ICCV 2025 论文列表

Ensemble-based attacks have been proven to be effective in enhancing adversarial transferability by aggregating the output of models with various architectures. However, existing research primarily focuses on refining ensemble weights or optimizing the ensemble path, overlooking the exploration of ensemble models to enhance the transferability of adversarial attacks. To address this gap, we propose applying adversarial augmentation to the surrogate models, aiming to boost overall generalization of ensemble models and reduce the risk of adversarial overfitting. Meanwhile, observing that ensemble Vision Transformers (ViTs) gain less attention, we propose ViT-EnsembleAttack based on the idea of model adversarial augmentation, the first ensemble-based attack method tailored for ViTs to the best of our knowledge. Our approach generates augmented models for each surrogate ViT using three strategies: Multi-head dropping, Attention score scaling, and MLP feature mixing, with the associated parameters optimized by Bayesian optimization. These adversarially augmented models are ensembled to generate adversarial examples. Furthermore, we introduce Automatic Reweighting and Step Size Enlargement modules to boost transferability. Extensive experiments demonstrate that ViT-EnsembleAttack significantly enhances the adversarial transferability of ensemble-based attacks on ViTs, outperforming existing methods by a substantial margin. Code is available at https://github.com/Trustworthy-AI-Group/TransferAttack.

Recent advances in Large Multi-modal Models (LMMs) are primarily focused on offline video understanding. Instead, streaming video understanding poses great challenges to recent models due to its time-sensitive, omni-modal and interactive characteristics. In this work, we aim to extend the streaming video understanding from a new perspective and propose a novel task named Visual Instruction Feedback in which models should be aware of visual contents and learn to extract instructions from them. For example, when users wave their hands to agents, agents should recognize the gesture and start conversations with welcome information. Thus, following instructions in visual modality greatly enhances user-agent interactions. To facilitate research, we define seven key subtasks highly relevant to visual modality and collect the ViSpeak-Instruct dataset for training and the ViSpeak-Bench for evaluation. Further, we propose the ViSpeak model, which is a SOTA streaming video understanding LMM with GPT-4o-level performance on various streaming video understanding benchmarks. After finetuning on our ViSpeak-Instruct dataset, ViSpeak is equipped with basic visual instruction feedback ability, serving as a solid baseline for future research.

Visual Mamba networks (ViMs) extend the selective state space model (Mamba) to various vision tasks and demonstrate significant potential. As a promising compression technique, vector quantization (VQ) decomposes network weights into codebooks and assignments, significantly reducing memory usage and computational latency, thereby enabling the deployment of ViMs on edge devices. Although existing VQ methods have achieved extremely low-bit quantization (e.g., 3-bit, 2-bit, and 1-bit) in convolutional neural networks and Transformer-based networks, directly applying these methods to ViMs results in unsatisfactory accuracy. We identify several key challenges: 1) The weights of Mamba-based blocks in ViMs contain numerous outliers, significantly amplifying quantization errors. 2) When applied to ViMs, the latest VQ methods suffer from excessive memory consumption, lengthy calibration procedures, and suboptimal performance in the search for optimal codewords. In this paper, we propose ViM-VQ, an efficient post-training vector quantization method tailored for ViMs. ViM-VQ consists of two innovative components: 1) a fast convex combination optimization algorithm that efficiently updates both the convex combinations and the convex hulls to search for optimal codewords, and 2) an incremental vector quantization strategy that incrementally confirms optimal codewords to mitigate truncation errors. Experimental results demonstrate that ViM-VQ achieves state-of-the-art performance in low-bit quantization across various visual tasks.

Reliable Uncertainty Quantification (UQ) and failure prediction remain open challenges for Vision-Language Models (VLMs). We introduce ViLU, a new Vision-Language Uncertainty quantification framework that contextualizes uncertainty estimates by leveraging all task-relevant textual representations. ViLU constructs an uncertainty-aware multi-modal representation by integrating the visual embedding, the predicted textual embedding, and an image-conditioned textual representation via cross-attention. Unlike traditional UQ methods based on loss prediction, ViLU trains an uncertainty predictor as a binary classifier to distinguish correct from incorrect predictions using a weighted binary cross-entropy loss, making it loss-agnostic. In particular, our proposed approach is well-suited for post-hoc settings, where only vision and text embeddings are available without direct access to the model itself. Extensive experiments on diverse datasets show the significant gains of our method compared to state-of-the-art failure prediction methods. We apply our method to standard classification datasets, such as ImageNet-1k, as well as large-scale image-caption datasets like CC12M and LAION-400M. Ablation studies highlight the critical role of our architecture and training in achieving effective uncertainty quantification.

