CVPR 2025 论文列表

Blind Face Video Restoration (BFVR) focuses on reconstructing high-quality facial image sequences from degraded video inputs. The main challenge is address unknown degradations, while maintaining temporal consistency across frames. Current blind face restoration methods are primarily designed for images, and directly applying these approaches to BFVR will encounter a significant drop in restoration performance. In this work, we proposed Dynamic Content Prediction with Motion-aware Priors, referred to as DCP-MP. We develop a motion-aware semantic dictionary by encoding the semantic information of high-quality videos into discrete elements, and capturing the motion information in terms of element relationships, which are derived from the dynamic temporal changes within videos. For the purpose of utilizing dictionary to represent the degraded video, we train a temporal-aware element predictor, conditioned on degraded content, to learn the prediction of discrete elements in dictionary. The predicted elements will be refined, conditioned on motion information captured by the motion-aware semantic dictionary, to enhance temporal coherence. To alleviate deviation from the original structure information, we propose a conditional structure feature correction module that corrects the features flowing from the encoder to the generator. Through extensive experiments, we validate the effectiveness of our design components and demonstrate the superior performance of DCP-MP in synthesizing high-quality video.

One important task of multimodal models is to translate referred image regions to human preferred language descriptions. Existing methods, however, ignore the resolution adaptability needs of different tasks, which hinders them to find out precise language descriptions. In this study, we propose a DynRefer approach, to pursue high-accuracy region-level referring through mimicking the resolution adaptability of human visual cognition. During training, DynRefer stochastically aligns language descriptions of multimodal tasks with images of multiple resolutions, which are constructed by nesting a set of random views around the referred region. This process essentially constructs a set of region representations, where suitable representations for specific tasks can be matched. During inference, DynRefer performs selectively multimodal referring by sampling proper region representations for tasks from the set of views based on image and task priors. This allows the visual information for referring to better match human preferences, thereby improving the representational adaptability of region-level multimodal models. Experiments show that DynRefer brings mutual improvement upon broad tasks including region-level captioning, open-vocabulary region recognition and attribute detection. Furthermore, DynRefer achieves state-of-the-art results on multiple region-level multimodal tasks using a single model. Code is available at https://github.com/callsys/DynRefer.

The challenge in LLM-based video understanding lies in preserving visual and semantic information in long videos while maintaining a memory-affordable token count. However, redundancy and correspondence in videos have hindered the performance potential of existing methods. Through statistical learning on current datasets, we observe that redundancy occurs in both repeated and answer-irrelevant frames, and the corresponding frames vary with different questions. This suggests the possibility of adopting dynamic encoding to balance detailed video information preservation with token budget reduction. To this end, we propose a dynamic cooperative network, DynFocus, for memory-efficient video encoding in this paper. Specifically, i) a Dynamic Event Prototype Estimation (DPE) module to dynamically select meaningful frames for question answering; (ii) a Compact Cooperative Encoding (CCE) module that encodes meaningful frames with detailed visual appearance and the remaining frames with sketchy perception separately. We evaluate our method on five publicly available benchmarks, and experimental results consistently demonstrate that our method achieves competitive performance.

Video large language models (VLLMs) have significantly advanced recently in processing complex video content. Yet, their inference efficiency remains constrained because of the high computational cost stemming from the thousands of visual tokens generated from the video inputs. We empirically observe that, unlike single image inputs, VLLMs typically attend visual tokens from different frames at different decoding iterations. This makes a one-shot pruning strategy prone to removing important tokens by mistake. Motivated by this, we present DyCoke, a training-free token compression method to optimize token representation and accelerate VLLMs. DyCoke incorporates a plug-and-play temporal compression module to minimize temporal redundancy by merging redundant tokens across frames and applying dynamic KV cache reduction to prune spatially redundant tokens selectively. It ensures high-quality inference by dynamically retaining the critical tokens at each decoding step. Extensive experimental results demonstrate that DyCoke can outperform the prior SoTA counterparts, achieving 1.5x inference speedup, and 1.4x memory reduction against the baseline VLLM, while still improving the performance, with no training.

