ICCV 2025 Papers

Current generative models struggle to synthesize dynamic 4D driving scenes that simultaneously support temporal extrapolation and spatial novel view synthesis (NVS) without per-scene optimization. A key challenge lies in finding an efficient and generalizable geometric representation that seamlessly connects temporal and spatial synthesis. To address this, we propose DiST-4D, the first disentangled spatiotemporal diffusion framework for 4D driving scene generation, which leverages metric depth as the core geometric representation. DiST-4D decomposes the problem into two diffusion processes: DiST-T, which predicts future metric depth and multi-view RGB sequences directly from past observations, and DiST-S, which enables spatial NVS by training only on existing viewpoints while enforcing cycle consistency. This cycle consistency mechanism introduces a forward-backward rendering constraint, reducing the generalization gap between observed and unseen viewpoints. Metric depth is essential for both accurate reliable forecasting and accurate spatial NVS, as it provides a view-consistent geometric representation that generalizes well to unseen perspectives. Experiments demonstrate that DiST-4D achieves state-of-the-art performance in both temporal prediction and NVS tasks, while also delivering competitive performance in planning-related evaluations. The Project is available at https://royalmelon0505.github.io/DiST-4D

3D semantic segmentation provides high-level scene understanding for applications in robotics, autonomous systems, etc. Traditional methods adapt exclusively to either task-specific goals (open-vocabulary segmentation) or scene content (unsupervised semantic segmentation). We propose DiSCO-3D, the first method addressing the broader problem of 3D Open-Vocabulary Sub-concepts Discovery, which aims to provide a 3D semantic segmentation that adapts to both the scene and user queries. We build DiSCO-3D on Neural Fields representations, combining unsupervised segmentation with weak open-vocabulary guidance. Our evaluations demonstrate that DiSCO-3D achieves effective performance in Open-Vocabulary Sub-concepts Discovery and exhibits state-of-the-art results in the edge cases of both open-vocabulary and unsupervised segmentation.

We introduce DiMPLe (Disentangled Multi-Modal Prompt Learning), a novel approach to disentangle invariant and spurious features across vision and language modalities in multi-modal learning. Spurious correlations in visual data often hinder out-of-distribution (OOD) performance. Unlike prior methods focusing solely on image features, DiMPLe disentangles features within and across modalities while maintaining consistent alignment, enabling better generalization to novel classes and robustness to distribution shifts.Our method combines three key objectives: (1) mutual information minimization between invariant and spurious features, (2) spurious feature regularization, and (3) contrastive learning on invariant features. Extensive experiments demonstrate DiMPLe demonstrates superior performance compared to CoOp-OOD, when averaged across 11 diverse datasets, and achieves absolute gains of 15.27 in base class accuracy and 44.31 in novel class accuracy. The code will be released publicly upon acceptance.

Developing a unified pipeline that enables users to remove, re-texture, or replace objects in a versatile manner is crucial for text-guided 3D inpainting. However, there are still challenges in performing multiple 3D inpainting tasks within a unified framework: 1) Single reference inpainting methods lack robustness when dealing with views that are far from the reference view; 2) Appearance inconsistency arises when independently inpainting multi-view images with 2D diffusion priors; 3) Geometry inconsistency limits performance when there are significant geometric changes in the inpainting regions. To tackle these challenges, we introduce DiGA3D, a novel and versatile 3D inpainting pipeline that leverages diffusion models to propagate consistent appearance and geometry in a coarse-to-fine manner. First, DiGA3D develops a robust strategy for selecting multiple reference views to reduce errors during propagation. Next, DiGA3D designs an Attention Feature Propagation (AFP) mechanism that propagates attention features from the selected reference views to other views via diffusion models to maintain appearance consistency. Furthermore, DiGA3D introduces a Texture-Geometry Score Distillation Sampling (TG-SDS) loss to further improve the geometric consistency of inpainted 3D scenes. Extensive experiments on multiple 3D inpainting tasks demonstrate the effectiveness of our method. The project page is available at https://rorisis.github.io/DiGA3D/.