Recent efforts in video reasoning segmentation (VRS) integrate large language models (LLMs) with perception models to localize and track objects via textual instructions, achieving barely satisfactory results in simple scenarios. However, they struggled to discriminate and deduce the objects from user queries in more real-world scenes featured by long durations, multiple objects, rapid motion, and heavy occlusions. In this work, we analyze the underlying causes of these limitations, and present **ViLLa**: **Vi**deo reasoning segmentation with **L**arge **La**nguage Model. Remarkably, our ViLLa manages to tackle these challenges through multiple core innovations: (1) a context synthesizer that dynamically encodes the user intent with video contexts for accurate reasoning, resolving ambiguities in complex queries, and (2) a hierarchical temporal synchronizer that disentangles multi-object interactions across complex temporal scenarios by modelling multi-object interactions at local and global temporal scales. To enable efficient processing of long videos, ViLLa incorporates (3) a key segment sampler that adaptively partitions long videos into shorter but semantically dense segments for less redundancy. What's more, to promote research in this unexplored area, we construct a VRS benchmark, **VideoReasonSeg**, featuring different complex scenarios. Our model also exhibits impressive state-of-the-art results on VideoReasonSeg, Ref-YouTube-VOS, Ref-DAVIS17, MeViS, and ReVOS. Both quantitative and qualitative experiments demonstrate that our method effectively enhances video reasoning segmentation capabilities for multimodal LLMs.

We introduce ViCTr (Vital Consistency Transfer), a framework for advancing medical image synthesis through a principled integration with Rectified Flow trajectories. Unlike traditional approaches, we modify the Tweedie formulation to accommodate linear trajectories within the Rectified Flow framework, enabling more accurate initial state approximation and consistent trajectory paths. ViCTr's design allows for precise control over anatomical accuracy and pathological attributes across CT and MRI modalities via a two-stage architecture. In Stage 1, it performs anatomical learning on the ATLAS-8k dataset using Elastic Weight Consolidation (EWC) to selectively train model weights tailored for medical data. In Stage 2, an adversarial fine-tuning strategy is applied: the base model from Stage 1 remains frozen while a LoRA adapter is exclusively applied to the weights tuned in Stage 1, allowing targeted adaptation for downstream tasks while preserving the core medical data properties learned during pretraining. ViCTr achieves notable improvements by utilizing segmentation maps and textual prompts to enable refined control over CT and MRI synthesis. Extensive experiments on benchmark datasets, including BTCV, AMOS, and CirrMRI600+, demonstrate ViCTr's superiority, showing significant enhancements in quantitative metrics and clinical detail, such as liver surface nodularity in cirrhosis synthesis. These results establish ViCTr as a major advancement in medical image synthesis with impactful applications in data augmentation and clinical training.

We introduce VertexRegen, a novel mesh generation framework that enables generation at a continuous level of detail. Existing autoregressive methods generate meshes in a partial-to-complete manner and thus intermediate steps of generation represent incomplete structures. VertexRegen takes inspiration from progressive meshes and reformulates the process as the reversal of edge collapse, i.e. vertex split, learned through a generative model. Experimental results demonstrate that VertexRegen produces meshes of comparable quality to state-of-the-art methods while uniquely offering anytime generation with the flexibility to halt at any step to yield valid meshes with varying levels of detail.

Leveraging text, images, structure maps, or motion trajectories as conditional guidance, diffusion models have achieved great success in automated and high-quality video generation. However, generating smooth and rational transition videos given the first and last video frames as well as descriptive text prompts is far underexplored. We present VTG, a Versatile Transition video Generation framework that can generate smooth, high-fidelity, and semantic-coherent video transitions. VTG introduces interpolation-based initialization that helps preserve object identity and handle abrupt content changes effectively. In addition, it incorporates dual-directional motion fine-tuning and representation alignment regularization to mitigate the limitations of pre-trained image-to-video diffusion models in motion smoothness and generation fidelity, respectively. To evaluate VTG and facilitate future studies on unified transition generation, we collected TransitBench, a comprehensive benchmark for transition generation covering two representative transition tasks: concept blending and scene transition. Extensive experiments show that VTG achieves superior transition performance consistently across all four tasks.

Humans recognize objects after observing only a few examples, a remarkable capability enabled by their inherent language understanding of the real-world environment. Developing verbalized and interpretable representation can significantly improve model generalization in low-data settings. In this work, we propose Verbalized Representation Learning (VRL), a novel approach for automatically extracting human-interpretable features for object recognition using few-shot data. Our method uniquely captures inter-class differences and intra-class commonalities in the form of natural language by employing a Vision-Language Model (VLM) to identify key discriminative features between different classes and shared characteristics within the same class. These verbalized features are then mapped to numeric vectors through the VLM. The resulting feature vectors can be further utilized to train and infer with downstream classifiers. Experimental results show that, at the same model scale, VRL achieves a 24% absolute improvement over prior state-of-the-art methods while using 95% less data and a smaller model. Furthermore, compared to human-labeled attributes, the features learned by VRL exhibit a 20% absolute gain when used for downstream classification tasks.

