CVPR 2024 论文列表

Diffusion models have recently gained unprecedented attention in the field of image synthesis due to their remarkable generative capabilities. Notwithstanding their prowess these models often incur substantial computational costs primarily attributed to the sequential denoising process and cumbersome model size. Traditional methods for compressing diffusion models typically involve extensive retraining presenting cost and feasibility challenges. In this paper we introduce DeepCache a novel training-free paradigm that accelerates diffusion models from the perspective of model architecture. DeepCache capitalizes on the inherent temporal redundancy observed in the sequential denoising steps of diffusion models which caches and retrieves features across adjacent denoising stages thereby curtailing redundant computations. Utilizing the property of the U-Net we reuse the high-level features while updating the low-level features in a very cheap way. This innovative strategy in turn enables a speedup factor of 2.3xfor Stable Diffusion v1.5 with only a 0.05 decline in CLIP Score and 4.1xfor LDM-4-G with a slight decrease of 0.22 in FID on ImageNet. Our experiments also demonstrate DeepCache's superiority over existing pruning and distillation methods that necessitate retraining and its compatibility with current sampling techniques. Furthermore we find that under the same throughput DeepCache effectively achieves comparable or even marginally improved results with DDIM or PLMS.

Diffusion model-based image restoration (IR) aims to use diffusion models to recover high-quality (HQ) images from degraded images achieving promising performance. Due to the inherent property of diffusion models most existing methods need long serial sampling chains to restore HQ images step-by-step resulting in expensive sampling time and high computation costs. Moreover such long sampling chains hinder understanding the relationship between inputs and restoration results since it is hard to compute the gradients in the whole chains. In this work we aim to rethink the diffusion model-based IR models through a different perspective i.e. a deep equilibrium (DEQ) fixed point system called DeqIR. Specifically we derive an analytical solution by modeling the entire sampling chain in these IR models as a joint multivariate fixed point system. Based on the analytical solution we can conduct parallel sampling and restore HQ images without training. Furthermore we compute fast gradients via DEQ inversion and found that initialization optimization can boost image quality and control the generation direction. Extensive experiments on benchmarks demonstrate the effectiveness of our method on typical IR tasks and real-world settings.

We demonstrate text as a strong cross-modal interface. Rather than relying on deep embeddings to connect image and language as the interface representation our approach represents an image as text from which we enjoy the interpretability and flexibility inherent to natural language. We employ an autoencoder that uses a pre-trained text-to-image diffusion model for decoding. The encoder is trained to transform an input image into text which is then fed into the fixed text-to-image diffusion decoder to reconstruct the original input a process we term De-Diffusion. Experiments validate both the precision and comprehensiveness of De-Diffusion text representing images such that it can be readily ingested by off-the-shelf text-to-image tools and LLMs for diverse multi-modal tasks. For example a single De-Diffusion model can generalize to provide transferable prompts for different text-to-image tools and also achieves a new state of the art on open-ended vision-language tasks by simply prompting large language models with few-shot examples. Project page: https://dediffusion.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 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 DanceCamera3D a transformer-based diffusion model that incorporates a novel body 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. Code and video demos are available at https://github.com/ Carmenw1203/DanceCamera3D-Official.

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.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^4M) an efficient framework for dataset distillation. Compared to architecture-dependent methods D^4M employs latent diffusion model to guarantee consistency and incorporates label information into category prototypes. The distilled datasets are versatile eliminating the need for repeated generation of distinct datasets for various architectures. Through comprehensive experiments D^4M demonstrates superior performance and robust generalization surpassing the SOTA methods across most aspects.

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 to faster training convergence and a to reduction in necessary sampling steps to achieve comparable or superior results. We hope DREAM will inspire a rethinking of diffusion model training paradigms.

The recovery of occluded human meshes poses challenges for current methods due to the difficulty in extracting effective image features under severe occlusion. In this paper we introduce DPMesh an innovative framework for occluded human mesh recovery that capitalizes on the profound knowledge about object structure and spatial relationships embedded in a pre-trained text-to-image diffusion model. Unlike previous methods reliant on conventional backbones for vanilla feature extraction DPMesh seamlessly integrates the pre-trained denoising U-Net with potent priors as its image backbone and performs a single-step inference to provide occlusion-aware information. To enhance the perception capability for occluded poses DPMesh incorporates judicious guidance via condition injection which produces effective controls from 2D observations for the denoising U-Net. Furthermore we explore a dedicated noisy key-point reasoning approach to mitigate disturbances arising from occlusion and crowded scenarios. This strategy fully unleashes the perceptual capability of the diffusion prior thereby enhancing accuracy. Extensive quantitative and qualitative experiments affirm the efficacy of our framework as we outperform state-of-the-art methods on both occlusion-specific and standard datasets underscoring its ability to achieve precise and robust 3D human mesh recovery particularly in challenging scenarios involving occlusion and crowded scenes. Code is available at https://github.com/EternalEvan/DPMesh.

We introduce Diffusion Parametric Head Models (DPHMs) a generative model that enables robust volumetric head reconstruction and tracking from monocular depth sequences. While recent volumetric head models such as NPHMs can now excel in representing high-fidelity head geometries tracking and reconstructing heads from real-world single-view depth sequences remains very challenging as the fitting to partial and noisy observations is underconstrained. To tackle these challenges we propose a latent diffusion-based prior to regularize volumetric head reconstruction and tracking. This prior-based regularizer effectively constrains the identity and expression codes to lie on the underlying latent manifold which represents plausible head shapes. To evaluate the effectiveness of the diffusion-based prior we collect a dataset of monocular Kinect sequences consisting of various complex facial expression motions and rapid transitions. We compare our method to state-of-the-art tracking methods and demonstrate improved head identity reconstruction as well as robust expression tracking.

