ICCV 2025 论文列表

Recent advances in motion diffusion models have led to remarkable progress in diverse motion generation tasks, including text-to-motion synthesis.However, existing approaches represent motions as dense frame sequences, requiring the model to process redundant or less informative frames.The processing of dense animation frames imposes significant training complexity, especially when learning intricate distributions of large motion datasets even with modern neural architectures. This severely limits the performance of generative motion models for downstream tasks.Inspired by professional animators who mainly focus on sparse keyframes, we propose a novel diffusion framework explicitly designed around sparse and geometrically meaningful keyframes.Our method reduces computation by masking non-keyframes and efficiently interpolating missing frames. We dynamically refine the keyframe mask during inference to prioritize informative frames in later diffusion steps.Extensive experiments show that our approach consistently outperforms state-of-the-art methods in text alignment and motion realism, while also effectively maintaining high performance at significantly fewer diffusion steps.We further validate the robustness of our framework by using it as a generative prior and adapting it to different downstream tasks.

The research focus of GUI agents is shifting from text-dependent to pure-vision-based approaches, which, though promising, prioritize comprehensive pre-training data collection while neglecting contextual modeling challenges. We probe the characteristics of element and history contextual modeling in GUI agents and summarize: **1) the high-density and loose-relation of element context** highlight the existence of many unrelated elements and their negative influence; **2) the high redundancy of history context** reveals the inefficient history modeling in current GUI agents. In this work, we propose a context-aware simplification framework for building an efficient and effective GUI Agent, termed **SimpAgent**. To mitigate potential interference from numerous unrelated elements, we introduce a **masking-based element pruning** method that circumvents the intractable relation modeling through an efficient masking mechanism. To reduce the redundancy in historical information, we devise a **consistency-guided history compression** module, which enhances implicit LLM-based compression through innovative explicit guidance, achieving an optimal balance between performance and efficiency. With the above components, SimpAgent reduces 27% FLOPs and achieves superior GUI navigation performances. Comprehensive navigation experiments across diverse web and mobile environments demonstrate the effectiveness and potential of our agent.

This work focuses on the task of privacy-preserving action recognition (PPAR), which aims to protect individual privacy in action videos without compromising recognition performance. Despite recent advancements, existing PPAR models still struggle with video domain shifts. To address this challenge, this work aims to develop transferable PPAR models by leveraging labeled videos from the source domain and unlabeled videos from the target domain. This work contributes a novel method named GenPriv, which improves the transferability of privacy-preserving models by generative decoupled learning. Inspired by the fact that privacy-sensitive information in action videos primarily comes from static human appearances, our GenPriv decouples video features into static and dynamic aspects and then removes privacy-sensitive content from static action features. We propose a generative architecture, ST-VAE, complemented by Spatial Consistency and Temporal Alignment losses, to enhance decoupled learning. Experimental results on three benchmarks with diverse domain shifts demonstrate the effectiveness of our proposed GenPriv.

Machine Unlearning has recently garnered significant attention, aiming to selectively remove knowledge associated with specific data while preserving the model's performance on the remaining data. A fundamental challenge in this process is balancing effective unlearning with knowledge retention, as naive optimization of these competing objectives can lead to conflicting gradients, hindering convergence and degrading overall performance. To address this issue, we propose Learning to Unlearn while Retaining, aimed to mitigate gradient conflicts between unlearning and retention objectives. Our approach strategically avoids conflicts through an implicit gradient regularization mechanism that emerges naturally within the proposed framework. This prevents conflicting gradients between unlearning and retention, leading to effective unlearning while preserving the model's utility. We validate our approach across both discriminative and generative tasks, demonstrating its effectiveness in achieving unlearning without compromising performance on remaining data. Our results highlight the advantages of avoiding such gradient conflicts, outperforming existing methods that fail to account for these interactions.

