality SR results.Since this model operates in the image space, the larger the resolution of image is produced, the more memory and inference time is required, and it also does not maintain scale-specific consistency.We propose a novel pipeline that can super-resolve an input image or generate from a random noise a novel image at arbitrary scales. The method consists of a pretrained auto-encoder, a latent diffusion model, and an implicit neural decoder, and their learning strategies.The proposed method adopts diffusion processes in a latent space, thus efficient, yet aligned with output image space decoded by MLPs at arbitrary scales. More specifically, our arbitrary-scale decoder is designed by the symmetric decoder w/o up-scaling from the pretrained auto-encoder, and Local Implicit Image Function (LIIF) in series. The latent diffusion process is learnt by the denoising and the alignment losses jointly. Errors in output images are backpropagated via the fixed decoder, improving the quality of output images.In the extensive experiments using multiple public benchmarks on the two tasks i.e. image super-resolution and novel image generation at arbitrary scales, the proposed method outperforms relevant methods in metrics of image quality, diversity and scale consistency. It is significantly better than the relevant prior-art in the inference speed and memory usage.

Text-to-video diffusion models have advanced video generation significantly. However,customizing these models to generate videos with tailored motions presents a substantial challenge. In specific,they encounter hurdles in (1) accurately reproducing motion from a target video, and (2) creating diverse visual variations. For example, straightforward extensions of static image customization methods to video often lead to intricate entanglements of appearance and motion data. To tackle this, here we present the Video Motion Customization (VMC) framework, a novel one-shot tuning approach crafted to adapt temporal attention layers within video diffusion models. Our approach introduces a novel motion distillation objective using residual vectors between consecutive noisy latent frames as a motion reference. The diffusion process then preserve low-frequency motion trajectories while mitigating high-frequency motion-unrelated noise in image space. We validate our method against state-of-the-art video generative models across diverse real-world motions and contexts. Our codes and data can be found at: https://video-motion-customization.github.io/

This paper proposes an image-vector dual diffusion model for generative layout design. Distinct from prior efforts that mostly ignores element-level visual information, our approach integrates the power of a pre-trained large image diffusion model to guide layout composition in a vector diffusion model by providing enhanced salient region understanding and high-level inter-element relationship reasoning. Our proposed model simultaneously operates in two domains: it generates the overall design appearance in the image domain while optimizing the size and position of each design element in the vector domain. The proposed method achieves the state-of-the-art results on several datasets and enables new layout design applications.

Diffusion models are just at a tipping point for image super-resolution task. Nevertheless, it is not trivial to capitalize on diffusion models for video super-resolution which necessitates not only the preservation of visual appearance from low-resolution to high-resolution videos, but also the temporal consistency across video frames. In this paper, we propose a novel approach, pursuing Spatial Adaptation and Temporal Coherence (SATeCo), for video super-resolution. SATeCo pivots on learning spatial-temporal guidance from low-resolution videos to calibrate both latent-space high-resolution video denoising and pixel-space video reconstruction. Technically, SATeCo freezes all the parameters of the pre-trained UNet and VAE, and only optimizes two deliberately-designed spatial feature adaptation (SFA) and temporal feature alignment (TFA) modules, in the decoder of UNet and VAE. SFA modulates frame features via adaptively estimating affine parameters for each pixel, guaranteeing pixel-wise guidance for high-resolution frame synthesis. TFA delves into feature interaction within a 3D local window (tubelet) through self-attention, and executes cross-attention between tubelet and its low-resolution counterpart to guide temporal feature alignment. Extensive experiments conducted on the REDS4 and Vid4 datasets demonstrate the effectiveness of our approach.