Large-scale pre-training technology has achieved remarkable performance in diversified object re-identification (Re-ID) downstream tasks. Nevertheless, to our best knowledge, the pre-training model specifically for vehicle Re-ID, which focuses on tackling the challenge of multi-view variations, has not been fully investigated. In this paper, we first leverage a diffusion model to build a large-scale vehicle Re-ID benchmark dataset, dubbed "DiffVERI", containing over 1700K images from abundant multi-view annotations. In terms of this dataset, we further present VehicleMAE, a novel masked image modeling pre-training paradigm that learns view-invariant representations by performing mutual-distillation in a self-supervised manner. To be specific, the pipeline of VehicleMAE unfolds two core modules, i.e., view-asymmetry masked image modeling (VMIM) and past-to-present mutual-distillation (PPMD). Technically, VMIM consists of two homogeneous masked autoencoders (MAE) that simultaneously reconstruct the RGB pixels and multi-view semantic information of the specific vehicle body region via paired asymmetric mask sampling strategies. To progressively distill the knowledge of the model itself, PPMD considers the two MAEs in the current epoch and the previous one as the student models and the teacher models, respectively, which leverages the knowledge learned by the current student and the historical teacher for mutual feature-level distillation. Extensive experimental results have verified that the proposed pre-training paradigm on DiffVERI gains compelling downstream task performance for vehicle Re-ID.

Contrastive learning (CL) has become a cornerstone of self-supervised pretraining (SSP) in foundation models; however, extending CL to pixel-wise representation--crucial for medical vision--remains an open problem. Standard CL formulates SSP as a binary optimization problem (binary CL) where the excessive pursuit of feature dispersion leads to an "over-dispersion" problem, breaking pixel-wise feature correlation thus disrupting the intra-class distribution. Our vector CL reformulates CL as a vector regression problem, enabling dispersion quantification in pixel-wise pretraining via modeling feature distances in regressing displacement vectors. To implement this novel paradigm, we propose the COntrast in VEctor Regression (COVER) framework. COVER establishes an extendable vector-based self-learning, enforces a consistent optimization flow from vector regression to distance modeling, and leverages a vector pyramid architecture for granularity adaptation, thus preserving pixel-wise feature correlations in SSP. Extensive experiments across 8 tasks, spanning 2 dimensions and 4 modalities, show that COVER significantly improves pixel-wise SSP, advancing generalizable medical visual foundation models. Codes will be publicly available at https://github.com/YutingHe-list/COVER.

Increasingly expensive training of ever larger models such as Vision Transfomers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge. To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are then used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over 70% of its original performance and requires only 10 epochs of fine-tuning to regain 99% of the original accuracy while simultaneously reducing MACs by 35% and model size by 36%, thus speeding up the model by 1.44 times.

State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames (640x360) on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least 50% reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by 4.3% on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks.

In recent years, video question answering based on multimodal large language models (MLLM) has garnered considerable attention, due to the benefits from the substantial advancements in LLMs. However, these models have a notable deficiency in the domains of video temporal grounding and reasoning, posing challenges to the development of effective real-world video understanding systems. Inspired by how humans use video players to interact with the progress bar for video comprehension, we introduce VTimeCoT, a simple yet effective training-free framework, designed for high-performance video grounding and reasoning. The proposed framework incorporates two novel visual tools of the progress bar: a plug-and-play progress bar integration tool and a high-efficiency highlighting tool. In addition, to address the limitations of conventional text-based chain-of-thought (CoT) approaches, we introduce a visuotemporal CoT process that integrates cross-modality reasoning across both video and text. Our approach demonstrates significant performance improvements on both Qwen2VL-7B and GPT4o baselines in tasks of video temporal grounding and reasoning-based question answering. Finally, we showcase that the proposed framework achieves a compositional and interpretable reasoning process. Project page: \href https://vtimecot.github.io https://vtimecot.github.io .

Vision transformers have significantly advanced the field of computer vision, offering robust modeling capabilities and global receptive field. However, their high computational demands limit their applicability in processing long sequences. To tackle this issue, State Space Models (SSMs) have gained prominence in vision tasks as they offer linear computational complexity. Recently, State Space Duality (SSD), an improved variant of SSMs, was introduced in Mamba2 to enhance model performance and efficiency. However, the inherent causal nature of SSD/SSMs restricts their applications in non-causal vision tasks. To address this limitation, we introduce Visual State Space Duality (VSSD) model, which has a non-causal format of SSD. Specifically, we propose to discard the magnitude of interactions between the hidden state and tokens while preserving their relative weights, which relieves the dependencies of token contribution on previous tokens. Together with the involvement of multi-scan strategies, we show that the scanning results can be integrated to achieve non-causality, which not only improves the performance of SSD in vision tasks but also enhances its efficiency. We conduct extensive experiments on various benchmarks including image classification, detection, and segmentation, where VSSD surpasses existing state-of-the-art SSM-based models. Code is available at https://github.com/YuHengsss/VSSD.

