ICCV 2025 论文列表

Current human motion synthesis frameworks rely on global action descriptions, creating a modality gap that limits both motion understanding and generation capabilities. A single coarse description, such as "run", fails to capture essential details like variations in speed, limb positioning, and kinematic dynamics, leading to significant ambiguities between text and motion modalities. To address this challenge, we introduce KinMo, a unified framework built on a hierarchical describable motion representation that extends beyond global action by incorporating kinematic group movements and their interactions.We design an automated annotation pipeline to generate high-quality, fine-grained descriptions for this decomposition, resulting in the KinMo dataset. To leverage these structured descriptions, we propose Hierarchical Text-Motion Alignment, improving spatial understanding by integrating additional motion details. Furthermore, we introduce a coarse-to-fine generation procedure to demonstrate how enhanced spatial understanding benefits motion synthesis. Experimental results show that KinMo significantly improves motion understanding, demonstrated by enhanced text-motion retrieval performance and enabling more fine-grained motion generation and editing capabilities. Project Page: https://andypinxinliu.github.io/KinMo

Vision language models (VLMs) demonstrate strong capabilities in jointly processing visual and textual data. However, they often incur substantial computational overhead due to redundant visual information, particularly in long-form video scenarios. Existing approaches predominantly focus on either vision token pruning, which may overlook spatio-temporal dependencies, or keyframe selection, which identifies informative frames but discards others, thus disrupting contextual continuity. In this work, we propose KVTP (Keyframe-oriented Vision Token Pruning), a novel framework that overcomes the drawbacks of token pruning and keyframe selection. By adaptively assigning pruning rates based on frame relevance to the query, KVTP effectively retains essential contextual information while significantly reducing redundant computation. To thoroughly evaluate the long-form video understanding capacities of VLMs, we curated and reorganized subsets from VideoMME, EgoSchema, and NextQA into a unified benchmark named SparseKV-QA that highlights real-world scenarios with sparse but crucial events. Our experiments with VLMs of various scales show that KVTP can reduce token usage by 80% without compromising spatiotemporal and contextual consistency, significantly cutting computation while maintaining the performance. These results demonstrate our approach's effectiveness in efficient long-video processing, facilitating more scalable VLM deployment.

In this paper, we introduce Part-Aware Point Grounded Description (PaPGD), a challenging task aimed at advancing 3D multimodal learning for fine-grained, part-aware segmentation grounding and detailed explanation of 3D objects. Existing 3D datasets largely focus on either vision-only part segmentation or vision-language scene segmentation, lacking the fine-grained multimodal segmentation needed for robotic navigation and interaction in real-world environments. To address this gap, we present the 3DCoMPaT Grounded Instructions (3DCoMPaT-GrIn) Dataset, a comprehensive resource that pairs rich point cloud descriptions with corresponding part-level segmentation masks. This dataset encompasses extensive samples designed for both PaPGD and fine-grained single-part grounding tasks. To tackle the inherent challenges of grounding objects and generating grounded descriptions at the part level, we propose Kestrel, a part-aware 3D multimodal large language model that integrates an advanced language model for nuanced language comprehension with multi-level point feature propagation and query refinement mechanism to enhance spatial reasoning at the part level. The extensive experiments demonstrate that Kestrel effectively bridges the gap between part-aware language understanding and 3D segmentation grounding, paving the way for more robust and interpretable 3D object comprehension that meets the demands of real-world robotic applications.

Based on pseudo-labels, voxel-wise contrastive learning (VCL) is a prominent approach designed to learn effective feature representations for semi-supervised medical image segmentation. However, in multi-organ segmentation (MoS), the complex anatomical structures of certain organs often lead to many unreliable pseudo-labels. Directly applying VCL can introduce confirmation bias, resulting in poor segmentation performance. A common practice is to first transform these unreliable pseudo-labels into complementary ones, which represent classes that voxels are least likely to belong to, and then push voxels away from the generated complementary labels. However, we find that this approach may fail to allow voxels with unreliable pseudo-labels (unreliable voxels) to fully benefit from the advantages of VCL. In this paper, we propose DVCL, a novel distance-aware VCL method for semi-supervised MoS. DVCL is based on the observation that unreliable voxels, which may not form discriminative feature boundaries, still form clear clusters. Hence, voxels close to each other in the feature space ('neighbors') likely belong to the same semantic class, while distant ones ('outsiders') likely belong to different classes. In DVCL, we first identify neighbors and outsiders for all unreliable voxels, and then pull their neighbors into the same clusters while pushing outsiders away. In this way, unreliable voxels can learn more discriminative features, thereby fully enjoying the advantages of VCL. However, DVCL itself will inevitably introduce the problem of noisy neighbors and outliers. To address these challenges, we further propose a neighbor partitioning strategy and a query outlier strategy to provide more stable feature representations for DVCL. Extensive experiments demonstrate the effectiveness of our method.

