CVPR 2025 Papers

With the rise of generative models, there is a growing interest in unifying all tasks within a generative framework. Anomaly detection methods also fall into this scope and utilize diffusion models to generate or reconstruct normal samples when given arbitrary anomaly images. However, our study found that the diffusion model suffers from severe faithfulness hallucination and catastrophic forgetting, which can't meet the unpredictable pattern increments. To mitigate the above problems, we propose a continual diffusion model that uses gradient projection to achieve stable continual learning. Gradient projection deploys a regularization on the model updating by modifying the gradient towards the direction protecting the learned knowledge. But as a double-edged sword, it also requires huge memory costs brought by the Markov process. Hence, we propose an iterative singular value decomposition method based on the transitive property of linear representation, which consumes tiny memory and incurs almost no performance loss. Finally, considering the risk of over-fitting to normal images of the diffusion model, we propose an anomaly-masked network to enhance the condition mechanism of the diffusion model. For continual anomaly detection, ours achieves first place in 17/18 settings on MVTec and VisA.

Text-to-Image (T2I) diffusion models have made remarkable advancements in generative modeling; however, they face a trade-off between inference speed and image quality, posing challenges for efficient deployment. Existing distilled T2I models can generate high-fidelity images with fewer sampling steps, but often struggle with diversity and quality, especially in one-step models. From our analysis, we observe redundant computations in the UNet encoders. Our findings suggest that, for T2I diffusion models, decoders are more adept at capturing richer and more explicit semantic information, while encoders can be effectively shared across decoders from diverse time steps.Based on these observations, we introduce the first Time-independent Unified Encoder (TiUE) for the student model UNet architecture, which is a loop-free image generation approach for distilling T2I diffusion models. Using a one-pass scheme, TiUE shares encoder features across multiple decoder time steps, enabling parallel sampling and significantly reducing inference time complexity. In addition, we incorporate a KL divergence term to regularize noise prediction, which enhances the perceptual realism and diversity of the generated images. Experimental results demonstrate that TiUE outperforms state-of-the-art methods, including LCM, SD-Turbo, and SwiftBrushv2, producing more diverse and realistic results while maintaining the computational efficiency.

Transformers today still struggle to generate one-minute videos because self-attention layers are inefficient for long context. Alternatives such as Mamba layers struggle to produce coherent scenes because their hidden states are small and less expressive. We experiment with Test-Time Training (TTT) layers, whose hidden states themselves can be neural networks, therefore larger and more expressive. Adding TTT layers into a pre-trained Transformer enables it to generate one-minute videos from text storyboards. We curate a dataset based on Tom and Jerry cartoons as a proof-of-concept benchmark. Compared to baselines such as Mamba 2, Gated DeltaNet, and sliding-window attention layers, TTT layers generate much more coherent videos that tell complete stories, leading by 34 Elo points in a human evaluation of 100 videos per method. Although promising, our results are still limited in physical realism, and the efficiency of our implementation can be further improved. Sample videos, code and annotations are available at: https://test-time-training.github.io/video-dit

Collaborative perception in autonomous driving significantly enhances the perception capabilities of individual agents. Immutable heterogeneity in collaborative perception, where agents have different and fixed perception networks, presents a major challenge due to the semantic gap in their exchanged intermediate features without modifying the perception networks. Most existing methods bridge the semantic gap through interpreters. However, they either require training a new interpreter for each new agent type, limiting extensibility, or rely on a two-stage interpretation via an intermediate standardized semantic space, causing cumulative semantic loss. To achieve both extensibility in immutable heterogeneous scenarios and low-loss feature interpretation, we propose PolyInter, a polymorphic feature interpreter. It contains an extension point through which emerging new agents can seamlessly integrate by overriding only their specific prompts, which are learnable parameters intended to guide the interpretation, while reusing PolyInter's remaining parameters. By leveraging polymorphism, our design ensures that single interpreter is sufficient to accommodate diverse agents and interpret their features into the ego agent's semantic space. Experiments conducted on OPV2V dataset demonstrate that PolyInter improves collaborative perception precision by up to 11.1% compared to SOTA interpreters, while comparable results can be achieved by training only 1.4% of PolyInter's parameters when adapting to new agents.

Advanced image fusion methods mostly prioritise high-level missions, where task interaction struggles with semantic gaps, requiring complex bridging mechanisms. In contrast, we propose to leverage low-level vision tasks from digital photography fusion, allowing for effective feature interaction through pixel-level supervision. This new paradigm provides strong guidance for unsupervised multimodal fusion without relying on abstract semantics, enhancing task-shared feature learning for broader applicability. Owning to the hybrid image features and enhanced universal representations, the proposed GIFNet supports diverse fusion tasks, achieving high performance across both seen and unseen scenarios with a single model. Uniquely, experimental results reveal that our framework also supports single-modality enhancement, offering superior flexibility for practical applications. Our code will be available at https://github.com/AWCXV/GIFNet.