As large models gain traction, vision-language models are enabling robots to tackle increasingly complex tasks. However, limited by the difficulty of data collection, progress has mainly focused on controlling simple gripper end-effectors. There is little research on functional grasping with large models for human-like dexterous hands. In this paper, we introduce DexVLG, a large Vision-Language-Grasp model for Dexterous grasp pose prediction aligned with language instructions using single-view RGBD input. To accomplish this, we generate a dataset of 170 million dexterous grasp poses mapped to semantic parts across 174,000 objects in simulation, paired with detailed part-level captions. This large-scale dataset, named DexGraspNet 3.0, is used to train a VLM with a flow-matching-based pose head producing instruction-aligned grasp poses for tabletop objects. To evaluate DexVLG's performance, we create benchmarks in simulations and conduct real-world experiments. Extensive experiments demonstrate DexVLG's strong zero-shot generalization capabilities, achieving an over 76% zero-shot execution success rate and state-of-the-art part-grasp accuracy in simulation, as well as successful part-aligned grasps on physical objects in real-world scenarios.

Handover between a human and a dexterous robotic hand is a fundamental yet challenging task in human-robot collaboration. It requires handling dynamic environments and a wide variety of objects and demands robust and adaptive grasping strategies. However, progress in developing effective dynamic dexterous grasping methods is limited by the absence of high-quality, real-world human-to-robot handover datasets. Existing datasets primarily focus on grasping static objects or rely on synthesized handover motions, which differ significantly from real-world robot motion patterns, creating a substantial gap in applicability. In this paper, we introduce DexH2R, a comprehensive real-world dataset for human-to-robot handovers, built on a dexterous robotic hand. Our dataset captures a diverse range of interactive objects, dynamic motion patterns, rich visual sensor data, and detailed annotations. Additionally, to ensure natural and human-like dexterous motions, we utilize teleoperation for data collection, enabling the robot's movements to align with human behaviors and habits, which is a crucial characteristic for intelligent humanoid robots. Furthermore, we propose an effective solution, DynamicGrasp, for human-to-robot handover and evaluate various state-of-the-art approaches, including auto-regressive models and diffusion policy methods, providing a thorough comparison and analysis. We believe our benchmark will drive advancements in human-to-robot handover research by offering a high-quality dataset, effective solutions, and comprehensive evaluation metrics. Project is at https://dexh2r.github.io.

Transformer-based approaches have gained significant attention in image restoration, where the core component, i.e, Multi-Head Attention (MHA), plays a crucial role in capturing diverse features and recovering high-quality results. In MHA, heads perform attention calculation independently from uniform split subspaces, and a redundancy issue is triggered to hinder the model from achieving satisfactory outputs. In this paper, we propose to improve MHA by exploring diverse learners and introducing various interactions between heads, which results in a Hierarchical multI-head atteNtion driven Transformer model, termed HINT, for image restoration. HINT contains two modules, i.e., the Hierarchical Multi-Head Attention (HMHA) and the Query-Key Cache Updating (QKCU) module, to address the redundancy problem that is rooted in vanilla MHA. Specifically, HMHA extracts diverse contextual features by employing heads to learn from subspaces of varying sizes and containing different information. Moreover, QKCU, comprising intra- and inter-layer schemes, further reduces the redundancy problem by facilitating enhanced interactions between attention heads within and across layers. Extensive experiments are conducted on 12 benchmarks across 5 image restoration tasks, including low-light enhancement, dehazing, desnowing, denoising, and deraining, to demonstrate the superiority of HINT. The source code is available in the supplementary materials.

