ICCV 2025 Papers

Many real-world applications, such as interactive photo retouching, artistic content creation, and product design, require flexible and iterative image editing. However, existing image editing methods primarily focus on achieving the desired modifications in a single step, which often struggles with ambiguous user intent, complex transformations, or the need for progressive refinements. As a result, these methods frequently produce inconsistent outcomes or fail to meet user expectations. To address these challenges, we propose a multi-turn image editing framework that enables users to iteratively refine their edits, progressively achieving more satisfactory results. Our approach leverages flow matching for accurate image inversion and a dual-objective Linear Quadratic Regulators (LQR) for stable sampling, effectively mitigating error accumulation. Additionally, by analyzing the layer-wise roles of transformers, we introduce a adaptive attention highlighting method that enhances editability while preserving multi-turn coherence. Extensive experiments demonstrate that our framework significantly improves edit success rates and visual fidelity compared to existing methods. The code is available at: https://zhouzj-dl.github.io/Multi-turn_Consistent_Image_Editing/.

Overlapping text poses significant challenges for text-related perception tasks, particularly in open scenes characterized by diverse fonts and visual effects. While existing research has primarily addressed the overlapping problem in documents, its applicability to other scenes remains limited. To bridge this gap, we propose a new task of multi-scenario overlapping text segmentation and introduce a corresponding real dataset in both English and Chinese, spanning various contexts such as printed text, bills, artistic designs, and house numbers. To further enhance the generalization of overlapping text segmentation models, we propose a hierarchical training data synthesis strategy that simulates diverse overlapping patterns across different scenarios. Furthermore, we found that depth maps can provide clear relative position relationships in three-dimensional space, assisting the model in capturing complex overlapping relationships between text instances. Building on this insight, we present a depth-guided decoder that seamlessly integrates image and depth features to capture overlapping interactions. Our proposed model achieves a 5.3% improvement in text mIoU and a 6.4% improvement in overall mIoU compared to existing SOTA methods on our benchmark and SignaTR6k datasets, respectively.

Camouflaged scenes, where objects blend seamlessly into their environments, pose significant challenges to both human observers and computer vision systems. These objects match the background in color, texture, and shape, making them difficult to detect. To this end, we propose leveraging the Segment Anything Model (SAM) to tackle this challenging task effectively. Specifically, we propose how to exploit SAM without requiring any manual prompts by proposing several ideas. At the core of our method lies the rich information extracted through multi-modal prompts. At first, we generate an image caption using the BLIP model and obtain its text embedding through the use of a text encoder. We then generate a visual embedding through the vision encoder of the BLIP model and use both as inputs to SAM to provide additional semantic information about the image. Finally, we propose a couple of architectural novelties, a) we effectively integrate the multi-modal information in SAM through a multi-level adapter and b) we replace the dense embedding of SAM with the image embedding of its image encoder. Our method achieves new state-of-the-art performance in 11 out of 12 metrics in three benchmark datasets for camouflaged detection. Additionally, our method can be successfully adapted to other tasks such as medical image segmentation performing on par or even outperforming the state-of-the-art methods. Our code is available in https://github.com/ic-qialanqian/Vision-Language-SAM.

Conventional person re-identification (ReID) research is often limited to single-modality sensor data from static cameras, which fails to address the complexities of real-world scenarios where multi-modal signals are increasingly prevalent. For instance, consider an urban ReID system integrating stationary RGB cameras, nighttime infrared sensors, and UAVs equipped with dynamic tracking capabilities. Such systems face significant challenges due to variations in camera perspectives, lighting conditions, and sensor modalities, hindering effective person ReID.To address these challenges, we introduce the MP-ReID benchmark, a novel dataset designed specifically for multi-modality and multi-platform ReID. This benchmark uniquely compiles data from 1,930 identities across diverse modalities, including RGB, infrared, and thermal imaging, captured by both UAVs and ground-based cameras in indoor and outdoor environments.Building on this benchmark, we introduce Uni-Prompt ReID, a framework with specific-designed prompts, tailored for cross-modality and cross-platform scenarios. Our method consistently outperforms state-of-the-art approaches, establishing a robust foundation for future research in complex and dynamic ReID environments. Additionally, our dataset will be made publicly available to support further advancements.

The success of multi-modal foundational models can be partly attributed to their diverse, billions scale training data. By nature, web data contains human faces and descriptions of individuals. Thus, these models pose potentially widespread privacy issues. Recently, identity membership inference attacks (IMIAs) against the CLIP model showed that membership of an individual's name and image within training data can be reliably inferred. This work formalizes the problem of identity extraction, wherein an attacker can reliably extract the names of individuals given their images only. We provide the following contributions (i) we adapt a previous IMIA to the problem of selecting the correct name among a large set and show that the method scales to millions of names (ii) we design an attack that outperforms the adapted baseline (iii) we show that an attacker can extract names via optimization only. To demonstrate the interest of our framework, we show how identity extraction can be used to audit model privacy. Indeed, a family of prominent models that advertise blurring faces before training to protect privacy is still highly vulnerable to attack.