Diffusion models represent a new paradigm in text-to-image generation. Beyond generating high-quality images from text prompts, models such as Stable Diffusion have been successfully extended to the joint generation of semantic segmentation pseudo-masks. However, current extensions primarily rely on extracting attentions linked to prompt words used for image synthesis. This approach limits the generation of segmentation masks derived from word tokens not contained in the text prompt. In this work, we introduce Open-Vocabulary Attention Maps (OVAM)—a training-free method for text-to-image diffusion models that enables the generation of attention maps for any word. In addition, we propose a lightweight optimization process based on OVAM for finding tokens that generate accurate attention maps for an object class with a single annotation. We evaluate these tokens within existing state-of-the-art Stable Diffusion extensions. The best-performing model improves its mIoU from 52.1 to 86.6 for the synthetic images' pseudo-masks, demonstrating that our optimized tokens are an efficient way to improve the performance of existing methods without architectural changes or retraining.

Implicit neural representations (INRs) were recently proposed as a new video compression paradigm, with existing approaches performing on par with H.264 and HEVC. However, such methods only perform well in limited settings, e.g., specific model sizes, fixed aspect ratios, and low-motion videos. We address this issue by proposing T-NeRV, a hybrid video INR that combines frame-specific embeddings with GOP-specific features, providing a lever for content-specific fine-tuning. We employ entropy-constrained training to jointly optimize our model for rate and distortion and demonstrate that T-NeRV can thereby automatically adjust this lever during training, effectively fine-tuning itself to the target content. We evaluate T-NeRV on the UVG dataset, where it achieves state-of-the-art results on the video regression task, outperforming previous works by up to 3dB PSNR on challenging high-motion sequences. Further, our method improves on the compression performance of previous methods and is the first video INR to outperform HEVC on all UVG sequences.

Recent advancements in large-scale pre-trained text-to-image models have led to remarkable progress in semantic image synthesis. Nevertheless, synthesizing high-quality images with consistent semantics and layout remains a challenge. In this paper, we propose the adaPtive LAyout-semantiC fusion modulE (PLACE) that harnesses pre-trained models to alleviate the aforementioned issues. Specifically, we first employ the layout control map to faithfully represent layouts in the feature space. Subsequently, we combine the layout and semantic features in a timestep-adaptive manner to synthesize images with realistic details. During fine-tuning, we propose the Semantic Alignment (SA) loss to further enhance layout alignment. Additionally, we introduce the Layout-Free Prior Preservation (LFP) loss, which leverages unlabeled data to maintain the priors of pre-trained models, thereby improving the visual quality and semantic consistency of synthesized images. Extensive experiments demonstrate that our approach performs favorably in terms of visual quality, semantic consistency, and layout alignment. The source code and model are available at \href{https://github.com/cszy98/PLACE/tree/main}{PLACE}.

Knowledge distillation methods have recently shown to be a promising direction to speedup the synthesis of large-scale diffusion models by requiring only a few inference steps. While several powerful distillation methods were recently proposed, the overall quality of student samples is typically lower compared to the teacher ones, which hinders their practical usage. In this work, we investigate the relative quality of samples produced by the teacher text-to-image diffusion model and its distilled student version. As our main empirical finding, we discover that a noticeable portion of student samples exhibit superior fidelity compared to the teacher ones, despite the ``approximate'' nature of the student. Based on this finding, we propose an adaptive collaboration between student and teacher diffusion models for effective text-to-image synthesis. Specifically, the distilled model produces the initial sample, and then an oracle decides whether it needs further improvements with a slow teacher model. Extensive experiments demonstrate that the designed pipeline surpasses state-of-the-art text-to-image alternatives for various inference budgets in terms of human preference. Furthermore, the proposed approach can be naturally used in popular applications such as text-guided image editing and controllable generation.

Neural Radiance Fields (NeRFs) have shown great potential in novel view synthesis. However, they struggle to render sharp images when the data used for training is affected by motion blur. On the other hand, event cameras excel in dynamic scenes as they measure brightness changes with microsecond resolution and are thus only marginally affected by blur. Recent methods attempt to enhance NeRF reconstructions under camera motion by fusing frames and events. However, they face challenges in recovering accurate color content or constrain the NeRF to a set of predefined camera poses, harming reconstruction quality in challenging conditions. This paper proposes a novel formulation addressing these issues by leveraging both model- and learning-based modules. We explicitly model the blur formation process, exploiting the event double integral as an additional model-based prior. Additionally, we model the event-pixel response using an end-to-end learnable response function, allowing our method to adapt to non-idealities in the real event-camera sensor. We show, on synthetic and real data, that the proposed approach outperforms existing deblur NeRFs that use only frames as well as those that combine frames and events by +6.13dB and +2.48dB, respectively.