We present a novel large-scale dataset for defect detection in a logistics setting. Recent work on industrial anomaly detection has primarily focused on manufacturing scenarios with highly controlled poses and a limited number of object categories. Existing benchmarks like MVTec-AD (Bergmann et al., 2021) and VisA (Zou et al., 2022) have reached saturation, with state-of-the-art methods achieving up to 99.9% AUROC scores. In contrast to manufacturing, anomaly detection in retail logistics faces new challenges, particularly in the diversity and variability of object pose and appearance. Leading anomaly detection methods fall short when applied to this new setting. To bridge this gap, we introduce a new benchmark that overcomes the current limitations of existing datasets. With over 230,000 images (and more than 29,000 defective instances), it is 40 times larger than MVTec and contains more than 48,000 distinct objects. To validate the difficulty of the problem, we conduct an extensive evaluation of multiple state-of-the-art anomaly detection methods, demonstrating that they do not surpass 56.96% AUROC on our dataset. Further qualitative analysis confirms that existing methods struggle to leverage normal samples under heavy pose and appearance variation. With our large-scale dataset, we set a new benchmark and encourage future research towards solving this challenging problem in retail logistics anomaly detection. The dataset is available for download under https://www.kaputt-dataset.com.

Camera pose estimation is a fundamental computer vision task that is essential for applications like visual localization and multi-view stereo reconstruction. In the object-centric scenarios with sparse inputs, the accuracy of pose estimation can be significantly influenced by background textures that occupy major portions of the images across different viewpoints. In light of this, we introduce the Kaleidoscopic Background Attack (KBA), which uses identical segments to form discs with multi-fold radial symmetry. These discs maintain high similarity across different viewpoints, enabling effective attacks on pose estimation models even with natural texture segments. Additionally, a projected orientation consistency loss is proposed to optimize the kaleidoscopic segments, leading to a significant enhancement in the attack effectiveness. Experimental results show that optimized adversarial kaleidoscopic backgrounds can effectively attack various camera pose estimation models.

Background consistency remains a significant challenge in image editing tasks. Despite extensive developments, existing works still face a trade-off between maintaining similarity to the original image and generating content that aligns with the target. Here, we propose KV-Edit, a training-free approach that uses KV cache in DiTs to maintain background consistency, where background tokens are preserved rather than regenerated, eliminating the need for complex mechanisms or expensive training, ultimately generating new content that seamlessly integrates with the background within user-provided regions. We further explore the memory consumption of KV cache during editing and optimize the space complexity from O(N) to O(1) using an inversion-free method. Our approach is compatible with any DiT-based generative model without additional training. Experiments demonstrate that KV-Edit significantly outperforms existing approaches in terms of both background and image quality, even surpassing training-based methods.

Data-free black-box attacks aim to attack a model without access to either the model parameters or training data. Existing methods use a generator to synthesize training samples and then train a substitute model to imitate the victim model. The adversarial examples (AEs) are finally generated using the substitute model to transfer to the victim model. To this end, how to generate diverse training samples for substitute model training and improve the transferability of AEs from the substitute model to victim model become the core challenges. In this paper, we propose a Knowledge-Orthogonalized Ensemble Attack, dubbed KOEnsAttack, to accomplish these two goals. We first use dual networks as the ensemble substitute model, and then propose a sample hardness enhancement to transform the samples from the generator into hard samples that exist in the controversial regions of the dual models for promoting the sample diversity. Next, during the substitute model training, we design a knowledge orthogonalization module to guide the dual networks in learning complementary and useful information from the black-box, thereby enhancing the transferability of adversarial samples generated on the final ensemble model. Extensive experiments on several datasets are conducted to evaluate the effectiveness of our method. The results show that the proposed method can achieve superior performance compared with the state-of-the-art competitors.

We present JointDiT, a diffusion transformer that models the joint distribution of RGB and depth. By leveraging the architectural benefit and outstanding image prior of the state-of-the-art diffusion transformer, JointDiT not only generates high-fidelity images but also produces geometrically plausible and accurate depth maps. This solid joint distribution modeling is achieved through two simple yet effective techniques that we propose, i.e., adaptive scheduling weights, which depend on the noise levels of each modality, and the unbalanced timestep sampling strategy. With these techniques, we train our model across all noise levels for each modality, enabling JointDiT to naturally handle various combinatorial generation tasks, including joint generation, depth estimation, and depth-conditioned image generation by simply controlling the timesteps of each branch. JointDiT demonstrates outstanding joint generation performance. Furthermore, it achieves comparable results in depth estimation and depth-conditioned image generation, suggesting that joint distribution modeling can serve as a replaceable alternative to conditional generation.