Learning-based methods have made promising advances in low-light RAW image enhancement, while their capability to extremely dark scenes where the environmental illuminance drops as low as 0.0001 lux remains to be explored due to the lack of corresponding datasets. To this end, we propose a paired-to-paired data synthesis pipeline capable of generating well-calibrated extremely low-light RAW images at three precise illuminance ranges of 0.01-0.1 lux, 0.001-0.01 lux, and 0.0001-0.001 lux, together with high-quality sRGB references to comprise a large-scale paired dataset named See-in-the-Extremely-Dark (SIED) to benchmark low-light RAW image enhancement approaches. Furthermore, we propose a diffusion-based framework that leverages the generative ability and intrinsic denoising property of diffusion models to restore visually pleasing results from extremely low-SNR RAW inputs, in which an Adaptive Illumination Correction Module (AICM) and a color consistency loss are introduced to ensure accurate exposure correction and color restoration. Extensive experiments on the proposed SIED and publicly available benchmarks demonstrate the effectiveness of our method. The code and dataset are available at https://github.com/JianghaiSCU/SIED.

Computer vision seeks to infer a wide range of information about objects and events. However, vision systems based on conventional imaging are limited to extracting information only from the visible surfaces of scene objects. For instance, a vision system can detect and identify a Coke can in the scene, but it cannot determine whether the can is full or empty. In this paper, we aim to expand the scope of computer vision to include the novel task of inferring the hidden liquid levels of opaque containers by sensing the tiny vibrations on their surfaces. Our method provides a first-of-a-kind way to inspect the fill level of multiple sealed containers remotely, at once, without needing physical manipulation and manual weighing. First, we propose a novel speckle-based vibration sensing system for simultaneously capturing scene vibrations on a 2D grid of points. We use our system to efficiently and remotely capture a dataset of vibration responses for a variety of everyday liquid containers. Then, we develop a transformer-based approach for analyzing the captured vibrations and classifying the container type and its hidden liquid level at the time of measurement. Our architecture is invariant to the vibration source, yielding correct liquid level estimates for controlled and ambient scene sound sources. Moreover, our model generalizes to unseen container instances within known classes (e.g., training on five Coke cans of a six-pack, testing on a sixth) and fluid levels. We demonstrate our method by recovering liquid levels from various everyday containers.

Multimodal Large Language Models (MLLMs) have shown impressive capabilities in visual reasoning, yet come with substantial computational cost, limiting their deployment in resource-constrained settings. Despite recent effort on improving the efficiency of MLLMs, prior solutions fall short in responding to varying runtime conditions, in particular changing resource availability (e.g., contention due to the execution of other programs on the device). To bridge this gap, we introduce AdaLLaVA, an adaptive inference framework that learns to dynamically reconfigure operations in an MLLM during inference, accounting for the input data and a latency budget. We conduct extensive experiments across benchmarks involving question-answering, reasoning, and hallucination. Our results show that AdaLLaVA effectively adheres to input latency budget, achieving varying accuracy and latency tradeoffs at runtime. Further, we demonstrate that AdaLLaVA adapts to both input latency and content, can be integrated with token selection for enhanced efficiency, and generalizes across MLLMs.

Leveraging the effective visual-text alignment and static generalizability from CLIP, recent video learners adopt CLIP initialization with further regularization or recombination for generalization in open-vocabulary action recognition in-context. However, due to the static bias of CLIP, such video learners tend to overfit on shortcut static features, thereby compromising their generalizability, especially to novel out-of-context actions. To address this issue, we introduce Open-MeDe, a novel Meta-optimization framework with static Debiasing for Open-vocabulary action recognition. From a fresh perspective of generalization, Open-MeDe adopts a meta-learning approach to improve known-to-open generalizing and image-to-video debiasing in a cost-effective manner. Specifically, Open-MeDe introduces a cross-batch meta-optimization scheme that explicitly encourages video learners to quickly generalize to arbitrary subsequent data via virtual evaluation, steering a smoother optimization landscape. In effect, the free of CLIP regularization during optimization implicitly mitigates the inherent static bias of the video meta-learner. We further apply self-ensemble over the optimization trajectory to obtain generic optimal parameters that can achieve robust generalization to both in-context and out-of-context novel data. Extensive evaluations show that Open-MeDe not only surpasses state-of-the-art regularization methods tailored for in-context open-vocabulary action recognition but also substantially excels in out-of-context scenarios.