6D pose confidence region estimation has emerged as a critical direction, aiming to perform uncertainty quantification for assessing the reliability of estimated poses. However, current sampling-based approach suffers from critical limitations that severely impede their practical deployment: 1) the sampling speed significantly decreases as the number of samples increases. 2) the derived confidence regions are often excessively large. To address these challenges, we propose a deterministic and efficient method for estimating pose confidence regions. Our approach uses inductive conformal prediction to calibrate the deterministically regressed Gaussian keypoint distributions into 2D keypoint confidence regions. We then leverage the implicit function theorem to propagate these keypoint confidence regions directly into 6D pose confidence regions. This method avoids the inefficiency and inflated region sizes associated with sampling and ensembling. It provides compact confidence regions that cover the ground-truth poses with a user-defined confidence level. Experimental results on the LineMOD Occlusion and SPEED datasets show that our method achieves higher pose estimation accuracy with reduced computational time. For the same coverage rate, our method yields significantly smaller confidence region volumes, reducing them by up to 99.9% for rotations and 99.8% for translations. The code will be available soon.

Human pose estimation methods work well on isolated people but struggle with multiple-bodies-in-proximity scenarios. Previous work has addressed this problem by conditioning pose estimation by detected bounding boxes or keypoints, but overlooked instance masks. We propose to iteratively enforce mutual consistency of bounding boxes, instance masks, and poses. The introduced BBox-Mask-Pose (BMP) method uses three specialized models that improve each other's output in a closed loop. All models are adapted for mutual conditioning, which improves robustness in multi-body scenes. MaskPose, a new mask-conditioned pose estimation model, is the best among top-down approaches on OCHuman. BBox-Mask-Pose pushes SOTA on OCHuman dataset in all three tasks - detection, instance segmentation, and pose estimation. It also achieves SOTA performance on COCO pose estimation. The method is especially good in scenes with large instances overlap, where it improves detection by 39% over the baseline detector. With small specialized models and faster runtime, BMP is an effective alternative to large human-centered foundational models. Code and weights are available on the project website: MiraPurkrabek.github.io/BBox-Mask-Pose/

Despite the success of deep learning in close-set 3D object detection, existing approaches struggle with zero-shot generalization to novel objects and camera configurations. We introduce DetAny3D, a promptable 3D detection foundation model capable of detecting any novel object under arbitrary camera configurations using only monocular inputs. Training a foundation model for 3D detection is fundamentally constrained by the limited availability of annotated 3D data, which motivates DetAny3D to leverage the rich prior knowledge embedded in extensively pre-trained 2D foundation models to compensate for this scarcity. To effectively transfer 2D knowledge to 3D, DetAny3D incorporates two core modules: the 2D Aggregator, which aligns features from different 2D foundation models, and the 3D Interpreter with Zero-Embedding Mapping, which stabilizes early training in 2D-to-3D knowledge transfer. Experimental results validate the strong generalization of our DetAny3D, which not only achieves state-of-the-art performance on unseen categories and novel camera configurations, but also surpasses most competitors on in-domain data. DetAny3D sheds light on the potential of the 3D foundation model for diverse applications in real-world scenarios, e.g., rare object detection in autonomous driving, and demonstrates promise for further exploration of 3D-centric tasks in open-world settings. More visualization results can be found at our code repository.

Unlike closed-vocabulary 3D instance segmentation that is often trained end-to-end, open-vocabulary 3D instance segmentation (OV-3DIS) often leverages vision-language models (VLMs) to generate 3D instance proposals and classify them. While various concepts have been proposed from existing research, we observe that these individual concepts are not mutually exclusive but complementary. In this paper, we propose a new state-of-the-art solution for OV-3DIS by carefully designing a recipe to combine the concepts together and refining them to address key challenges. Our solution follows the two-stage scheme: 3D proposal generation and instance classification. We employ robust 3D tracking-based proposal aggregation to generate 3D proposals and remove overlapped or partial proposals by iterative merging/removal. For the classification stage, we replace the standard CLIP model with Alpha-CLIP, which incorporates object masks as an alpha channel to reduce background noise and obtain object-centric representation. Additionally, we introduce the standardized maximum similarity (SMS) score to normalize text-to-proposal similarity, effectively filtering out false positives and boosting precision. Our framework achieves state-of-the-art performance on ScanNet200 and S3DIS across all AP and AR metrics, even surpassing an end-to-end closed-vocabulary method.