In the realm of AI, data serves as a pivotal resource. Real-world hyperspectral images (HSIs), bearing wide spectral characteristics, are particularly valuable. However, the acquisition of HSIs is always costly and time-intensive, resulting in a severe data-thirsty issue in HSI research and applications. Current solutions have not been able to generate a sufficient volume of diverse and reliable synthetic HSIs. To this end, our study formulates a novel, generalized paradigm for HSI synthesis, i.e., unmixing before fusion, that initiates with unmixing across multi-source data and follows by fusion-based synthesis. By integrating unmixing, this work maps unpaired HSI and RGB data to a low-dimensional abundance space, greatly alleviating the difficulty of generating high-dimensional samples. Moreover, incorporating abundances inferred from unpaired RGB images into generative models allows for cost-effective supplementation of various realistic spatial distributions in abundance synthesis. Our proposed paradigm can be instrumental with a series of deep generative models, filling a significant gap in the field and enabling the generation of vast high-quality HSI samples for large-scale downstream tasks. Extension experiments on downstream tasks demonstrate the effectiveness of synthesized HSIs. The code is available at: HSI-Synthesis.github.io.

As with many machine learning problems, the progress of image generation methods hinges on good evaluation metrics. One of the most popular is the Frechet Inception Distance (FID). FID estimates the distance between a distribution of Inception-v3 features of real images, and those of images generated by the algorithm. We highlight important drawbacks of FID: Inception's poor representation of the rich and varied content generated by modern text-to-image models, incorrect normality assumptions, and poor sample complexity. We call for a reevaluation of FID's use as the primary quality metric for generated images. We empirically demonstrate that FID contradicts human raters, it does not reflect gradual improvement of iterative text-to-image models, it does not capture distortion levels, and that it produces inconsistent results when varying the sample size. We also propose an alternative new metric, CMMD, based on richer CLIP embeddings and the maximum mean discrepancy distance with the Gaussian RBF kernel. It is an unbiased estimator that does not make any assumptions on the probability distribution of the embeddings and is sample efficient. Through extensive experiments and analysis, we demonstrate that FID-based evaluations of text-to-image models may be unreliable, and that CMMD offers a more robust and reliable assessment of image quality.

Modern text-to-image generation models produce high-quality images that are both photorealistic and faithful to the text prompts. However, this quality comes at significant computational cost: nearly all of these models are iterative and require running inference multiple times with large models. This iterative process is needed to ensure that different regions of the image are not only aligned with the text prompt, but also compatible with each other. In this work, we propose a light-weight approach to achieving this compatibility between different regions of an image, using a Markov Random Field (MRF) model. We demonstrate the effectiveness of this method on top of the latent token-based Muse text-to-image model. The MRF richly encodes the compatibility among image tokens at different spatial locations to improve quality and significantly reduce the required number of Muse sampling steps. Inference with the MRF is significantly cheaper, and its parameters can be quickly learned through back-propagation by modeling MRF inference as a differentiable neural-network layer. Our full model, MarkovGen, uses this proposed MRF model to both speed up Muse by $1.5\times$ and produce higher quality images by decreasing undesirable image artifacts.