Object navigation tasks require an agent to locate a target object using visual observations in unseen environments, where unfamiliar layouts and novel object appearances can hinder navigation. Most existing methods lack the adaptability needed to handle these uncertainties, as their navigation models remain fixed during testing. In this paper, we address this challenge by examining object-conditioned trajectory distribution shifts in navigation caused by changes in environmental dynamics. We propose learning a central conditional distribution as a prior that approximates the specific distributions of diverse environments. To retain environment-specific information during navigation, we allow each environment-specific distribution to approximate this central distribution rather than relying on it directly. To implement this, we introduce a meta-learning mechanism that integrates with traditional navigation methods, offering tailored solutions for various types of navigation approaches. Our approach, Learning on the Go (LOG), enables agents to learn on the go, allowing for flexible, adaptive, real-time learning during navigation. Our theoretical analysis highlights the benefits of learning a central distribution for effective generalization across environments, and empirical results confirm the proposed method's effectiveness, demonstrating superior performance compared to existing approaches.

Humans possess an exceptional ability to imagine 4D scenes, encompassing both motion and 3D geometry, from a single still image. This ability is rooted in our accumulated observations of similar scenes and an intuitive understanding of physics. In this paper, we aim to replicate this capacity in neural networks, specifically focusing on natural fluid imagery. Existing methods for this task typically employ simplistic 2D motion estimators to animate the image, leading to motion predictions that often defy physical principles, resulting in unrealistic animations. Our approach introduces a novel method for generating 4D scenes with physics-consistent animation from a single image. We propose the use of a physics-informed neural network that predicts motion for each surface point, guided by a loss term derived from fundamental physical principles, including the Navier-Stokes equations. To capture appearance, we predict feature-based 3D Gaussians from the input image and its estimated depth, which are then animated using the predicted motions and rendered from any desired camera perspective. Experimental results highlight the effectiveness of our method in producing physically plausible animations, showcasing significant performance improvements over existing methods. Our project page is \href https://physfluid.github.io/ https://physfluid.github.io/ .

Class-Incremental Semantic Segmentation (CISS) requires continuous learning of newly introduced classes while retaining knowledge of past classes. By abstracting mainstream methods into two stages (visual feature extraction and prototype-feature matching), we identify a more fundamental challenge termed catastrophic semantic entanglement. This phenomenon involves Prototype-Feature Entanglement caused by semantic misalignment during the incremental process, and Background-Increment Entanglement due to dynamic data evolution. Existing techniques, which rely on visual feature learning without sufficient cues to distinguish targets, introduce significant noise and errors. To address these issues, we introduce a Language-inspired Bootstrapped Disentanglement framework (LBD). We leverage the prior class semantics of pre-trained visual-language models (e.g., CLIP) to guide the model in autonomously disentangling features through Language-guided Prototypical Disentanglement and Manifold Mutual Background Disentanglement. The former guides the disentangling of new prototypes by treating hand-crafted text features as topological templates, while the latter employs multiple learnable prototypes and mask-pooling-based supervision for background-incremental class disentanglement. By incorporating soft prompt tuning and encoder adaptation modifications, we further bridge the capability gap of CLIP between dense and sparse tasks, achieving state-of-the-art performance on both Pascal VOC and ADE20k, particularly in multi-step scenarios.

Compositional Zero-Shot Learning (CZSL) aims to recognize novel attribute-object compositions by leveraging knowledge from seen compositions. Existing methods typically align textual prototypes with visual features using Vision-Language Models (VLMs), but they face two key limitations: (1) modality gaps hinder the ability to distinguish semantically similar attribute-object pairs, and (2) textual prototypes alone lack the fine-grained visual cues needed for accurate recognition. To address these challenges, we propose Visual Proxy Learning, a method that reduces modality gaps and enhances compositional generalization by initializing visual proxies for attributes, objects, and their compositions from text representations and optimizing the visual space to better capture fine-grained visual cues. To further strengthen cross-modal understanding, we introduce Cross-Modal Joint Learning (CMJL), which enforces consistency between text-image embeddings and fine-grained visual representations. This dual strategy improves generalization to unseen compositions and enhances the discrimination of similar pairs. Extensive experiments demonstrate that our method achieves state-of-the-art performance in closed-world settings and competitive results in open-world scenarios across four CZSL benchmarks, validating its effectiveness in improving compositional generalization.