Semi-supervised learning in medical image segmentation leverages unlabeled data to reduce annotation burdens through consistency learning. However, current methods struggle with class imbalance and high uncertainty from pathology variations, leading to inaccurate segmentation in 3D medical images. To address these challenges, we present DyCON, a Dynamic Uncertainty-aware Consistency and Contrastive Learning framework that enhances the generalization of consistency methods with two complementary losses: Uncertainty-aware Consistency Loss (UnCL) and Focal Entropy-aware Contrastive Loss (FeCL). UnCL enforces global consistency by dynamically weighting the contribution of each voxel to the consistency loss based on its uncertainty, preserving high-uncertainty regions instead of filtering them out. Initially, UnCL prioritizes learning from uncertain voxels with lower penalties, encouraging the model to explore challenging regions. As training progress, the penalty shift towards confident voxels to refine predictions and ensure global consistency. Meanwhile, FeCL enhances local feature discrimination in imbalanced regions by introducing dual focal mechanisms and adaptive confidence adjustments into the contrastive principle. These mechanisms jointly prioritizes hard positives and negatives while focusing on uncertain sample pairs, effectively capturing subtle lesion variations under class imbalance. Extensive evaluations on four diverse medical image segmentation datasets (ISLES'22, BraTS'19, LA, Pancreas) show DyCON's superior performance against SOTA methods(\href https://dycon25.github.io/) Project Page: https://dycon25.github.io/ .

Pansharpening aims at integrating complementary information from panchromatic and multispectral images. Available deep-learning based pansharpening methods typically perform exceptionally with particular satellite datasets. At the same time, it has been observed that these models also exhibit scene dependence, for example, if the majority of the training samples come from the urban scenes, the model's performance may decline in the river scene. To address the domain gap produced by varying satellite sensors and distinct scenes, we propose a dual-granularity semantic guided sparse routing diffusion model for general pansharpening. By utilizing the large Vision-Language Models (VLMs) in the field of geoscience, e.g, GeoChat, we introduce the dual granularity semantics to generate dynamic sparse routing scores for adaptation of different satellite sensors and scenes. This scene-level and region-level dual-granularity semantic information serves as guidance for dynamically activating specialized experts within the diffusion model. Extensive experiments on WorldView-3, QuickBird, and GaoFen-2 datasets show the effectiveness of our proposed method. Notably, the proposed method outperforms the comparison approaches in adapting to new satellite sensors and scenes. The codes are available at https://github.com/codgodtao/SGDiff.

Open-vocabulary semantic segmentation aims to enable models to segment arbitrary categories. Currently, though pre-trained Vision-Language Models (VLMs) like CLIP have established a robust foundation for this task by learning to match text and image representations from large-scale data, their lack of pixel-level recognition necessitates further fine-tuning. Most existing methods leverage text as a guide to achieve pixel-level recognition. However, the inherent biases in text semantic descriptions and the lack of pixel-level supervisory information make it challenging to fine-tune CLIP-based models effectively. This paper considers leveraging image-text data to simultaneously capture the semantic information contained in both image and text, thereby constructing Dual Semantic Guidance and corresponding pixel-level pseudo annotations. Particularly, the visual semantic guidance is enhanced via explicitly exploring foreground regions and minimizing the influence of background. The dual semantic guidance is then jointly utilized to fine-tune CLIP-based segmentation models, achieving decent fine-grained recognition capabilities. As the comprehensive evaluation shows, our method outperforms state-of-art results with large margins, on eight commonly used datasets with/without background.

Achieving robust stereo 3D imaging under diverse illumination conditions is an important however challenging task, largely due to the limited dynamic ranges (DRs) of cameras, which are significantly smaller than real world DR. As a result, the accuracy of existing stereo depth estimation methods is often compromised by under- or over-exposed images.In this work, we introduce dual-exposure stereo for extended dynamic range 3D imaging. We develop automatic dual-exposure control method that adjusts the dual exposures, diverging them when the scene DR exceeds the camera DR, thereby providing information about broader DR. From the captured dual-exposure stereo images, we estimate depth by developing a motion-aware dual-exposure stereo depth network.To validate our proposed method, we develop a robot-vision system, collect real-world stereo video datasets, and generate a synthetic dataset. Our approach outperforms traditional exposure control and depth estimation methods.