Generating human videos from a single image while ensuring high visual quality and precise control is a challenging task, especially in complex scenarios involving multiple individuals and interactions with objects. Existing methods, while effective for single-human cases, often fail to handle the intricacies of multi-identity interactions because they struggle to associate the correct pairs of human appearance and pose condition and model the distribution of 3D-aware dynamics. To address these limitations, we present Structural Video Diffusion, a novel framework designed for generating realistic multi-human videos. Our approach introduces two core innovations: identity-specific embeddings to maintain consistent appearances across individuals and a structural learning mechanism that incorporates depth and surface-normal cues to model human-object interactions. Additionally, we expand existing human video dataset with 25K new videos featuring diverse multi-human and object interaction scenarios, providing a robust foundation for training. Experimental results demonstrate that Structural Video Diffusion achieves superior performance in generating lifelike, coherent videos for multiple subjects with dynamic and rich interactions, advancing the state of human-centric video generation. Code is available at https://github.com/zhenzhiwang/Multi-HumanVid

Recent face anti-spoofing (FAS) methods have shown remarkable cross-domain performance by employing vision-language models like CLIP. However, existing CLIP-based FAS models do not fully exploit CLIP's patch embedding tokens, failing to detect critical spoofing clues. Moreover, these models rely on a single text prompt per class (e.g, 'live' or 'fake'), which limits generalization. To address these issues, we propose MVP-FAS, a novel framework incorporating two key modules: Multi-View Slot attention (MVS) and Multi-Text Patch Alignment (MTPA). Both modules utilize multiple paraphrased texts to generate generalized features and reduce dependence on domain-specific text. MVS extracts local detailed spatial features and global context from patch embeddings by leveraging diverse texts with multiple perspectives. MTPA aligns patches with multiple text representations to improve semantic robustness. Extensive experiments demonstrate that MVP-FAS achieves superior generalization performance, outperforming previous state-of-the-art methods on cross-domain datasets.

We introduce the first data-driven multi-view 3D point tracker, designed to track arbitrary points in dynamic scenes using multiple camera views. Unlike existing monocular trackers, which struggle with depth ambiguities and occlusion, or prior multi-camera methods that require over 20 cameras and tedious per-sequence optimization, our feed-forward model directly predicts 3D correspondences using a practical number of cameras (e.g., four), enabling robust and accurate online tracking. Given known camera poses and either sensor-based or estimated multi-view depth, our tracker fuses multi-view features into a unified point cloud and applies k-nearest-neighbors correlation alongside a transformer-based update to reliably estimate long-range 3D correspondences, even under occlusion. We train on 5K synthetic multi-view Kubric sequences and evaluate on two real-world benchmarks--Panoptic Studio and DexYCB--achieving median trajectory errors of 3.1 cm and 2.0cm, respectively. Our method generalizes well to diverse camera setups of 1-8 views with varying vantage points and video lengths of 24-150 frames. By releasing our tracker alongside training and evaluation datasets, we aim to set a new standard for multi-view 3D tracking research and provide a practical tool for real-world applications. Project page: https://ethz-vlg.github.io/mvtracker.

Composed Image Retrieval (CIR) aims to retrieve a target image using a query that combines a reference image and a textual description, benefiting users to express their intent more effectively. Despite significant advances in CIR methods, two unresolved problems remain: 1) existing methods overlook multi-schema interaction due to the lack of fine-grained explicit visual supervision, which hinders the capture of complex correspondences, and 2) existing methods overlook noisy negative pairs formed by potential corresponding query-target pairs, which increases confusion. To address these problems, we propose a Multi-schemA Proximity Network (MAPNet) for CIR, consisting of two key components: Multi-Schema Interaction (MSI) and Relaxed Proximity Loss (RPLoss). Specifically, MSI leverages textual descriptions as an implicit guide to establish correspondences between multiple objects and attributes in the reference and target images, enabling multi-schema interactions. Then, RPLoss further aligns the query and target features while avoiding the poison of noisy negative pairs by denoising and reweighting strategy. Comprehensive experiments conducted on CIRR, FashionIQ, and LaSCo demonstrate that MAPNet achieves competitive results against state-of-the-art CIR methods.