We introduce \texttt OneDiffusion - a single large-scale diffusion model designed to tackle a wide range of image synthesis and understanding tasks. It can generate images conditioned on text, depth, pose, layout, or semantic maps. It also handles super-resolution, multi-view generation, instant personalization, and text-guided image editing. Additionally, the same model is also effective for image understanding tasks such as depth estimation, pose estimation, open-vocab segmentation and camera pose estimation, etc.Our unified model is trained with simple but effective recipe: we casting all tasks as modeling a sequence of frames where we inject different noise scales during training to each frame. During inference, we can set any frame to a clean image for tasks like conditional image generation, camera pose estimation, or generating paired depth and images from a caption. Experiment results shows our proposed method achieve competitive performance on wide range of tasks. By eliminating the need for specialized architectures or finetuning, OneDiffusion provides enhanced flexibility and scalability, reflecting the emerging trend towards general-purpose diffusion models in the field.

Vision-language model (VLM) is one of the most important models for multi-modal tasks. Real industrial applications often meet the challenge of adapting VLMs to different scenarios, such as varying hardware platforms or performance requirements. Traditional methods involve training or fine-tuning to adapt multiple unique VLMs or using model compression techniques to create multiple compact models. These approaches are complex and resource-intensive. This paper introduces a novel paradigm called Once-Tuning-Multiple-Variants (OTMV). OTMV requires only a single tuning process to inject dynamic weight expansion capacity into the VLM with dynamic expansion capacity. This tuned VLM can then be expanded into multiple variants tailored for different scenarios in inference. The tuning mechanism of OTMV is inspired by the mathematical series expansion theorem, which helps to reduce the parameter size and memory requirements while maintaining accuracy for VLM. Experiment results show that OTMV-tuned models achieve comparable accuracy to baseline VLMs across various visual-language tasks. The experiments also demonstrate the dynamic expansion capability of OTMV-tuned VLMs, outperforming traditional model compression and adaptation techniques in terms of accuracy and efficiency.

Event cameras provide low-latency perception for only milliwatts of power. This makes them highly suitable for resource-restricted, agile robots such as small flying drones. Self-supervised learning based on contrast maximization holds great potential for event-based robot vision, as it foregoes the need for high-frequency ground truth and allows for online learning in the robot's operational environment. However, online, on-board learning raises the major challenge of achieving sufficient computational efficiency for real-time learning, while maintaining competitive visual perception performance. In this work, we improve the time and memory efficiency of the contrast maximization pipeline, making on-device learning of low-latency monocular depth possible. We demonstrate that online learning on board a small drone yields more accurate depth estimates and more successful obstacle avoidance behavior compared to only pre-training. Benchmarking experiments show that the proposed pipeline is not only efficient, but also achieves state-of-the-art depth estimation performance among self-supervised approaches. Our work taps into the unused potential of online, on-device robot learning, promising smaller reality gaps and better performance.

Pre-trained Vision-Language Models (VLMs) like CLIP, have demonstrated strong zero-shot generalization capabilities. Despite their effectiveness on various downstream tasks, they remain vulnerable to adversarial samples. Existing methods fine-tune VLMs to improve their performance via performing adversarial training on a certain dataset. However, this can lead to model overfitting and is not a true zero-shot scenario. In this paper, we propose a truly zero-shot and training-free approach that can significantly improve the VLM's zero-shot adversarial robustness. Specifically, we first discover that simply adding Gaussian noise greatly enhances the VLM's zero-shot performance. Then, we treat the adversarial examples with added Gaussian noise as anchors and strive to find a path in the embedding space that leads from the adversarial examples to the cleaner samples. We improve the VLMs' generalization abilities in a truly zero-shot and training-free manner compared to previous methods. Extensive experiments on 16 datasets demonstrate that our method can achieve state-of-the-art zero-shot robust performance, improving the top-1 robust accuracy by an average of 9.77%. The code will be publicly available.

We investigate the generalization boundaries of current Large Multimodal Models (LMMs) via comprehensive evaluation under out-of-distribution scenarios and domain-specific tasks. We evaluate their zero-shot generalization across synthetic images, real-world distributional shifts, and specialized datasets like medical and molecular imagery. Empirical results indicate that LMMs struggle with generalization beyond common training domains, limiting their direct application without adaptation. To understand the cause of unreliable performance, we analyze three hypotheses: semantic misinterpretation, visual feature extraction insufficiency, and mapping deficiency. Results identify mapping deficiency as the primary hurdle. To address this problem, we show that in-context learning (ICL) can significantly enhance LMMs' generalization, opening new avenues for overcoming generalization barriers. We further explore the robustness of ICL under distribution shifts and show its vulnerability to domain shifts, label shifts, and spurious correlation shifts between in-context examples and test data.

