ICCV 2025 Papers

The early stage of multi-modal pre-training plays a pivotal role in aligning two modalities for Large Vision-Language Models (LVLMs), while evaluating its training quality usually requires the costly supervised fine-tuning (SFT) stage to verify the downstream benchmark scores. Loss, perplexity, and in-context evaluation results are commonly used pre-training metrics for Large Language Models (LLMs), while we observed that these metrics are less indicative when quantifying the pre-trained LVLMs. Due to the lack of proper metrics, the research of LVLMs in the multi-modal fusion stage is hindered greatly, including the training data choice, efficient module design, etc.In this paper, we first present Modality Integration Rate (MIR), an effective, robust, and generalized metric to indicate the multi-modal alignment quality of LVLMs without SFT. This metric evaluates LVLM pre-training from the inter-modal distribution distance perspective, which is 1) Effective to represent the fusion quality and show a positive relation with the benchmark performance after SFT, 2) Robust toward different training/evaluation data, and 3) Generalize across training configurations and architecture choices. Complementing MIR, we further propose learnable Modality Calibration (MoCa), a lightweight module to narrow the modality gap at each language model layer during training. A series of experiments are conducted to explore the effectiveness of MIR and MoCa, demonstrating that MIR is highly indicative about training data selection, training strategy schedule, and model architecture design to get better pre-training results. The code is avaliable at \href https://github.com/shikiw/Modality-Integration-Rate shikiw/Modality-Integration-Rate

Existing distillation-based and reconstruction-based methods have a critical limitation: Autoencoder-based frameworks trained exclusively on normal samples unexpectedly well reconstruct abnormal features, leadingto degraded detection performance. We identify this phenomenon as 'anomaly leakage' (AL): the decoder optimized by reconstruction loss tends to directly copy the encoded input, regardless of whether the input is a normal or abnormal feature. To address this issue, we propose a novel framework that explicitly decouples encoded features into normal and abnormal components through a special invertible mapping in a prior latent space. Next, we remove abnormal components and leverage normal remainders for feature reconstruction. Compared to previous methods, the invertible structure can eliminate anomalous information point-to-point without damaging the information of neighboring patches, improving reconstruction. In this process, effective synthetic abnormal features are essential for training the decoupling process. Therefore, we propose applying adversarial training to find suitable perturbations to simulate feature-level anomalies. Extensive experimental evaluations on benchmark datasets, including MVTec AD, VisA, and Real-IAD, demonstrate that our method achieves competitive performance compared to state-of-the-art approaches. Code is available at: DecAD.

The leak of anomalous information from input condition poses a great challenge to reconstruction-based anomaly detection. Recent diffusion-based methods respond to this issue by suppressing anomaly information for condition injection or in-sampling inversion. However, since they treat conditions as a time-invariant prior, they fall into a trade-off problem between anomaly suppression and normal pattern consistency. To address this probelm, we propose Debiasing Trace Guidance (DTG) framework based on Flow Matching towards debiasing generation for more accurate unsupervised multi-class anomaly detection. Generally, DTG distills a low-dimensional generation sub-trace robust to anomalies by Top-down Trace Distillation, and then utilizes its time-varying velocity features to guide a debiasing generation by Bottom-up Velocity Alignment. The trace distillation filters out high-frequency anomalies via learnable wavelet filters and reserving structural information by keeping global consistency across samples using Skinhorn Distance. Subsequently, the velocity field of original trace is aligned with the one of sub-trace through KV-Injection Attention mechanism. The model is forced to generate normal details from corresponding low-dimensional contexts via Alignment Mask. Experimental results on several benchmarks have demonstrated the effectiveness of the proposed method.