Sketch animation, which brings static sketches to life by generating dynamic video sequences, has found widespread applications in GIF design, cartoon production, and daily entertainment. While current methods for sketch animation perform well in single-object sketch animation, they struggle in multi-object scenarios. By analyzing their failures, we identify two major challenges of transitioning from single-object to multi-object sketch animation: object-aware motion modeling and complex motion optimization. For multi-object sketch animation, we propose MoSketch based on iterative optimization through Score Distillation Sampling (SDS) and thus animating a multi-object sketch in a training-data free manner. To tackle the two challenges in a divide-and-conquer strategy, MoSketch has four novel modules, i.e., LLM-based scene decomposition, LLM-based motion planning, multi-grained motion refinement, and compositional SDS. Extensive qualitative and quantitative experiments demonstrate the superiority of our method over existing sketch animation approaches. MoSketch takes a pioneering step towards multi-object sketch animation, opening new avenues for future research and applications.

We present Multi-Modal Multi-Task Unified Embedding Model (M3T-UEM), a framework that advances vision-language matching and retrieval by leveraging a large language model (LLM) backbone. While concurrent LLM-based approaches like VLM2VEC, MM-Embed, NV-Embed, and MM-GEM have demonstrated impressive capabilities in multi-modal and multi-task scenarios, our work introduces novel mechanisms for task-adaptive learning and embedding extraction that further enhance the potential of LLM-based retrieval systems. Our key technical contribution lies in the development of a task-aware contrastive learning framework with an automated Bayesian weighing mechanism. This approach provides a principled way to balance multiple tasks during training, departing from conventional contrastive learning strategies. We further enhance the framework through a multiple-token summarization strategy and an auxiliary language modeling objective, which together significantly improve retrieval performance.Comprehensive experiments on M-BEIR and ICinW benchmarks demonstrate M3T-UEM's effectiveness, showing competitive or superior performance compared to both traditional encoder-based methods and recent LLM-based approaches. Furthermore, we demonstrate particular strengths in handling compositional conceptual changes and multilingual scenarios owing to the incorporation of an LLM backbone where the method drastically outperforms CLIP in zero-shot settings, often by orders of magnitude.

Procedural videos are critical for learning new tasks. Temporal action segmentation (TAS), which classifies the action in every video frame, has become essential for understanding procedural videos. Existing TAS models, however, learn a fixed-set of tasks at training and unable to adapt to novel tasks at test time. Thus, we introduce the new problem of Multi-Modal Few-shot Temporal Action Segmentation (MMF-TAS) to learn open-set models that can generalize to novel procedural tasks with minimal visual/textual examples. We propose the first MMF-TAS framework, by designing a Prototype Graph Network (PGNet). In PGNet, a Prototype Building Block summarizes action information from support videos of the novel tasks via an Action Relation Graph, and encodes this information into action prototypes via a Dynamic Graph Transformer. Next, a Matching Block compares action prototypes with query videos to infer framewise action labels. To exploit the advantages of both visual and textual modalities, we compute separate action prototypes for each modality and combine the two modalities through prediction fusion to avoid overfitting on one modality. By extensive experiments on procedural datasets, we show our method successfully adapts to novel tasks during inference and significantly outperforms baselines. Our code is available at https://github.com/ZijiaLewisLu/ICCV2025-MMF-TAS.

Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2x speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3x speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The code is available at https://github.com/HYUNJS/STTM.

In zero-shot setting, test-time adaptation adjusts pre-trained models using unlabeled data from the test phase to enhance performance on unknown test distributions. Existing cache-enhanced TTA methods rely on a low-entropy criterion to select samples for prototype construction, assuming intra-class compactness. However, low-entropy samples may be unreliable under distribution shifts, and the resulting prototypes may not ensure compact intra-class distributions. This study identifies a positive correlation between cache-enhanced performance and intra-class compactness. Based on this observation, we propose a Multi-Cache enhanced Prototype-based Test-Time Adaptation (MCP) featuring three caches: an entropy cache for initializing prototype representations with low-entropy samples, an align cache for integrating visual and textual information to achieve compact intra-class distributions, and a negative cache for prediction calibration using high-entropy samples. We further developed MCP++, a framework incorporating cross-modal prototype alignment and residual learning, introducing prototype residual fine-tuning. Comparative and ablation experiments across 15 downstream tasks demonstrate that the proposed method and framework achieve state-of-the-art generalization performance.

We present Multi-Baseline Gaussian Splatting (MuGS), a generalized feed-forward approach for novel view synthesis that effectively handles diverse baseline settings, including sparse input views with both small and large baselines. Specifically, we integrate features from Multi-View Stereo (MVS) and Monocular Depth Estimation (MDE) to enhance feature representations for generalizable reconstruction. Next, We propose a projection-and-sampling mechanism for deep depth fusion, which constructs a fine probability volume to guide the regression of the feature map. Furthermore, We introduce a reference-view loss to improve geometry and optimization efficiency. We leverage 3D Gaussian representations to accelerate training and inference time while enhancing rendering quality. MuGS achieves state-of-the-art performance across multiple baseline settings and diverse scenarios ranging from simple objects (DTU) to complex indoor and outdoor scenes (RealEstate10K). We also demonstrate promising zero-shot performance on the LLFF and Mip-NeRF 360 datasets. Code is available at https://github.com/EuclidLou/MuGS.