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 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 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. Our project page is https://tianhao-qi.github.io/DEADiff/.

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 do 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 build 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 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. The resulting assistant can perform customized generation in 2-5 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.

Crowd counting is a fundamental problem in crowd analysis which is typically accomplished by estimating a crowd density map and summing over the density values. However this approach suffers from background noise accumulation and loss of density due to the use of broad Gaussian kernels to create the ground truth density maps. This issue can be overcome by narrowing the Gaussian kernel. However existing approaches perform poorly when trained with ground truth density maps with broad kernels. To deal with this limitation we propose using conditional diffusion models to predict density maps as diffusion models show high fidelity to training data during generation. With that we present CrowdDiff that generates the crowd density map as a reverse diffusion process. Furthermore as the intermediate time steps of the diffusion process are noisy we incorporate a regression branch for direct crowd estimation only during training to improve the feature learning. In addition owing to the stochastic nature of the diffusion model we introduce producing multiple density maps to improve the counting performance contrary to the existing crowd counting pipelines. We conduct extensive experiments on publicly available datasets to validate the effectiveness of our method. CrowdDiff outperforms existing \sota crowd counting methods on several public crowd analysis benchmarks with significant improvements. CrowdDiff project is available at: https://dylran.github.io/crowddiff.github.io.

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 5000 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.

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 models: (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 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 1488x1255 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) training framework. It seamlessly decomposes the cross-attention features in 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. Project page: https://cosmicman-cvpr2024.github.io/.

Co-salient object detection (CoSOD) aims to identify the common and salient (usually in the foreground) regions across a given group of images. Although achieving significant progress state-of-the-art CoSODs could be easily affected by some adversarial perturbations leading to substantial accuracy reduction. The adversarial perturbations can mislead CoSODs but do not change the high-level semantic information (e.g. concept) of the co-salient objects. In this paper we propose a novel robustness enhancement framework by first learning the concept of the co-salient objects based on the input group images and then leveraging this concept to purify adversarial perturbations which are subsequently fed to CoSODs for robustness enhancement. Specifically we propose CosalPure containing two modules i.e. group-image concept learning and concept-guided diffusion purification. For the first module we adopt a pre-trained text-to-image diffusion model to learn the concept of co-salient objects within group images where the learned concept is robust to adversarial examples. For the second module we map the adversarial image to the latent space and then perform diffusion generation by embedding the learned concept into the noise prediction function as an extra condition. Our method can effectively alleviate the influence of the SOTA adversarial attack containing different adversarial patterns including exposure and noise. The extensive results demonstrate that our method could enhance the robustness of CoSODs significantly.

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. Our code is available at https://github.com/UCSB-NLP-Chang/diffusion_resampling.git.

Gestures play a key role in human communication. Recent methods for co-speech gesture generation while managing to generate beat-aligned motions struggle generating gestures that are semantically aligned with the utterance. Compared to beat gestures that align naturally to the audio signal semantically coherent gestures require modeling the complex interactions between the language and human motion and can be controlled by focusing on certain words. Therefore we present ConvoFusion a diffusion-based approach for multi-modal gesture synthesis which can not only generate gestures based on multi-modal speech inputs but can also facilitate controllability in gesture synthesis. Our method proposes two guidance objectives that allow the users to modulate the impact of different conditioning modalities (e.g. audio vs text) as well as to choose certain words to be emphasized during gesturing. Our method is versatile in that it can be trained either for generating monologue gestures or even the conversational gestures. To further advance the research on multi-party interactive gestures the DnD Group Gesture dataset is released which contains 6 hours of gesture data showing 5 people interacting with one another. We compare our method with several recent works and demonstrate effectiveness of our method on a variety of tasks. We urge the reader to watch our supplementary video at https://vcai.mpi-inf.mpg.de/projects/ConvoFusion/.

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.

Score distillation sampling (SDS) and its variants have greatly boosted the development of text-to-3D generation but are vulnerable to geometry collapse and poor textures yet. To solve this issue we first deeply analyze the SDS and find that its distillation sampling process indeed corresponds to the trajectory sampling of a stochastic differential equation (SDE): SDS samples along an SDE trajectory to yield a less noisy sample which then serves as a guidance to optimize a 3D model. However the randomness in SDE sampling often leads to a diverse and unpredictable sample which is not always less noisy and thus is not a consistently correct guidance explaining the vulnerability of SDS. Since for any SDE there always exists an ordinary differential equation (ODE) whose trajectory sampling can deterministically and consistently converge to the desired target point as the SDE we propose a novel and effective "Consistent3D" method that explores the ODE deterministic sampling prior for text-to-3D generation. Specifically at each training iteration given a rendered image by a 3D model we first estimate its desired 3D score function by a pre-trained 2D diffusion model and build an ODE for trajectory sampling. Next we design a consistency distillation sampling loss which samples along the ODE trajectory to generate two adjacent samples and uses the less noisy sample to guide another more noisy one for distilling the deterministic prior into the 3D model. Experimental results show the efficacy of our Consistent3D in generating high-fidelity and diverse 3D objects and large-scale scenes as shown in Fig. 1. The codes are available at https://github.com/sail-sg/Consistent3D.