Diffusion models have gained tremendous success in text-to-image generation, yet still struggle with visual understanding tasks, an area dominated by autoregressive vision-language models. We propose a large-scale and fully end-to-end diffusion model for multi-modal understanding and generation that significantly improves on existing diffusion-based multimodal models, and is the first of its kind to support the full suite of vision-language modeling capabilities. Inspired by the multimodal diffusion transformer (MM-DiT) and recent advances in discrete diffusion language modeling, we leverage a cross-modal maximum likelihood estimation framework that simultaneously trains the conditional likelihoods of both images and text jointly under a single loss function, which is back-propagated through both branches of the diffusion transformer. The resulting model is highly flexible and capable of a wide range of tasks including image generation, captioning, and visual question answering. Our model attained competitive performance compared to recent unified image understanding and generation models, demonstrating the potential of multimodal diffusion modeling as a promising alternative to autoregressive next-token prediction models.

Aiming to achieve class-agnostic perception in visual object tracking, current trackers commonly formulate tracking as a one-shot detection problem with the template-matching architecture. Despite the success, severe environmental variations in long-term tracking raise challenges to generalizing the tracker in novel situations. Temporal trackers try to fix it by preserving the time-validity of target information with historical predictions, e.g., updating the template. However, solely transmitting the previous observations instead of learning from them leads to an inferior capability of understanding the tracking scenario from past experience, which is critical for the generalization in new frames. To address this issue, we reformulate temporal learning in visual tracking as a History-to-Future process and propose a novel tracking framework DreamTrack. Our DreamTrack learns the temporal dynamics from past observations to dream the future variations of the environment, which boosts the generalization with the extended future information from history. Considering the uncertainty of future variation, multimodal prediction is designed to infer the target trajectory of each possible future situation. The experiments demonstrate that our DreamTrack achieves leading performance with real-time inference speed. In particular, DreamTrack obtains SUC scores of 76.6%/87.9% on LaSOT/TrackingNet, surpassing all recent SOTA trackers.

Adapting generative models to specific domains presents an effective solution for satisfying specialized requirements. However, adapting to some complex domains remains challenging, especially when these domains require substantial paired data to capture the targeted distributions. Since unpaired data from a single modality, such as vision or language, is more readily available, we utilize the bidirectional mappings between vision and language learned by the unified generative model to enable training on unpaired data for domain adaptation. Specifically, we propose DoraCycle, which integrates two multimodal cycles: text-to-image-to-text and image-to-text-to-image. The model is optimized through cross-entropy loss computed at the cycle endpoints, where both endpoints share the same modality. This facilitates self-evolution of the model without reliance on annotated text-image pairs. Experimental results demonstrate that for tasks independent of paired knowledge, such as stylization, DoraCycle can effectively adapt the unified model using only unpaired data. For tasks involving new paired knowledge, such as specific identities, a combination of a small set of paired image-text examples and larger-scale unpaired data is sufficient for effective domain-oriented adaptation. The code will be released at https://github.com/showlab/DoraCycle.

Domain generalization (DG) aims to train a model on source domains that can generalize well to unseen domains. Recent advances in Vision-Language Models (VLMs), such as CLIP, exhibit remarkable generalization capabilities across a wide range of data distributions, benefiting tasks like DG. However, CLIP is pre-trained by aligning images with their descriptions, which inevitably captures domain-specific details. Moreover, adapting CLIP to source domains with limited feature diversity introduces bias. These limitations hinder the model's ability to generalize across domains. In this paper, we propose a new DG approach by learning with diverse text prompts. These text prompts incorporate varied contexts to imitate different domains, enabling DG model to learn domain-invariant features. The text prompts guide DG model learning in three aspects: feature suppression, which uses these prompts to identify domain-sensitive features and suppress them; feature consistency, which ensures the model's features are robust to domain variations imitated by the diverse prompts; and feature diversification, which diversifies features based on the prompts to mitigate bias. Experimental results show that our approach improves domain generalization performance on five datasets on the DomainBed benchmark, achieving state-of-the-art results.