Despite the progress in text-to-image generation, semantic image editing remains a challenge. Inversion-based algorithms unavoidably introduce reconstruction errors, while instruction-based models mainly suffer from limited dataset quality and scale. To address these problems, we propose a descriptive-prompt-based editing framework, named DescriptiveEdit. The core idea is to re-frame `instruction-based image editing' as `reference-image-based text-to-image generation', which preserves the generative power of well-trained Text-to-Image models without architectural modifications or inversion. Specifically, taking the reference image and a prompt as input, we introduce a Cross-Attentive UNet, which newly adds attention bridges to inject reference image features into the prompt-to-edit-image generation process. Owing to its text-to-image nature, DescriptiveEdit overcomes limitations in instruction dataset quality, integrates seamlessly with ControlNet, IP-Adapter, and other extensions, and is more scalable. Experiments on the Emu Edit benchmark show it improves editing accuracy and consistency.

Open-set 3D object retrieval (3DOR) is an emerging task aiming to retrieve 3D objects of unseen categories beyond the training set. Existing methods typically utilize all modalities (i.e., voxels, point clouds, multi-view images) and train specific backbones before fusion. However, they still struggle to produce generalized representations due to insufficient 3D training data. Being contrastively pre-trained on web-scale image-text pairs, CLIP inherently produces generalized representations for a wide range of downstream tasks. Building upon it, we present a simple yet effective framework named Describe, Adapt and Combine (DAC) by taking only multi-view images for open-set 3DOR. DAC innovatively synergizes a CLIP model with a multi-modal large language model (MLLM) to learn generalized 3D representations, where the MLLM is used for dual purposes. First, it describes the seen category information to align with CLIP's training objective for adaptation during training. Second, it provides external hints about unknown objects complementary to visual cues during inference. To improve the synergy, we introduce an Additive-Bias Low-Rank adaptation (AB-LoRA), which alleviates overfitting and further enhances the generalization to unseen categories. With only multi-view images, DAC significantly surpasses prior arts by an average of +10.01% mAP on four open-set 3DOR datasets. Moreover, its generalization is also validated on image-based and cross-dataset setups. Code is available at https://github.com/wangzhichuan123/DAC.

Generating detailed and accurate descriptions for specific regions in images and videos remains a fundamental challenge for vision-language models. We introduce the Describe Anything Model (DAM), a model designed for detailed localized captioning (DLC). DAM preserves both local details and global context through two key innovations: a focal prompt, which ensures high-resolution encoding of targeted regions, and a localized vision backbone, which integrates precise localization with its broader context. To tackle the scarcity of high-quality DLC data, we propose a Semi-supervised learning (SSL)-based Data Pipeline (DLC-SDP). DLC-SDP starts with existing segmentation datasets and expands to unlabeled web images using SSL. We introduce DLC-Bench, a benchmark designed to evaluate DLC without relying on reference captions. DAM sets new state-of-the-art on 7 benchmarks spanning keyword-level, phrase-level, and detailed multi-sentence localized image and video captioning.

The emergence of vision-language models has transformed medical AI, enabling unprecedented advances in diagnostic capability and clinical applications. However, progress in dermatology has lagged behind other medical domains due to the lack of standard image-text pairs. Existing dermatological datasets are limited in both scale and depth, offering only single-label annotations across a narrow range of diseases instead of rich textual descriptions, and lacking the crucial clinical context needed for real-world applications. To address these limitations, we present Derm1M, the first large-scale vision-language dataset for dermatology, comprising 1,029,761 image-text pairs. Built from diverse educational resources and structured around a standard ontology collaboratively developed by experts, Derm1M provides comprehensive coverage for over 390 skin conditions across four hierarchical levels and 130 clinical concepts with rich contextual information such as medical history, symptoms, and skin tone. To demonstrate Derm1M's potential in advancing both AI research and clinical application, we pretrained a series of CLIP-like models, collectively called DermLIP, on this dataset. The DermLIP family significantly outperforms state-of-the-art foundation models on eight diverse datasets across multiple tasks, including zero-shot skin disease classification, clinical and artifacts concept identification, few-shot/full-shot learning, and cross-modal retrieval. Our dataset and code are available at https://github.com/SiyuanYan1/Derm1M.