Recent advances in Handwritten Text Recognition (HTR) have led to significant reductions in transcription errors on standard benchmarks under the i.i.d. assumption, thus focusing on minimizing in-distribution (ID) errors.However, this assumption does not hold in real-world applications, which has motivated HTR research to explore Transfer Learning and Domain Adaptation techniques. In this work, we investigate the unaddressed limitations of HTR models in generalizing to out-of-distribution (OOD) data. We adopt the challenging setting of Domain Generalization, where models are expected to generalize to OOD data without any prior access. To this end, we analyze 336 OOD cases from eight state-of-the-art HTR models across seven widely used datasets, spanning five languages. Additionally, we study how HTR models leverage synthetic data to generalize. We reveal that the most significant factor for generalization lies in the textual divergence between domains, followed by visual divergence. We demonstrate that the error of HTR models in OOD scenarios can be reliably estimated, with discrepancies falling below 10 points in 70% of cases. We identify the underlying limitations of HTR models, laying the foundation for future research to address this challenge.

Video large language models (Video-LLMs) can temporally ground language queries and retrieve video moments. Yet, such temporal comprehension capabilities are neither well-studied nor understood. So we conduct a study on prediction consistency -- a key indicator for robustness and trustworthiness of temporal grounding. After the model identifies an initial moment within the video content, we apply a series of probes to check if the model's responses align with this initial grounding as an indicator of reliable comprehension. Our results reveal that current Video-LLMs are sensitive to variations in video contents, language queries, and task settings, unveiling severe deficiencies in maintaining consistency. We further explore common prompting and instruction-tuning methods as potential solutions, but find that their improvements are often unstable. To that end, we propose event temporal verification tuning that explicitly accounts for consistency, and demonstrate significant improvements for both grounding and consistency. Our data and code are open-sourced at https://github.com/minjoong507/Consistency-of-Video-LLM.

To capture individual gait patterns, excluding identity-irrelevant cues in walking videos, such as clothing texture and color, remains a persistent challenge for vision-based gait recognition. Traditional silhouette and pose-based methods, though theoretically effective at removing such distractions, often fall short of high accuracy due to their sparse and less informative inputs. To address this, emerging end-to-end methods focus on directly denoising RGB videos using global optimization and human-defined priors. Building on this trend, we propose a novel gait denoising method, DenosingGait. Inspired by the philosophy that "what I cannot create, I do not understand", we turn to generative diffusion models, uncovering how these models can partially filter out irrelevant factors for improved gait understanding. Based on this generation-driven denoising, we introduce feature matching, a kind of popular geometrical constraint in optical flow and depth estimation, to compact multi-channel float-encoded RGB information into two-channel direction vectors that represent local structural features, where within-frame matching captures spatial details and cross-frame matching conveys temporal dynamics. Experiments on the CCPG, CAISA-B*, and SUSTech1K datasets demonstrate that DenoisingGait achieves a new SoTA performance in most cases for both within-domain and cross-domain evaluations. Code is available at https://github.com/ShiqiYu/OpenGait.

Panoramic imagery, with its 360deg field of view, offers comprehensive information to support Multi-Object Tracking (MOT) in capturing spatial and temporal relationships of surrounding objects. However, most MOT algorithms are tailored for pinhole images with limited views, impairing their effectiveness in panoramic settings. Additionally, panoramic image distortions, such as resolution loss, geometric deformation, and uneven lighting, hinder direct adaptation of existing MOT methods, leading to significant performance degradation. To address these challenges, we propose OmniTrack, an omnidirectional MOT framework that incorporates Tracklet Management to introduce temporal cues, FlexiTrack Instances for object localization and association, and the CircularStatE Module to alleviate image and geometric distortions. This integration enables tracking in panoramic field-of-view scenarios, even under rapid sensor motion. To mitigate the lack of panoramic MOT datasets, we introduce the QuadTrack dataset--a comprehensive panoramic dataset collected by a quadruped robot, featuring diverse challenges such as panoramic fields of view, intense motion, and complex environments. Extensive experiments on the public JRDB dataset and the newly introduced QuadTrack benchmark demonstrate the state-of-the-art performance of the proposed framework. OmniTrack achieves a HOTA score of 26.92% on JRDB, representing an improvement of 3.43%, and further achieves 23.45% on QuadTrack, surpassing the baseline by 6.81%. The established dataset and source code are available at https://github.com/xifen523/OmniTrack.