Large multimodal models (LMMs) have achieved impressive progress in vision-language understanding, yet they face limitations in real-world applications requiring complex reasoning over a large number of images. Existing benchmarks for multi-image question-answering are limited in scope, each question is paired with only up to 30 images, which does not fully capture the demands of large-scale retrieval tasks encountered in the real-world usages. To reduce these gaps, we introduce two document haystack benchmarks, dubbed DocHaystack and InfoHaystack, designed to evaluate LMM performance on large-scale visual document retrieval and understanding. Additionally, we propose V-RAG, a novel, vision-centric retrieval-augmented generation (RAG) framework that leverages a suite of multimodal vision encoders, each optimized for specific strengths, and a dedicated question-document relevance module. V-RAG sets a new standard, with a 9% and 11% improvement in Recall@1 on the challenging DocHaystack-1000 and InfoHaystack-1000 benchmarks, respectively, compared to the previous best baseline models. Additionally, integrating V-RAG with LMMs enables them to efficiently operate across thousands of images, yielding significant improvements on our DocHaystack and InfoHaystack benchmarks. Our code and datasets will be made publicly available.

Despite significant progress in multimodal large language models (MLLMs), their performance on complex, multi-page document comprehension remains inadequate, largely due to the lack of high-quality, document-level datasets. While current retrieval-augmented generation (RAG) methods offer partial solutions, they suffer from issues, such as fragmented retrieval contexts, multi-stage error accumulation, and extra time costs of retrieval. In this work, we present a high-quality document-level dataset, Doc-750K, designed to support in-depth understanding of multimodal documents. This dataset includes diverse document structures, extensive cross-page dependencies, and real question-answer pairs derived from the original documents. Building on the dataset, we develop a native multimodal model--Docopilot, which can accurately handle document-level dependencies without relying on RAG. Experiments demonstrate that Docopilot achieves superior coherence, accuracy, and efficiency in document understanding tasks and multi-turn interactions, setting a new baseline for document-level multimodal understanding. Data, code, and models are released at https://github.com/OpenGVLab/Docopilot.

Vision-Language Models (VLMs) excel in diverse visual tasks but face challenges in document understanding, which requires fine-grained text processing. While typical visual tasks perform well with low-resolution inputs, reading-intensive applications demand high-resolution, resulting in significant computational overhead. Using OCR-extracted text in VLM prompts partially addresses this issue but underperforms compared to full-resolution counterpart, as it lacks the complete visual context needed for optimal performance.We introduce DocVLM, a method that integrates an OCR-based modality into VLMs to enhance document processing while preserving original weights. Our approach employs an OCR encoder to capture textual content and layout, compressing these into a compact set of learned queries incorporated into the VLM. Comprehensive evaluations across leading VLMs show that DocVLM significantly reduces reliance on high-resolution images for document understanding.In limited-token regimes (448x448), DocVLM with 64 learned queries improves DocVQA results from 56.0% to 86.6% when integrated with InternVL2 and from 84.4% to 91.2% with Qwen2-VL. In LLaVA-OneVision, DocVLM achieves improved results while using 80% less image tokens. The reduced token usage allows processing multiple pages effectively, showing impressive zero-shot results on DUDE and state-of-the-art performance on MP-DocVQA, highlighting DocVLM's potential for applications requiring high-performance and efficiency.

Computer vision foundation models, such as DINO or OpenCLIP, are trained in a self-supervised manner on large image datasets. Analogously, substantial evidence suggests that the human visual system (HVS) is influenced by the statistical distribution of colors and patterns in the natural world, characteristics also present in the training data of foundation models. The question we address in this paper is whether foundation models trained on natural images mimic some of the low-level characteristics of the human visual system, such as contrast detection, contrast masking, and contrast constancy. Specifically, we designed a protocol comprising nine test types to evaluate the image encoders of 45 foundation and generative models. Our results indicate that some foundation models (e.g., DINO, DINOv2, and OpenCLIP), share some of the characteristics of human vision, but other models show little resemblance. Foundation models tend to show smaller sensitivity to low contrast and rather irregular responses to contrast across frequencies. The foundation models show the best agreement with human data in terms of contrast masking. Our findings suggest that human vision and computer vision may take both similar and different paths when learning to interpret images of the real world. Overall, while differences remain, foundation models trained on vision tasks start to align with low-level human vision, with DINOv2 showing the closest resemblance. Our code is available on https://github.com/caiyancheng/VFM_HVS_CVPR2025.