Chest X-ray classification is extensively utilized within the field of medical image analysis. However, manually labeling chest X-ray images is time-consuming and costly. Domain adaptation, which is designed to transfer knowledge from related domains, could offer a promising solution. Existing methods employ feature adaptation or self-training for knowledge transfer. Nonetheless, negative transfer is observed due to the entanglement of class imbalance and distribution shift in chest X-ray classification. In this paper, we propose Debiased Curriculum Adaptation framework to mitigate negative transfer in two aspects: (1) Curriculum Adaptation, which is designed to transfer knowledge in an easy-to-hard way, is proposed to alleviate confirmation bias in self-training. (2) Spectral Debiasing is introduced to harmonize the feature space between the source and target domains, as well as balance the feature space of positive and negative samples. Extensive experiments on 72 transfer tasks (including 6 diseases and 4 domains) demonstrate our superiority over state-of-the-art methods. In comparison to advanced methods, our approach effectively mitigates negative transfer, ensuring safe knowledge transfer.

Despite LiDAR (Light Detection and Ranging) being an effective privacy-preserving alternative to RGB cameras to perceive human activities, it remains largely underexplored in the context of multi-modal contrastive pre-training for human activity understanding tasks, such as human activity recognition (HAR), retrieval, or person re-identification (RE-ID). To close this gap, our work explores learning the correspondence between LiDAR point clouds, human skeleton poses, IMU data, and text in a joint embedding space. More specifically, we present DeSPITE, a \underline De ep \underline S keleton-\underline P ointcloud-\underline I MU-\underline T ext \underline E mbedding model, which effectively learns a joint embedding space across these four modalities. At the heart of our empirical exploration, we have combined the existing LIPD and Babel datasets, which enabled us to synchronize data of all four modalities, allowing us to explore the learning of a new joint embedding space. Our experiments demonstrate novel human activity understanding tasks for point cloud sequences enabled through DeSPITE, including Skeleton\leftrightarrowPointcloud\leftrightarrowIMU matching, retrieval, and temporal moment retrieval. Furthermore, we show that DeSPITE is an effective pre-training strategy for point cloud HAR through experiments in MSR-Action3D and HMPEAR. Code and models are publicly available at https://github.com/thkreutz/despite.

Referring Image Segmentation (RIS) is a challenging task that aims to segment objects in an image based on natural language expressions. While prior studies have predominantly concentrated on improving vision-language interactions and achieving fine-grained localization, a systematic analysis of the fundamental bottlenecks in existing RIS frameworks remains underexplored. To bridge this gap, we propose DeRIS , a novel framework that decomposes RIS into two key components: perception and cognition . This modular decomposition facilitates a systematic analysis of the primary bottlenecks impeding RIS performance. Our findings reveal that the predominant limitation lies not in perceptual deficiencies, but in the insufficient multi-modal cognitive capacity of current models. To mitigate this, we propose a Loopback Synergy mechanism, which enhances the synergy between the perception and cognition modules, thereby enabling precise segmentation while simultaneously improving robust image-text comprehension. Additionally, we analyze and introduce a simple non-referent sample conversion data augmentation to address the long-tail distribution issue related to target existence judgement in general scenarios. Notably, DeRIS demonstrates inherent adaptability to both non- and multi-referents scenarios without requiring specialized architectural modifications, enhancing its general applicability. The codes and models are available at https://github.com/Dmmm1997/DeRIS

Reconstructing clean, distractor-free 3D scenes from real-world captures remains a significant challenge, particularly in highly dynamic and cluttered settings such as egocentric videos. To tackle this problem, we introduce DeGauss, a simple and robust self-supervised framework for dynamic scene reconstruction based on a decoupled dynamic-static Gaussian Splatting design. DeGauss models dynamic elements with foreground Gaussians and static content with background Gaussians, using a probabilistic mask to coordinate their composition and enable independent yet complementary optimization. DeGauss generalizes robustly across a wide range of real-world scenarios, from casual image collections to long, dynamic egocentric videos, without relying on complex heuristics or extensive supervision. Experiments on benchmarks including NeRF-on-the-go, ADT, AEA, Hot3D, and EPIC-Fields demonstrate that DeGauss consistently outperforms existing methods, establishing a strong baseline for generalizable, distractor-free 3D reconstruction in highly dynamic, interaction-rich environments.