Recent works propose extending 3DGS with semantic feature vectors for simultaneous semantic segmentation and image rendering. However, these methods often treat the semantic and rendering branches separately, relying solely on 2D supervision while ignoring the 3D Gaussian geometry. Moreover, current adaptive strategies adapt the Gaussian set depending solely on rendering gradients, which can be insufficient in subtle or textureless regions. In this work, we propose a joint enhancement framework for 3D semantic Gaussian modeling that synergizes both semantic and rendering branches. Firstly, unlike conventional point cloud shape encoding, we introduce an anisotropic 3D Gaussian Chebyshev descriptor using the Laplace-Beltrami operator to capture fine-grained 3D shape details, thereby distinguishing objects with similar appearances and reducing reliance on potentially noisy 2D guidance. In addition, without relying solely on rendering gradient, we adaptively adjust Gaussian allocation and spherical harmonics (SH) with local semantic and shape signals, enhancing rendering efficiency through selective resource allocation. Finally, we employ a cross-scene knowledge transfer module to continuously update learned shape patterns, enabling faster convergence and robust representations without relearning shape information from scratch for each new scene. Experiments on multiple datasets demonstrate improvements in segmentation accuracy and rendering quality while maintaining high rendering frame rates.

We introduce a novel approach for simultaneous self-supervised video alignment and action segmentation based on a unified optimal transport framework. In particular, we first tackle self-supervised video alignment by developing a fused Gromov-Wasserstein optimal transport formulation with a structural prior, which trains efficiently on GPUs and needs only a few iterations for solving the optimal transport problem. Our single-task method achieves the state-of-the-art performance on multiple video alignment benchmarks and outperforms VAVA, which relies on a traditional Kantorovich optimal transport formulation with an optimality prior. Furthermore, we extend our approach by proposing a unified optimal transport framework for joint self-supervised video alignment and action segmentation, which requires training and storing a single model and saves both time and memory consumption as compared to two different single-task models. Extensive evaluations on several video alignment and action segmentation datasets demonstrate that our multi-task method achieves comparable video alignment yet superior action segmentation results over previous methods in video alignment and action segmentation respectively. Finally, to the best of our knowledge, this is the first work to unify video alignment and action segmentation into a single model.

Latest diffusion models have shown promising results in category-level 6D object pose estimation by modeling the conditional pose distribution with depth image input. The existing methods, however, suffer from slow convergence during training, learning its encoder with the diffusion denoising network in end-to-end fashion, and require an additional network that evaluates sampled pose hypotheses to filter out low-quality pose candidates. In this paper, we propose a novel pipeline that tackles these limitations by two key components. First, the proposed method pre-trains the encoder with the direct pose regression head, and jointly learns the networks via the regression head and the denoising diffusion head, significantly accelerating training convergence while achieving higher accuracy. Second, sampling guidance via time-dependent score scaling is proposed s.t. the exploration-exploitation trade-off is effectively taken, eliminating the need for the additional evaluation network. The sampling guidance maintains multi-modal characteristics of symmetric objects at early denoising steps while ensuring high-quality pose generation at final steps. Extensive experiments on multiple benchmarks including REAL275, HouseCat6D, and ROPE, demonstrate that the proposed method, simple yet effective, achieves state-of-the-art accuracies even with single-pose inference, while being more efficient in both training and inference.

In this work, we introduce JDCL - a new method for continual learning with generative rehearsal based on joint diffusion models. Neural networks suffer from catastrophic forgetting defined as abrupt loss in the model's performance when retrained with additional data coming from a different distribution. Generative-replay-based continual learning methods try to mitigate this issue by retraining a model with a combination of new and rehearsal data sampled from a generative model. In this work, we propose to extend this idea by combining a continually trained classifier with a diffusion-based generative model into a single - jointly optimized neural network. We show that such shared parametrization, combined with the knowledge distillation technique allows for stable adaptation to new tasks without catastrophic forgetting. We evaluate our approach on several benchmarks, where it outperforms recent state-of-the-art generative replay techniques. Additionally, we extend our method to the semi-supervised continual learning setup, where it outperforms competing buffer-based replay techniques, and evaluate, in a self-supervised manner, the quality of trained representations.