Large Multimodal Models (LMMs) have emerged as powerful models capable of understanding various data modalities, including text, images, and videos. LMMs encode both text and visual data into tokens that are then combined and processed by an integrated Large Language Model (LLM). Including visual tokens substantially increases the total token count, often by thousands. The increased input length for LLM significantly raises the complexity of inference, resulting in high latency in LMMs. To address this issue, token pruning methods, which remove part of the visual tokens, are proposed. The existing token pruning methods either require extensive calibration and fine-tuning or rely on suboptimal importance metrics which results in increased redundancy among the retained tokens. In this paper, we first formulate token pruning as Max-Min Diversity Problem (MMDP) where the goal is to select a subset such that the diversity among the selected tokens is maximized. Then, we solve the MMDP to obtain the selected subset and prune the rest. The proposed method, DivPrune, reduces redundancy and achieves the highest diversity of the selected tokens. By ensuring high diversity, the selected tokens better represent the original tokens, enabling effective performance even at high pruning ratios without requiring fine-tuning. Extensive experiments with various LMMs show that DivPrune achieves state-of-the-art accuracy over 16 image- and video-language datasets. Additionally, DivPrune reduces both the end-to-end latency and GPU memory usage for the tested models. The code is available here.

Multimodal Large Language Models (MLLMs) bridge the gap between visual and textual data, enabling a range of advanced applications. However, complex internal interactions among visual elements and their alignment with text can introduce vulnerabilities, which may be exploited to bypass safety mechanisms. To address this, we analyze the relationship between image content and task and find that the complexity of subimages, rather than their content, is key. Building on this insight, we propose the Distraction Hypothesis, followed by a novel framework called Contrasting Subimage Distraction Jailbreaking (CS-DJ), to achieve jailbreaking by disrupting MLLMs alignment through multi-level distraction strategies. CS-DJ consists of two components: structured distraction, achieved through query decomposition that induces a distributional shift by fragmenting harmful prompts into sub-queries, and visual-enhanced distraction, realized by constructing contrasting subimages to disrupt the interactions among visual elements within the model. This dual strategy disperses the model's attention, reducing its ability to detect and mitigate harmful content. Extensive experiments across five representative scenarios and four popular closed-source MLLMs, including \texttt GPT-4o-mini , \texttt GPT-4o , \texttt GPT-4V , and \texttt Gemini-1.5-Flash , demonstrate that CS-DJ achieves average success rates of 52.40% for the attack success rate and 74.10% for the ensemble attack success rate. These results reveal the potential of distraction-based approaches to exploit and bypass MLLMs' defenses, offering new insights for attack strategies.

Audio Descriptions (ADs) aim to provide a narration of a movie in text form, describing non-dialogue-related narratives, such as characters, actions, or scene establishment. Automatic generation of ADs remains challenging due to: i) the domain gap between movie-AD data and existing data used to train vision-language models, and ii) the issue of contextual redundancy arising from highly similar neighboring visual clips in a long movie. In this work, we propose **DistinctAD**, a novel two-stage framework for generating ADs that emphasize distinctiveness to produce better narratives. To address the domain gap, we introduce a CLIP-AD adaptation strategy that does not require additional AD corpora, enabling more effective alignment between movie and AD modalities at both global and fine-grained levels. In Stage-II, DistinctAD incorporates two key innovations: (i) a Contextual Expectation-Maximization Attention (EMA) module that reduces redundancy by extracting common bases from consecutive video clips, and (ii) an explicit distinctive word prediction loss that filters out repeated words in the context, ensuring the prediction of unique terms specific to the current AD. Comprehensive evaluations on MAD-Eval, CMD-AD, and TV-AD benchmarks demonstrate the superiority of DistinctAD, with the model consistently outperforming baselines, particularly in Recall@k/N, highlighting its effectiveness in producing high-quality, distinctive ADs.