Generative AI has made significant strides in computer vision, particularly in text-driven image/video synthesis (T2I/T2V). Despite the notable advancements, it remains challenging in human-centric content synthesis such as realistic dance generation. Current methodologies, primarily tailored for human motion transfer, encounter difficulties when confronted with real-world dance scenarios (e.g., social media dance), which require to generalize across a wide spectrum of poses and intricate human details. In this paper, we depart from the traditional paradigm of human motion transfer and emphasize two additional critical attributes for the synthesis of human dance content in social media contexts: (i) Generalizability: the model should be able to generalize beyond generic human viewpoints as well as unseen human subjects, backgrounds, and poses; (ii) Compositionality: it should allow for the seamless composition of seen/unseen subjects, backgrounds, and poses from different sources. To address these challenges, we introduce DisCo, which includes a novel model architecture with disentangled control to improve the compositionality of dance synthesis, and an effective human attribute pre-training for better generalizability to unseen humans. Extensive qualitative and quantitative results demonstrate that DisCo can generate high-quality human dance images and videos with diverse appearances and flexible motions.

The task of manipulating real image attributes through StyleGAN inversion has been extensively researched. This process involves searching latent variables from a well-trained StyleGAN generator that can synthesize a real image, modifying these latent variables, and then synthesizing an image with the desired edits. A balance must be struck between the quality of the reconstruction and the ability to edit. Earlier studies utilized the low-dimensional W-space for latent search, which facilitated effective editing but struggled with reconstructing intricate details. More recent research has turned to the high-dimensional feature space F, which successfully inverses the input image but loses much of the detail during editing. In this paper, we introduce StyleFeatureEditor -- a novel method that enables editing in both w-latents and F-latents. This technique not only allows for the reconstruction of finer image details but also ensures their preservation during editing. We also present a new training pipeline specifically designed to train our model to accurately edit F-latents. Our method is compared with state-of-the-art encoding approaches, demonstrating that our model excels in terms of reconstruction quality and is capable of editing even challenging out-of-domain examples.

Most neural compression models are trained on large datasets of images or videos in order to generalize to unseen data. Such generalization typically requires large and expressive architectures with a high decoding complexity. Here we introduce C3, a neural compression method with strong rate-distortion (RD) performance that instead overfits a small model to each image or video separately. The resulting decoding complexity of C3 can be an order of magnitude lower than neural baselines with similar RD performance. C3 builds on COOL-CHIC (Ladune et al.) and makes several simple and effective improvements for images. We further develop new methodology to apply C3 to videos. On the CLIC2020 image benchmark, we match the RD performance of VTM, the reference implementation of the H.266 codec, with less than 3k MACs/pixel for decoding. On the UVG video benchmark, we match the RD performance of the Video Compression Transformer (Mentzer et al.), a well-established neural video codec, with less than 5k MACs/pixel for decoding.

We introduce LightIt, a method for explicit illumination control for image generation. Recent generative methods lack lighting control, which is crucial to numerous artistic aspects of image generation such as setting the overall mood or cinematic appearance. To overcome these limitations, we propose to condition the generation on shading and normal maps. We model the lighting with single bounce shading, which includes cast shadows. We first train a shading estimation module to generate a dataset of real-world images and shading pairs. Then, we train a control network using the estimated shading and normals as input. Our method demonstrates high-quality image generation and lighting control in numerous scenes. Additionally, we use our generated dataset to train an identity-preserving relighting model, conditioned on an image and a target shading. Our method is the first that enables the generation of images with controllable, consistent lighting and performs on par with specialized relighting state-of-the-art methods.

Classifier-Free Guidance (CFG) has been widely used in text-to-image diffusion models, where the CFG scale is introduced to control the strength of text guidance on the whole image space. However, we argue that a global CFG scale results in spatial inconsistency on varying semantic strengths and suboptimal image quality. To address this problem, we present a novel approach, Semantic-aware Classifier-Free Guidance (S-CFG), to customize the guidance degrees for different semantic units in text-to-image diffusion models. Specifically, we first design a training-free semantic segmentation method to partition the latent image into relatively independent semantic regions at each denoising step. In particular, the cross-attention map in the denoising U-net backbone is renormalized for assigning each patch to the corresponding token, while the self-attention map is used to complete the semantic regions. Then, to balance the amplification of diverse semantic units, we adaptively adjust the CFG scales across different semantic regions to rescale the text guidance degrees into a uniform level. Finally, extensive experiments demonstrate the superiority of S-CFG over the original CFG strategy on various text-to-image diffusion models, without requiring any extra training cost.