Deep learning algorithms are highly data-intensive, particularly for tasks requiring pixel-level annotations, such as semantic segmentation, which makes achieving pixel-level image understanding costly. Few-shot segmentation seeks to address this challenge by enabling models to segment novel objects using only a limited number of labeled support images as references. In this paper, we argue that the traditional image-to-mask decoding framework places excessive reliance on the quality of the support sample, which is prone to errors when encountering class bias. Thus, we propose a novel image-to-mask denoising learning paradigm for few-shot segmentation, transforming mask decoding into a denoising process to reduce the support reliance problem with the help of denoising diffusion models. We formulate our image-to-mask denoising learning process in two stages: an image corruption stage and a mask denoising stage. In the first stage, we introduce an adaptive image corruption method that perturbs the image based on regional semantics, motivated by the insight of perturbing data to populate low data density regions. In the second stage, we employ an in-model denoising paradigm, designing a network to facilitate support-to-query semantic propagation and mask denoising in a single forward pass. To enhance categorical discrimination for the denoising network, we incorporate discriminative attribute learning, which leverages base classes to train the model in distinguishing object categories and generalizing to novel classes. Extensive experiments and ablation studies validate the effectiveness of our approach, demonstrating that the proposed method achieves competitive performance across various benchmarks.

High-quality open-source datasets have emerged as a pivotal catalyst driving the swift advancement of deep learning, while facing the looming threat of potential exploitation. Protecting these datasets is of paramount importance for the interests of their owners. The verification of dataset ownership has evolved into a crucial approach in this domain; however, existing verification techniques are predominantly tailored to supervised models and contrastive pre-trained models, rendering them ill-suited for direct application to the increasingly prevalent masked models. In this work, we introduce the inaugural methodology addressing this critical, yet unresolved challenge, termed Dataset Ownership Verification for Masked Modeling (DOV4MM). The central objective is to ascertain whether a suspicious black-box model has been pre-trained on a particular unlabeled dataset, thereby assisting dataset owners in safeguarding their rights. DOV4MM is grounded in our empirical observation that when a model is pre-trained on the target dataset, the difficulty of reconstructing masked information within the embedding space exhibits a marked contrast to models not pre-trained on that dataset. We validated the efficacy of DOV4MM through ten masked image models on ImageNet-1K and four masked language models on WikiText-103. The results demonstrate that DOV4MM rejects the null hypothesis, with a p-value considerably below 0.05, surpassing all prior approaches. Code is available at https://github.com/xieyc99/DOV4MM.

Dataset Distillation (DD) aims to generate a compact synthetic dataset that enables models to achieve performance comparable to training on the full large dataset, significantly reducing computational costs. Drawing from optimal transport theory, we introduce WMDD (Wasserstein Metric-based Dataset Distillation), a straightforward yet powerful method that employs the Wasserstein metric to enhance distribution matching. We compute the Wasserstein barycenter of features from a pretrained classifier to capture essential characteristics of the original data distribution. By optimizing synthetic data to align with this barycenter in feature space and leveraging per-class BatchNorm statistics to preserve intra-class variations, WMDD maintains the efficiency of distribution matching approaches while achieving state-of-the-art results across various high-resolution datasets. Our extensive experiments demonstrate WMDD's effectiveness and adaptability, highlighting its potential for advancing machine learning applications at scale. Code is available at https://github.com/Liu-Hy/WMDD and website at https://liu-hy.github.io/WMDD/.

Dataset distillation (DD) condenses large datasets into compact yet informative substitutes, preserving performance comparable to the original dataset while reducing storage, transmission costs, and computational consumption. However, previous DD methods mainly focus on distilling information from images, often overlooking the semantic information inherent in the data. The disregard for context hinders the model's generalization ability, particularly in tasks involving complex datasets, which may result in illogical outputs or the omission of critical objects. In this study, we integrate vision-language methods into DD by introducing text prototypes to distill language information and collaboratively synthesize data with image prototypes, thereby enhancing dataset distillation performance. Notably, the text prototypes utilized in this study are derived from descriptive text information generated by an open-source vision-language model. This framework demonstrates broad applicability across datasets without pre-existing text descriptions, expanding the potential of dataset distillation beyond traditional image-based approaches. Compared to other methods, the proposed approach generates logically coherent images containing target objects, achieving state-of-the-art validation performance and demonstrating robust generalization. Source code and generated data are available in https://github.com/zou-yawen/Dataset-Distillation-via-Vision-Language-Category-Prototype/.