Learning with noisy labels is a crucial task for training accurate deep neural networks. To mitigate label noise, prior studies have proposed various robust loss functions, particularly symmetric losses. Nevertheless, symmetric losses usually suffer from the underfitting issue due to the overly strict constraint. To address this problem, the Active Passive Loss (APL) jointly optimizes an active and a passive loss to mutually enhance the overall fitting ability. Within APL, symmetric losses have been successfully extended, yielding advanced robust loss functions. Despite these advancements, emerging theoretical analyses indicate that asymmetric losses, a new class of robust loss functions, possess superior properties compared to symmetric losses. However, existing asymmetric losses are not compatible with advanced optimization frameworks such as APL, limiting their potential and applicability. Motivated by this theoretical gap and the prospect of asymmetric losses, we extend the asymmetric loss to the more complex passive loss scenario and propose the Asymetric Mean Square Error (AMSE), a novel asymmetric loss. We rigorously establish the necessary and sufficient condition under which AMSE satisfies the asymmetric condition. By substituting the traditional symmetric passive loss in APL with our proposed AMSE, we introduce a novel robust loss framework termed Joint Asymmetric Loss (JAL). Extensive experiments demonstrate the effectiveness of our method in mitigating label noise. Code available at: https://github.com/cswjl/joint-asymmetric-loss

The automatic assembly problem has attracted increasing interest due to its complex challenges that involve 3D representation. This paper introduces Jigsaw++, a novel generative method designed to tackle the multifaceted challenges of reconstructing complete shape for the reassembly problem. Existing approach focusing primarily on piecewise information for both part and fracture assembly, often overlooking the integration of complete object prior. Jigsaw++ distinguishes itself by learning a shape prior of complete objects. It employs the proposed "retargeting" strategy that effectively leverages the output of any existing assembly method to generate complete shape reconstructions. This capability allows it to function orthogonally to the current methods. Through extensive evaluations on Breaking Bad dataset and PartNet, Jigsaw++ has demonstrated its effectiveness, reducing reconstruction errors and enhancing the precision of shape reconstruction, which sets a new direction for future reassembly model developments.

Multimodal Large Language Models (MLLMs) have achieved impressive performance and have been put into practical use in commercial applications, but they still have potential safety mechanism vulnerabilities. Jailbreak attacks are red teaming methods that aim to bypass safety mechanisms and discover MLLMs' potential risks. Existing MLLMs' jailbreak methods often bypass the model's safety mechanism through complex optimization methods or carefully designed image and text prompts. Despite achieving some progress, they have a low attack success rate on commercial closed-source MLLMs. Unlike previous research, we empirically find that there exists a Shuffle Inconsistency between MLLMs' comprehension ability and safety ability for the shuffled harmful instruction. That is, from the perspective of comprehension ability, MLLMs can understand the shuffled harmful text-image instructions well. However, they can be easily bypassed by the shuffled harmful instructions from the perspective of safety ability, leading to harmful responses. Then we innovatively propose a text-image jailbreak attack named SI-Attack. Specifically, to fully utilize the Shuffle Inconsistency and overcome the shuffle randomness, we apply a query-based black-box optimization method to select the most harmful shuffled inputs based on the feedback of the toxic judge model. A series of experiments show that SI-Attack can effectively improve the attack's performance on three benchmarks for both open-source and closed-source MLLMs.

Diffusion models are widely used in real-world applications, but ensuring their safety remains a major challenge. Despite many efforts to enhance the security of diffusion models, jailbreak and adversarial attacks can still bypass these defenses, generating harmful content. However, the lack of standardized evaluation makes it difficult to assess the robustness of diffusion model system. To address this, we introduce JailbreakDiffBench, a comprehensive benchmark for systematically evaluating the safety of diffusion models against various attacks and under different defenses. Our benchmark includes a high-quality, human-annotated prompt and image dataset covering diverse attack scenarios. It consists of two key components: (1) an evaluation protocol to measure the effectiveness of moderation mechanisms and (2) an attack assessment module to benchmark adversarial jailbreak strategies. Through extensive experiments, we analyze existing filters and reveal critical weaknesses in current safety measures. JailbreakDiffBench is designed to support both text-to-image and text-to-video models, ensuring extensibility and reproducibility.

Despite significant advances in learning-based lossy compression algorithms, standardizing codecs remains a critical challenge. In this paper, we present the JPEG Processing Neural Operator (JPNeO), a next-generation JPEG algorithm that maintains full backward compatibility with the current JPEG format. Our JPNeO improves chroma component preservation and enhances reconstruction fidelity compared to existing artifact removal methods by incorporating neural operators in both the encoding and decoding stages. JPNeO achieves practical benefits in terms of reduced memory usage and parameter count. We further validate our hypothesis about the existence of a space with high mutual information through empirical evidence. In summary, the JPNeO functions as a high-performance out-of-the-box image compression pipeline without changing source coding's protocol. Our source code is available at https://github.com/WooKyoungHan/JPNeO.