Structuring latent representations in a hierarchical manner enables models to learn patterns at multiple levels of abstraction. However, most prevalent image understanding models focus on visual similarity, and learning visual hierarchies is relatively unexplored. In this work, for the first time, we introduce a learning paradigm that can encode user-defined multi-level complex visual hierarchies in hyperbolic space without requiring explicit hierarchical labels. As a concrete example, first, we define a part-based image hierarchy using object-level annotations within and across images. Then, we introduce an approach to enforce the hierarchy using contrastive loss with pairwise entailment metrics. Finally, we discuss new evaluation metrics to effectively measure hierarchical image retrieval. Encoding these complex relationships ensures that the learned representations capture semantic and structural information that transcends mere visual similarity. Experiments in part-based image retrieval show significant improvements in hierarchical retrieval tasks, demonstrating the capability of our model in capturing visual hierarchies.

Understanding continuous video streams plays a fundamental role in real-time applications, including embodied AI and autonomous driving. Unlike offline video processing, streaming video understanding requires the ability to process video streams frame by frame, preserve historical information, and make low-latency decisions. To address these challenges, our main contributions are three-fold. (i) we develop a novel streaming video backbone, termed as StreamFormer, by incorporating causal temporal attention into a pre-trained vision transformer. This enables efficient streaming video processing while maintaining image representation capability. (ii) to train StreamFormer, we propose to unify diverse spatial-temporal video understanding tasks within a multitask visual-language alignment framework. Hence, StreamFormer learns global semantics, temporal dynamics, and fine-grained spatial relationships simultaneously. (iii) we conduct extensive experiments for online action detection, online video instance segmentation, and video question answering. StreamFormer achieves competitive performance while maintaining efficiency, demonstrating its potential for real-time applications.

Learning fine-grained representations from coarse labels for fine-grained visual recognition (FGVR) is a challenging yet valuable task, as it alleviates the reliance on labor-intensive fine-grained annotations. Early approaches focused primarily on minimizing intra-fine-grained-class variation but overlooked inter-fine-grained-class separability, resulting in limited FGVR performance. Subsequent studies employed a top-down paradigm to enhance separability via deep clustering, yet these methods require predefining the number of fine-grained classes, which is often impractical to obtain. Here, we introduce a bottom-up learning paradigm that constructs a hierarchical dendrogram by iteratively merging similar instances/clusters, inferring higher-level semantics from lowest-level instances without predefining class numbers. Leveraging this, we propose BuCSFR, a novel method that integrates a Bottom-up Construction (BuC) module to build the dendrogram based on a minimal information loss criterion, and a Separable Fine-grained Representation (SFR) module that treats dendrogram nodes as pseudo-labels to ensure representation separability. The synergistic interaction between these modules enables iterative enhancement, grounded theoretically in the Expectation-Maximization (EM) framework. Extensive experiments on five benchmark datasets demonstrate the superiority of our approach, showcasing its effectiveness in learning separable representations for FGVR. The source code is available at: https://github.com/BeCarefulOfYournaoke/BuCSFR.

Recently, learning-based stereo matching networks have advanced significantly.However, they often lack robustness and struggle to achieve impressive cross-domain performance due to domain shifts and imbalanced disparity distributions among diverse datasets.Leveraging Vision Foundation Models (VFMs) can intuitively enhance the model's robustness, but integrating such a model into stereo matching cost-effectively to fully realize their robustness remains a key challenge.To address this, we propose SMoEStereo, a novel framework that adapts VFMs for stereo matching through a tailored, scene-specific fusion of Low-Rank Adaptation (LoRA) and Mixture-of-Experts (MoE) modules.SMoEStereo introduces MoE-LoRA with adaptive ranks and MoE-Adapter with adaptive kernel sizes. The former dynamically selects optimal experts within MoE to adapt varying scenes across domains, while the latter injects inductive bias into frozen VFMs to improve geometric feature extraction.Importantly, to mitigate computational overhead, we further propose a lightweight decision network that selectively activates MoE modules based on input complexity, balancing efficiency with accuracy. Extensive experiments demonstrate that our method exhibits state-of-the-art cross-domain and joint generalization across multiple benchmarks without dataset-specific adaptation. The code is available at https://github.com/cocowy1/SMoE-Stereo.