Driven by the "scale-is-everything" paradigm, modern machine learning increasingly demands ever-larger datasets and models, yielding prohibitive computational and storage requirements. Dataset distillation mitigates this by compressing an original dataset into a small set of synthetic samples, while preserving its full utility. Yet, existing methods either maximize performance under fixed storage budgets or pursue suitable synthetic data representations for redundancy removal, without jointly optimizing both objectives. In this work, we propose a joint rate-utility optimization method for dataset distillation. We parameterize synthetic samples as optimizable latent codes decoded by extremely lightweight networks. We estimate the Shannon entropy of quantized latents as the rate measure and plug any existing distillation loss as the utility measure, trading them off via a Lagrange multiplier. To enable fair, cross-method comparisons, we introduce bits per class (bpc), a precise storage metric that accounts for sample, label, and decoder parameter costs. On CIFAR-10, CIFAR-100, and ImageNet-128, our method achieves up to 170xgreater compression than standard distillation at comparable accuracy. Across diverse bpc budgets, distillation losses, and backbone architectures, our approach consistently establishes better rate-utility trade-offs.

Low-light Object detection is crucial for many real-world applications but remains challenging due to degraded image quality. While recent studies have shown that RAW images offer superior potential over RGB images, existing approaches either use RAW-RGB images with information loss or employ complex frameworks. To address these, we propose a lightweight and self-adaptive Image Signal Processing (ISP) plugin, Dark-ISP, which directly processes Bayer RAW images in dark environments, enabling seamless end-to-end training for object detection. Our key innovations are: (1) We deconstruct conventional ISP pipelines into sequential linear (sensor calibration) and nonlinear (tone mapping) sub-modules, recasting them as differentiable components optimized through task-driven losses. Each module is equpped with content-aware adaptability and physics-informed priors, enabling automatic RAW-to-RGB conversion aligned with detection objectives. (2) By exploiting the ISP pipeline's intrinsic cascade structure, we devise a self-boosting strategy that facilitates cooperation between sub-modules. Through extensive experiments on three RAW image datasets, we demonstrate that our method outperforms state-of-the-art RGB- and RAW-based detection approaches, achieving superior results with minimal parameters in challenging low-light environments.

Generating coherent and diverse human dances from music signals has gained tremendous progress in animating virtual avatars. While existing methods support direct dance synthesis, they fail to recognize that enabling users to edit dance movements is far more practical in real-world choreography scenarios. Moreover, the lack of high-quality dance datasets incorporating iterative editing also limits addressing this challenge. To achieve this goal, we first construct DanceRemix, a large-scale multi-turn editable dance dataset comprising the prompt featuring over 25.3M dance frames and 84.5K pairs. In addition, we propose a novel framework for iterative and editable dance generation coherently aligned with given music signals, namely DanceEditor. Considering the dance motion should be both musical rhythmic and enable iterative editing by user descriptions, our framework is built upon a prediction-then-editing paradigm unifying multi-modal conditions. At the initial prediction stage, our framework improves the authority of generated results by directly modeling dance movements from tailored, aligned music. Moreover, at the subsequent iterative editing stages, we incorporate text descriptions as conditioning information to draw the editable results through a specifically designed Cross-modality Edition Module (CEM). Specifically, CEM adaptively integrates the initial prediction with music and text prompts as temporal motion cues to guide the synthesized sequences.Thereby, the results display music harmony while preserving fine-grained semantic alignment with text descriptions. Extensive experiments demonstrate that our method outperforms the state-of-the-art models on our newly collected DanceRemix dataset. Code is available at https://lzvsdy.github.io/DanceEditor/.