Diffusion-based video depth estimation methods have achieved remarkable success with strong generalization ability. However, predicting depth for long videos remains challenging. Existing methods typically split videos into overlapping sliding windows, leading to accumulated scale discrepancies across different windows, particularly as the number of windows increases. Additionally, these methods rely solely on 2D diffusion priors, overlooking the inherent 3D geometric structure of video depths, which results in geometrically inconsistent predictions.In this paper, we propose DepthSync, a novel, training-free framework using diffusion guidance to achieve scale- and geometry-consistent depth predictions for long videos. Specifically, we introduce scale guidance to synchronize the depth scale across windows and geometry guidance to enforce geometric alignment within windows based on the inherent 3D constraints in video depths. These two terms work synergistically, steering the denoising process toward consistent depth predictions. Experiments on various datasets validate the effectiveness of our method in producing depth estimates with improved scale and geometry consistency, particularly for long videos.

Event cameras capture sparse, high-temporal-resolution visual information, making them particularly suitable for challenging environments with high-speed motion and strongly varying lighting conditions. However, the lack of large datasets with dense ground-truth depth annotations hinders learning-based monocular depth estimation from event data. To address this limitation, we propose a cross-modal distillation paradigm to generate dense proxy labels leveraging a Vision Foundation Model (VFM). Our strategy requires an event stream spatially aligned with RGB frames, a simple setup even available off-the-shelf, and exploits the robustness of large-scale VFMs.Additionally, we propose to adapt VFMs, either a vanilla one like Depth Anything v2 (DAv2), or deriving from it a novel recurrent architecture to infer depth from monocular event cameras. We evaluate our approach using synthetic and real-world datasets, demonstrating that i) our cross-modal paradigm achieves competitive performance compared to fully supervised methods without requiring expensive depth annotations, and ii) our VFM-based models achieve state-of-the-art performance

With the superior sensitivity of event cameras to high-speed motion and extreme lighting conditions, event-based monocular depth estimation has gained popularity to predict structural information about surrounding scenes in challenging environments. However, the scarcity of labeled event data constrains prior supervised learning methods. Unleashing the promising potential of the existing RGB-based depth foundation model, DAM, we propose Depth Any Event stream (EventDAM) to achieve high-performance event based monocular depth estimation in an annotation-free manner. EventDAM effectively combines paired dense RGB images with sparse event data by incorporating three key cross-modality components: Sparsity-aware Feature Mixture (SFM), Sparsity-aware Feature Distillation (SFD), and Sparsity-invariant Consistency Module (SCM). With the proposed sparsity metric, SFM mixes features from RGB images and event data to generate auxiliary depth predictions, while SFD facilitates adaptive feature distillation. Furthermore, SCM ensures output consistency across varying sparsity levels in event data, thereby endowing EventDAM with zero shot capabilities across diverse scenes. Extensive experiments across a variety of benchmark datasets, compared to approaches using diverse input modalities, robustly substantiate the generalization and zero-shot capabilities of EventDAM.

We propose DepR, a depth-guided single-view scene reconstruction framework that integrates instance-level diffusion within a compositional paradigm. Instead of reconstructing the entire scene holistically, DepR generates individual objects and subsequently composes them into a coherent 3D layout. Unlike previous methods that use depth solely for object layout estimation during inference and therefore fail to fully exploit its rich geometric information, DepR leverages depth throughout both training and inference. Specifically, we introduce depth-guided conditioning to effectively encode shape priors into diffusion models. During inference, depth further guides DDIM sampling and layout optimization, enhancing alignment between the reconstruction and the input image. Despite being trained on limited synthetic data, DepR achieves state-of-the-art performance and demonstrates strong generalization in single-view scene reconstruction, as shown through evaluations on both synthetic and real-world datasets.