Embodied scene understanding requires not only comprehending visual-spatial information that has been observed but also determining where to explore next in the 3D physical world. Existing 3D Vision-Language (3D-VL) models primarily focus on grounding objects in static observations from 3D reconstruction, such as meshes and point clouds, but lack the ability to actively perceive and explore their environment. To address this limitation, we introduce Move to Understand (MTU3D), a unified framework that integrates active perception with 3D vision-language learning, enabling embodied agents to effectively explore and understand their environment. This is achieved by three key innovations 1) Online query-based representation learning, enabling direct spatial memory construction from RGB-D frames, eliminating the need for explicit 3D reconstruction. 2) A unified objective for grounding and exploration that represents unexplored locations as frontier queries and jointly optimizes object grounding and frontier selection. 3) End-to-end trajectory learning that combines Vision-Language-Exploration pre-training over a million diverse trajectories collected from both simulated and real-world RGB-D sequences. Extensive evaluations across various embodied navigation and question-answering benchmarks show that MTU3D outperforms state-of-the-art reinforcement learning and modular navigation approaches by 14%, 23%, 9%, and 2% in success rate on HM3D-OVON, GOAT-Bench, SG3D, and A-EQA, respectively. MTU3D's versatility enables navigation using diverse input modalities, including categories, language descriptions, and reference images. Additionally, we deploy it on a real robot to demonstrate its effectiveness in handling real-world data. These findings highlight the importance of bridging visual grounding and exploration for embodied intelligence.

Recent developments in Large Language Models (LLMs) pre-trained on extensive corpora have shown significant success in various natural language processing (NLP) tasks with minimal fine-tuning. This success offers new promise for robotics, which has long been constrained by the high cost of action-labeled data. We ask: given the abundant video data containing interaction-related knowledge available as a rich "corpus", can a similar generative pre-training approach be effectively applied to enhance robot learning? The key challenge is to identify an effective representation for autoregressive pre-training that benefits robot manipulation tasks.Inspired by the way humans learn new skills through observing dynamic environments, we propose that effective robotic learning should emphasize motion-related knowledge, which is closely tied to low-level actions and is hardware-agnostic, facilitating the transfer of learned motions to actual robot actions. To this end, we introduce Moto, which converts video content into latent Motion Token sequences by a Latent Motion Tokenizer, learning a bridging "language" of motion from videos in an unsupervised manner. We pre-train Moto-GPT through motion token autoregression, enabling it to capture diverse visual motion knowledge. After pre-training, Moto-GPT demonstrates the promising ability to produce semantically interpretable motion tokens, predict plausible motion trajectories, and assess trajectory rationality through output likelihood.To transfer learned motion priors to real robot actions, we implement a co-fine-tuning strategy that seamlessly bridges latent motion token prediction and real robot control. Extensive experiments show that the fine-tuned Moto-GPT exhibits superior robustness and efficiency on robot manipulation benchmarks, underscoring its effectiveness in transferring knowledge from video data to downstream visual manipulation tasks.

This paper addresses the challenge of text-conditioned streaming motion generation, which requires us to predict the next-step human pose based on variable-length historical motions and incoming texts. Existing methods struggle to achieve streaming motion generation, e.g., diffusion models are constrained by pre-defined motion lengths, while GPT-based methods suffer from delayed response and error accumulation problem due to discretized non-causal tokenization. To solve these problems, we propose MotionStreamer, a novel framework that incorporates a continuous causal latent space into a probabilistic autoregressive model. The continuous latents mitigate information loss caused by discretization and effectively reduce error accumulation during long-term autoregressive generation. In addition, by establishing temporal causal dependencies between current and historical motion latents, our model fully utilizes the available information to achieve accurate online motion decoding. Experiments show that our method outperforms existing approaches while offering more applications, including multi-round generation, long-term generation, and dynamic motion composition. Project Page: https://zju3dv.github.io/MotionStreamer/

Existing text-to-video methods struggle to transfer motion smoothly from a reference object to a target object with significant differences in appearance or structure between them. To address this challenge, we introduce MotionShot, a training-free framework capable of parsing reference-target correspondences in a fine-grained manner, thereby achieving high-fidelity motion transfer while preserving coherence in appearance. To be specific, MotionShot first performs semantic feature matching to ensure high-level alignments between the reference and target objects. It then further establishes low-level morphological alignments through reference-to-target shape retargeting. By encoding motion with temporal attention, our MotionShot can coherently transfer motion across objects, even in the presence of significant appearance and structure disparities, demonstrated by extensive experiments. The project page is available at: https://motionshot.github.io/.