Visual reasoning (VR), which is crucial in many fields for enabling human-like visual understanding, remains highly challenging. Recently, compositional visual reasoning approaches, which leverage the reasoning abilities of large language models (LLMs) with integrated tools to solve problems, have shown promise as more effective strategies than end-to-end VR methods. However, these approaches face limitations, as frozen LLMs lack tool awareness in VR, leading to performance bottlenecks. While leveraging LLMs for reasoning is widely used in other domains, they are not directly applicable to VR due to limited training data, imperfect tools that introduce errors and reduce data collection efficiency in VR, and challenging in fine-tuning on noisy workflows. To address these challenges, we propose DWIM: i) Discrepancy-aware training Workflow generation, which assesses tool usage and extracts more viable workflows for training; and ii) Instruct-Masking fine-tuning, which guides the model to only clone effective actions, enabling the generation of more practical solutions. Our experiments demonstrate that DWIM achieves state-of-the-art performance across various VR tasks, exhibiting strong generalization on multiple widely-used datasets.

Most 3D object generators prioritize aesthetic quality, often neglecting the physical constraints necessary for practical applications. One such constraint is that a 3D object should be self-supporting, i.e., remain balanced under gravity. Previous approaches to generating stable 3D objects relied on differentiable physics simulators to optimize geometry at test time, which is slow, unstable, and prone to local optima. Inspired by the literature on aligning generative models with external feedback, we propose Direct Simulation Optimization (DSO). This framework leverages feedback from a (non-differentiable) simulator to increase the likelihood that the 3D generator directly outputs stable 3D objects. We construct a dataset of 3D objects labeled with stability scores obtained from the physics simulator. This dataset enables fine-tuning of the 3D generator using the stability score as an alignment metric, via direct preference optimization (DPO) or direct reward optimization (DRO)---a novel objective we introduce to align diffusion models without requiring pairwise preferences. Our experiments demonstrate that the fine-tuned feed-forward generator, using either the DPO or DRO objective, is significantly faster and more likely to produce stable objects than test-time optimization. Notably, the DSO framework functions even without any ground-truth 3D objects for training, allowing the 3D generator to self-improve by automatically collecting simulation feedback on its own outputs.

We explore a quaternion adjugate matrix-based representation for rotational motion in the Perspective-n-Point (PnP) problem. Leveraging quadratic quaternion terms within a Determinant Ratio Matrix (DRaM) estimation framework, we extend its application to perspective scenarios, providing a robust and efficient initialization for iterative PnP pose estimation. Notably, by solving the orthographic projection least-squares problem, DRaM provides a reliable initialization that enhances the accuracy and stability of iterative PnP solvers. Experiments on synthetic and real data demonstrate its efficiency, accuracy, and robustness, particularly under high noise conditions. Furthermore, our non-minimal formulation ensures numerical stability, making it effective for real-world applications.

We present DPoser-X, a diffusion-based prior model for 3D whole-body human poses. Building a versatile and robust full-body human pose prior remains challenging due to the inherent complexity of articulated human poses and the scarcity of high-quality whole-body pose datasets. To address these limitations, we introduce a Diffusion model as body Pose prior (DPoser) and extend it to DPoser-X for expressive whole-body human pose modeling.Our approach unifies various pose-centric tasks as inverse problems, solving them through variational diffusion sampling. To enhance performance on downstream applications, we introduce a novel truncated timestep scheduling method specifically designed for pose data characteristics. We also propose a masked training mechanism that effectively combines whole-body and part-specific datasets, enabling our model to capture interdependencies between body parts while avoiding overfitting to specific actions.Extensive experiments demonstrate DPoser-X's robustness and versatility across multiple benchmarks for body, hand, face, and full-body pose modeling. Our model consistently outperforms state-of-the-art alternatives, establishing a new benchmark for whole-body human pose prior modeling.