Recent strides in the development of diffusion models, exemplified by advancements such as Stable Diffusion, have underscored their remarkable prowess in generating visually compelling images. However, the imperative of achieving a seamless alignment between the generated image and the provided prompt persists as a formidable challenge. This paper traces the root of these difficulties to invalid initial noise, and proposes a solution in the form of Initial Noise Optimization (InitNO), a paradigm that refines this noise. Considering text prompts, not all random noises are effective in synthesizing semantically-faithful images. We design the cross-attention response score and the self-attention conflict score to evaluate the initial noise, bifurcating the initial latent space into valid and invalid sectors. A strategically crafted noise optimization pipeline is developed to guide the initial noise towards valid regions. Our method, validated through rigorous experimentation, shows a commendable proficiency in generating images in strict accordance with text prompts. Our code is available at https://github.com/xiefan-guo/initno.

Contemporary machine learning, which involves training large neural networks on massive datasets, faces significant computational challenges. Dataset distillation, as a recent emerging strategy, aims to compress real-world datasets for efficient training. However, this line of research currently struggles with large-scale and high-resolution datasets, hindering its practicality and feasibility. Thus, we re-examine existing methods and identify three properties essential for real-world applications: realism, diversity, and efficiency. As a remedy, we propose RDED, a novel computationally-efficient yet effective data distillation paradigm, to enable both diversity and realism of the distilled data. Extensive empirical results over various model architectures and datasets demonstrate the advancement of RDED: we can distill a dataset to 10 images per class from full ImageNet-1K within 7 minutes, achieving a notable 42% accuracy with ResNet-18 on a single RTX-4090 GPU (while the SOTA only achieves 21% but requires 6 hours). Code: https://github.com/LINs-lab/RDED.

Scaling up model and data size has been quite successful for the evolution of LLMs. However, the scaling law for the diffusion based text-to-image (T2I) models is not fully explored. It is also unclear how to efficiently scale the model for better performance at reduced cost. The different training settings and expensive training cost make a fair model comparison extremely difficult. In this work, we empirically study the scaling properties of diffusion based T2I models by performing extensive and rigours ablations on scaling both denoising backbones and training set, including training scaled UNet and Transformer variants ranging from 0.4B to 4B parameters on datasets upto 600M images. For model scaling, we find the location and amount of cross attention distinguishes the performance of existing UNet designs. And increasing the transformer blocks is more parameter-efficient for improving text-image alignment than increasing channel numbers. We then identify an efficient UNet variant, which is 45% smaller and 28% faster than SDXL’s UNet. On the data scaling side, we show the quality and diversity of the training set matters more than simply dataset size. Increasing caption density and diversity improves text-image alignment performance and the learning efficiency. Finally, we provide scaling functions to predict the text-image alignment performance as functions of the scale of model size, compute and dataset size.

Diffusion models have recently gained prominence as a novel category of generative models. Despite their success, these models face a notable drawback in terms of slow sampling speeds, requiring a high number of function evaluations (NFE) in the order of hundreds or thousands. In response, both learning-free and learning-based sampling strategies have been explored to expedite the sampling process. Learning-free sampling employs various ordinary differential equation (ODE) solvers based on the formulation of diffusion ODEs. However, it encounters challenges in faithfully tracking the true sampling trajectory, particularly for small NFE. Conversely, learning-based sampling methods, such as knowledge distillation, demand extensive additional training, limiting their practical applicability. To overcome these limitations, we introduce Distilled-ODE solvers (D-ODE solvers), a straightforward distillation approach grounded in ODE solver formulations. Our method seamlessly integrates the strengths of both learning-free and learning-based sampling.D-ODE solvers are constructed by introducing a single parameter adjustment to existing ODE solvers. Furthermore, we optimize D-ODE solvers with smaller steps using knowledge distillation from ODE solvers with larger steps across a batch of samples. Comprehensive experiments demonstrate the superior performance of D-ODE solvers compared to existing ODE solvers, including DDIM, PNDM, DPM-Solver, DEIS, and EDM, particularly in scenarios with fewer NFE. Notably, our method incurs negligible computational overhead compared to previous distillation techniques, facilitating straightforward and rapid integration with existing samplers. Qualitative analysis reveals that D-ODE solvers not only enhance image quality but also faithfully follow the target ODE trajectory.