Understanding fine-grained temporal dynamics is crucial in egocentric videos, where continuous streams capture frequent, close-up interactions with objects. In this work, we bring to light that current egocentric video question-answering datasets often include questions that can be answered using only few frames or commonsense reasoning, without being necessarily grounded in the actual video. Our analysis shows that state-of-the-art Multi-Modal Large Language Models (MLLMs) on these benchmarks achieve remarkably high performance using just text or a single frame as input. To address these limitations, we introduce EgoTempo, a dataset specifically designed to evaluate temporal understanding in the egocentric domain. EgoTempo emphasizes tasks that require integrating information across the entire video, ensuring that models would need to rely on temporal patterns rather than static cues or pre-existing knowledge. Extensive experiments on EgoTempo show that current MLLMs still fall short in temporal reasoning on egocentric videos, and thus we hope EgoTempo will catalyze new research in the field and inspire models that better capture the complexity of temporal dynamics. Dataset and code are available at https://github.com/google-research-datasets/egotempo.git.

In this paper, we introduce OmniStyle-1M, a large-scale paired style transfer dataset comprising over one million content-style-stylized image triplets across 1,000 diverse style categories, each enhanced with textual descriptions and instruction prompts. We show that OmniStyle-1M can not only improve the generalization of style transfer models through supervised training but also facilitates precise control over target stylization. Especially, to ensure the quality of the dataset, we introduce OmniFilter, a comprehensive style transfer quality assessment framework, which filters high-quality triplets based on content preservation, style consistency, and aesthetic appeal. Building upon this foundation, we propose OmniStyle, a framework based on the Diffusion Transformer (DiT) architecture designed for high-quality and efficient style transfer. This framework supports both instruction-guided and image-guided style transfer, generating high resolution outputs with exceptional detail. Extensive qualitative and quantitative evaluations demonstrate OmniStyle's superior performance compared to existing approaches, highlighting its efficiency and versatility. OmniStyle-1M and its accompanying methodologies provide a significant contribution to advancing high-quality style transfer, offering a valuable resource for the research community.

Fast and reliable omnidirectional 3D sensing is essential to many applications such as autonomous driving, robotics and drone navigation. While many well-recognized methods have been developed to produce high-quality omnidirectional 3D information, they are too slow for real-time computation, limiting their feasibility in practical applications. Motivated by these shortcomings, we propose an efficient omnidirectional depth sensing framework, called OmniStereo, which generates high-quality 3D information in real-time. Unlike prior works, OmniStereo employs Cassini projection to simplify the photometric matching and introduces a lightweight stereo matching network to minimize computational overhead. Additionally, OmniStereo proposes a novel fusion method to handle depth discontinuities and invalid pixels complemented by a refinement module to reduce mapping-introduced errors and recover fine details. As a result, OmniStereo achieves state-of-the-art (SOTA) accuracy, surpassing the second-best method over 32% in MAE, while maintaining real-time efficiency. It operates more than 16.5xfaster than the second-best method in accuracy on TITAN RTX, achieving 12.3 FPS on embedded device Jetson AGX Orin, underscoring its suitability for real-world deployment. The code is available at https://github.com/DengJiaxi1/OmniStereo.

The rapid advancement of multi-modal language models (MLLMs) like GPT-4o has propelled the development of Omni language models, designed to process and proactively respond to continuous streams of multi-modal data. Despite their potential, evaluating their real-world interactive capabilities in streaming video contexts remains a formidable challenge. In this work, we introduce OmniMMI, a comprehensive multi-modal interaction benchmark tailored for OmniLLMs in streaming video contexts. OmniMMI encompasses over 1,121 videos and 2,290 questions, addressing two critical yet underexplored challenges in existing video benchmarks: streaming video understanding and proactive reasoning, across six distinct subtasks. Moreover, we propose a novel framework, Multi-modal Multiplexing Modeling (M4), designed to enable an inference-efficient streaming model that can see, listen while generating.

With the rapid growth of generative AI and its widespread application in image editing, new risks have emerged regarding the authenticity and integrity of digital content. Existing versatile watermarking approaches suffer from trade-offs between tamper localization precision and visual quality. Constrained by the limited flexibility of previous framework, their localized watermark must remain fixed across all images. Under AIGC-editing, their copyright extraction accuracy is also unsatisfactory. To address these challenges, we propose OmniGuard, a novel augmented versatile watermarking approach that integrates proactive embedding with passive, blind extraction for robust copyright protection and tamper localization. OmniGuard employs a hybrid forensic framework that enables flexible localization watermark selection and introduces a degradation-aware tamper extraction network for precise localization under challenging conditions. Additionally, a lightweight AIGC-editing simulation layer is designed to enhance robustness across global and local editing. Extensive experiments show that OmniGuard achieves superior fidelity, robustness, and flexibility. Compared to the recent state-of-the-art approach EditGuard, our method outperforms it by 4.25dB in PSNR of the container image, 20.7% in F1-Score under noisy conditions, and 14.8% in average bit accuracy.