Video super-resolution remains a significant challenge in low-level vision tasks. To date, CNN- and Transformer-based methods have delivered impressive results. However, CNNs are limited by local receptive fields, while Transformers struggle with quadratic complexity, posing challenges for processing long sequences in VSR. Recently, Mamba has drawn attention for its long-sequence modeling, linear complexity, and large receptive fields. In this work, we propose VSRM, a novel Video Super-Resolution framework that leverages the power of Mamba. VSRM introduces Spatial-to-Temporal Mamba and Temporal-to-Spatial Mamba blocks to extract long-range spatio-temporal features and enhance receptive fields efficiently. To better align adjacent frames, we propose Deformable Cross-Mamba Alignment module. This module utilizes a deformable cross-mamba mechanism to make the compensation stage more dynamic and flexible, preventing feature distortions. Finally, we minimize the frequency domain gaps between reconstructed and ground-truth frames by proposing a simple yet effective Frequency Charbonnier-like loss that better preserves high-frequency content and enhances visual quality. Through extensive experiments, VSRM achieves state-of-the-art results on diverse benchmarks, establishing itself as a solid foundation for future research.

Multimodal large language models are an exciting emerging class of language models (LMs) that have merged classic LM capabilities with those of image processing systems. However, how these capabilities integrate is often not intuitive and warrants direct investigation. One understudied capability in MLLMs is visual spatial planning---the ability to comprehend the spatial arrangements of objects and devise action plans to achieve desired outcomes in visual scenes. It is unclear why MLLMs fall short on these tasks generally considered easy for humans, given their successes across other diverse scenarios. To this end, we introduce VSP, a benchmark that 1) evaluates the spatial planning capability in MLLMs in general, and 2) diagnoses this capability via finer-grained sub-tasks, including perception and reasoning, and measure the capabilities of models through these sub-tasks. Our evaluation confirms that both open-source and private MLLMs fail to generate effective plans for even simple spatial planning tasks. Evaluations on the fine-grained analytical tasks further reveal fundamental deficiencies in the models' visual perception and bottlenecks in reasoning abilities, explaining their worse performance in the general spatial planning tasks. Our work illuminates future directions for improving MLLMs' abilities in spatial planning. Our benchmark is publicly available.

Text-to-image diffusion models have shown impressive capabilities in generating realistic visuals from natural-language prompts, yet they often struggle with accurately binding attributes to corresponding objects, especially in prompts containing multiple attribute-object pairs. This challenge primarily arises from the limitations of commonly used text encoders, such as CLIP, which can fail to encode complex linguistic relationships and modifiers effectively. Existing approaches have attempted to mitigate these issues through attention map control during inference and the use of layout information or fine-tuning during training, yet they face performance drops with increased prompt complexity. In this work, we introduce a novel compositional generation method that leverages pairwise image embeddings to improve attribute-object binding. Our approach decomposes complex prompts into sub-prompts, generates corresponding images, and computes visual prototypes that fuse with text embeddings to enhance representation. By applying segmentation-based localization training, we address cross-attention misalignment, achieving improved accuracy in binding multiple attributes to objects. Our approaches outperform existing compositional text-to-image diffusion models on the benchmark T2I CompBench, achieving better image quality, evaluated by humans, and emerging robustness under scaling the number of binding pairs in the prompt.

Knowledge distillation (KD) aims to transfer the knowledge of a more capable yet cumbersome teacher model to a lightweight student model. In recent years, relation-based KD methods have fallen behind, as their instance-matching counterparts dominate in performance. In this paper, we revive relational KD by identifying and tackling several key issues in relation-based methods, including their susceptibility to overfitting and spurious responses. Specifically, we transfer novelly constructed affinity graphs that compactly encapsulate a wealth of beneficial inter-sample, inter-class, and inter-view correlations by exploiting virtual views and relations as a new kind of knowledge. As a result, the student has access to richer guidance signals and stronger regularisation throughout the distillation process. To further mitigate the adverse impact of spurious responses, we prune the affinity graphs by dynamically detaching redundant and unreliable edges. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO datasets demonstrate the superior performance of the proposed virtual relation matching (VRM) method, where it consistently sets new state-of-the-art records over a range of models, architectures, tasks, and set-ups. For instance, VRM for the first time hits 74.0% accuracy for ResNet50-to-MobileNetV2 distillation on ImageNet, and improves DeiT-T by 14.44% on CIFAR-100 with a ResNet56 teacher.