We introduce a 3D-aware diffusion model, ZeroNVS, for single-image novel view synthesis for in-the-wild scenes. While existing methods are designed for single objects with masked backgrounds, we propose new techniques to address challenges introduced by in-the-wild multi-object scenes with complex backgrounds. Specifically, we train a generative prior on a mixture of data sources that capture object-centric, indoor, and outdoor scenes. To address issues from data mixture such as depth-scale ambiguity, we propose a novel camera conditioning parameterization and normalization scheme. Further, we observe that Score Distillation Sampling (SDS) tends to truncate the distribution of complex backgrounds during distillation of 360-degree scenes, and propose "SDS anchoring" to improve the diversity of synthesized novel views. Our model sets a new state-of-the-art result in LPIPS on the DTU dataset in the zero-shot setting, even outperforming methods specifically trained on DTU. We further adapt the challenging Mip-NeRF 360 dataset as a new benchmark for single-image novel view synthesis, and demonstrate strong performance in this setting. Code and models are available at https://kylesargent.github.io/zeronvs/.

We introduce the Fixed Point Diffusion Model (FPDM), a novel approach to image generation that integrates the concept of fixed point solving into the framework of diffusion-based generative modeling. Our approach embeds an implicit fixed point solving layer into the denoising network of a diffusion model, transforming the diffusion process into a sequence of closely-related fixed point problems. Combined with a new stochastic training method, this approach significantly reduces model size, reduces memory usage, and accelerates training. Moreover, it enables the development of two new techniques to improve sampling efficiency: reallocating computation across timesteps and reusing fixed point solutions between timesteps. We conduct extensive experiments with state-of-the-art models on ImageNet, FFHQ, CelebA-HQ, and LSUN-Church, demonstrating substantial improvements in performance and efficiency. Compared to the state-of-the-art DiT model, FPDM contains 87\% fewer parameters, consumes 60\% less memory during training, and improves image generation quality in situations where sampling computation or time is limited.

Efficient generation of 3D digital humans is important in several industries, including virtual reality, social media, and cinematic production. 3D generative adversarial networks (GANs) have demonstrated state-of-the-art (SOTA) quality and diversity for generated assets. Current 3D GAN architectures, however, typically rely on volume representations, which are slow to render, thereby hampering the GAN training and requiring multi-view-inconsistent 2D upsamplers. Here, we introduce Gaussian Shell Maps (GSMs) as a framework that connects SOTA generator network architectures with emerging 3D Gaussian rendering primitives using an articulable multi shell--based scaffold. In this setting, a CNN generates a 3D texture stack with features that are mapped to the shells. The latter represent inflated and deflated versions of a template surface of a digital human in a canonical body pose. Instead of rasterizing the shells directly, we sample 3D Gaussians on the shells whose attributes are encoded in the texture features. These Gaussians are efficiently and differentiably rendered. The ability to articulate the shells is important during GAN training and, at inference time, to deform a body into arbitrary user-defined poses. Our efficient rendering scheme bypasses the need for view-inconsistent upsamplers and achieves high-quality multi-view consistent renderings at a native resolution of $512 \times 512$ pixels. We demonstrate that GSMs successfully generate 3D humans when trained on single-view datasets, including SHHQ and DeepFashion. Project Page: https://rameenabdal.github.io/GaussianShellMaps/