Watermarking as a traceable authentication technology has been widely applied in image copyright protection. However, most existing watermarking methods embed watermarks by adding irremovable perturbations to the cover image, causing permanent distortion. To address this issue, we propose a novel watermarking approach termed Cover-Recoverable Watermark (CRMark). CRMark can losslessly recover the cover image and watermark in lossless channels and enables robust watermark extraction in lossy channels. CRMark leverages an integer Invertible Watermarking Network (iIWN) to achieve a lossless invertible mapping between the cover-image-watermark pair and the stego image. During the training phase, CRMark employs an encoder-noise-layer-decoder architecture to enhance its robustness against distortions. In the inference phase, CRMark first maps the cover-image-watermark pair into an overflowed stego image and a latent variable. Subsequently, the overflowed pixels and the latent variable are losslessly compressed into an auxiliary bitstream, which is then embedded into the clipped stego image using reversible data hiding. During extraction, in lossy channels, the noised stego image can directly undergo inverse mapping via iIWN to extract the watermark. In lossless channels, the latent variable and overflowed stego image are first recovered using reversible data hiding, followed by watermark extraction through iIWN. Extensive experimental results demonstrate that CRMark can be perfectly recovered in lossless channels while remaining robust to common distortions. Code is available at https://github.com/chenoly/CRMark.

Affordance, defined as the potential actions that an object offers, is crucial for embodied AI agents. For example, such knowledge directs an agent to grasp a knife by the handle for cutting or by the blade for safe handover. While existing approaches have made notable progress, affordance research still faces three key challenges: data scarcity, poor generalization, and real-world deployment. Specifically, there is a lack of large-scale affordance datasets with precise segmentation maps, existing models struggle to generalize across different domains or novel object and affordance classes, and little work demonstrates deployability in real-world scenarios. In this work, we address these issues by proposing a complete affordance learning system that (1) takes in egocentric videos and outputs precise affordance annotations without human labeling, (2) leverages geometric information and vision foundation models to improve generalization, and (3) introduces a framework that facilitates affordance-oriented robotic manipulation such as tool grasping and robot-to-human tool handover. Experimental results show that our model surpasses the state-of-the-art by 13.8% in mIoU, and the framework achieves 77.1% successful grasping among 179 trials, including evaluations on seen, unseen classes, and cluttered scenes. Project page: https://reagan1311.github.io/affgrasp.

Deep learning-based bilateral grid processing has emerged as a promising solution for image enhancement, inherently encoding spatial and intensity information while enabling efficient full-resolution processing through slicing operations. However, existing approaches are limited to linear affine transformations, hindering their ability to model complex color relationships. Meanwhile, while multi-layer perceptrons (MLPs) excel at non-linear mappings, traditional MLP-based methods employ globally shared parameters, which is hard to deal with localized variations. To overcome these dual challenges, we propose a Bilateral Grid-based Pixel-Adaptive Multi-layer Perceptron (BPAM) framework. Our approach synergizes the spatial modeling of bilateral grids with the non-linear capabilities of MLPs. Specifically, we generate bilateral grids containing MLP parameters, where each pixel dynamically retrieves its unique transformation parameters and obtain a distinct MLP for color mapping based on spatial coordinates and intensity values. In addition, we propose a novel grid decomposition strategy that categorizes MLP parameters into distinct types stored in separate subgrids. Multi-channel guidance maps are used to extract category-specific parameters from corresponding subgrids, ensuring effective utilization of color information during slicing while guiding precise parameter generation. Extensive experiments on public datasets demonstrate that our method outperforms state-of-the-art methods in performance while maintaining real-time processing capabilities.