Despite recent advances in inversion-based editing, text-guided image manipulation remains challenging for diffusion models. The primary bottlenecks include 1) the time-consuming nature of the inversion process; 2) the struggle to balance consistency with accuracy; 3) the lack of compatibility with efficient consistency sampling methods used in consistency models. To address the above issues, we start by asking ourselves if the inversion process can be eliminated for editing. We show that when the initial sample is known, a special variance schedule reduces the denoising step to the same form as the multi-step consistency sampling. We name this Denoising Diffusion Consistent Model (DDCM), and note that it implies a virtual inversion strategy without explicit inversion in sampling. We further unify the attention control mechanisms in a tuning-free framework for text-guided editing. Combining them, we present inversion-free editing (InfEdit), which allows for consistent and faithful editing for both rigid and non-rigid semantic changes, catering to intricate modifications without compromising on the image's integrity and explicit inversion. Through extensive experiments, InfEdit shows strong performance in various editing tasks and also maintains a seamless workflow (less than 3 seconds on one single A40), demonstrating the potential for real-time applications.

Recently, diffusion models have emerged as a new pow-erful generative method for 3D point cloud generationtasks. However, few works study the effect of the architec-ture of the diffusion model in the 3D point cloud, resortingto the typical UNet model developed for 2D images. In-spired by the wide adoption of Transformers, we study thecomplementary role of convolution (from UNet) and atten-tion (from Transformers). We discover that their respectiveimportance change according to the timestep in the diffu-sion process. At an early stage, the attention is given alarger weight to generate the overall shape quickly, andat a later stage, the convolution has a larger weight to re-fine the local surface’s quality of the generated point cloud.We propose a time-varying two-stream denoising modelcombined with convolution layers and transformer blocks.We generate an optimizable mask from each timestep toreweigh global and local features, obtaining time-varyingfused features. Experimentally, we demonstrate that ourproposed method quantitatively outperforms other state-of-the-art methods regarding visual quality and diversity. Thecode and models will be released upon publication.

Sampling from the posterior distribution poses a major computational challenge in solving inverse problems using latent diffusion models.Common methods rely on Tweedie's first-order moments, which are known to induce a quality-limiting bias. Existing second-order approximations are impractical due to prohibitive computational costs, making standard reverse diffusion processes intractable for posterior sampling. This paper introduces Second-order Tweedie sampler from Surrogate Loss (STSL), a novel sampler that offers efficiency comparable to first-order Tweedie with a tractable reverse process using second-order approximation. Our theoretical results reveal that the second-order approximation is lower bounded by our surrogate loss that only requires O(1) compute using the trace of the Hessian, and by the lower bound we derive a new drift term to make the reverse process tractable. Our method surpasses SoTA solvers PSLD and P2L , achieving 4X and 8X reduction in neural function evaluations, respectively, while notably enhancing sampling quality on FFHQ, ImageNet, and COCO benchmarks. In addition, we show STSL extends to text-guided image editing and addresses residual distortions present from corrupted images in leading text-guided image editing methods.To our best knowledge, this is the first work to offer an efficient second-order approximation in solving inverse problems using latent diffusion.

Concept personalization methods enable large text-to-image models to learn specific subjects (e.g., objects/poses/3D models) and synthesize renditions in new contexts. Given that the image references are highly biased towards visual attributes, state-of-the-art personalization models tend to overfit the whole subject and cannot disentangle visual characteristics in pixel space. In this study, we proposed a more challenging setting, namely fine-grained visual appearance personalization. Different from existing methods, we allow users to provide a sentence describing the desired attributes. A novel decoupled self-augmentation strategy is proposed to generate target-related and non-target samples to learn user-specified visual attributes. These augmented data allow for refining the model's understanding of the target attribute while mitigating the impact of unrelated attributes. At the inference stage, adjustments are conducted on semantic space through the learned target and non-target embeddings to further enhance the disentanglement of target attributes. Extensive experiments on various kinds of visual attributes with SOTA personalization methods shows the ability of the proposed method to mimic target attributes in novel contexts, thus improving the controllability and flexibility of personalization.