ICCV 2025 Papers

Text-based diffusion models have made significant breakthroughs in generating high-quality images and videos from textual descriptions. However, the lengthy sampling time of the denoising process remains a significant bottleneck in practical applications. Previous methods either ignore the statistical relationships between adjacent steps or rely on attention or feature similarity between them, which often only works with specific network structures. To address this issue, we discover a new statistical relationship in the transition operator between adjacent steps, focusing on the relationship of the outputs from the network. This relationship does not impose any requirements on the network structure. Based on this observation, we propose a novel training-free acceleration method called LTC-Accel, which uses the identified relationship to estimate the current transition operator based on adjacent steps. Due to no specific assumptions regarding the network structure, LTC-Accel is applicable to almost all diffusion-based methods and orthogonal to almost all existing acceleration techniques, making it easy to combine with them. Experimental results demonstrate that LTC-Accel significantly speeds up sampling in text-to-image and text-to-video synthesis while maintaining competitive sample quality. Specifically, LTC-Accel achieves a speedup of 1.67 times in Stable Diffusion v2 and a speedup of 1.55 times in video generation models. When combined with distillation models, LTC-Accel achieves a remarkable 10 times speedup in video generation, allowing real-time generation of more than 16 FPS. Our code (include colab version) is available on https://zhushangwen.github.io/LTC-accel.io/.

Query-based methods with dense features have demonstrated remarkable success in 3D object detection tasks. However, the computational demands of these models, particularly with large image sizes and multiple transformer layers, pose significant challenges for efficient running on edge devices. Existing pruning and distillation methods either need retraining or are designed for ViT models, which are hard to migrate to 3D detectors. To address this issue, we propose a zero-shot runtime pruning method for transformer decoders in 3D object detection models. The method, termed tgGBC (trim keys gradually Guided By Classification scores), systematically trims keys in transformer modules based on their importance. We expand the classification score to multiply it with the attention map to get the importance score of each key and then prune certain keys after each transformer layer according to their importance scores. Our method achieves a 1.99x speedup in the transformer decoder of the latest ToC3D model, with only a minimal performance loss of less than 1%. Interestingly, for certain models, our method even enhances their performance. Moreover, we deploy 3D detectors with tgGBC on an edge device, further validating the effectiveness of our method. The code can be found at https://github.com/iseri27/tg_gbc.

Tissue segmentation in pathology images is crucial for computer-aided diagnostics of human cancers. Traditional tissue segmentation models rely heavily on large-scale labeled datasets, where every tissue type must be annotated by experts. However, due to the complexity of tumor micro-environment, collecting annotations for all possible tissue types is challenging, which makes the traditional methods ineffective in segmenting unseen tissue types with zero training samples. With the rapid development of vision-language models (VLMs), recent studies extend their powerful zero-shot capabilities to pixel-level segmentation tasks, where the model is trained only on seen classes but can perform tissue segmentation on both seen and unseen categories in the testing phase. However, these VLM-based zero-shot segmentation models still require substantial annotation efforts on seen classes. To attach desirable segmentation performance on both seen and unseen categories with limited labeled data, we propose AcZeroTS, a novel active learning framework for zero-shot tissue segmentation in pathology images. Specifically, AcZeroTS is built on a VLM-based prototype-guided zero-shot segmentation model called ProZS. We introduce a novel active selection criterion to select the most valuable samples for annotation on seen classes, which not only considers both uncertainty and diversity of unlabeled samples, but also ensures that the generated prototypes of ProZS can effectively summarize both seen and unseen classes during inference. We evaluate our method on two pathology datasets (TNBC and HPBC) as well as a natural dataset (Pascal VOC 2012), and the experimental results demonstrate the superiority of our method in comparison with the existing studies.

With the rapid advancement of Multi-modal Large Language Models (MLLMs), several diagnostic benchmarks have recently been developed to assess these models' multimodal reasoning proficiency. However, these benchmarks are restricted to assessing primarily the visual aspect and do not examine the holistic audio-visual (AV) understanding. Moreover, currently, there are no benchmarks that investigate the capabilities of AVLLMs to calibrate their responses when presented with perturbed inputs. To this end, we introduce Audio-Visual Trustworthiness assessment Benchmark (AVTrustBench), comprising 600K samples spanning over 9 meticulously crafted tasks, evaluating the capabilities of AVLLMs across three distinct dimensions: Adversarial Attack, Compositional Reasoning, and Modality-specific Dependency. Using our benchmark, we extensively evaluate 16 state-of-the-art AVLLMs. The findings reveal that the majority of existing models fall significantly short of achieving human like comprehension, offering valuable insights for future research directions. To alleviate the limitations in the existing approaches, we further propose a robust, model agnostic calibrated audio-visual preference optimization based training strategy CAVPref, obtaining a gain up to 30.19% across all 9 tasks.

The advancement of Multimodal Large Language Models (MLLMs) has driven significant progress in Visual Question Answering (VQA), evolving from Single to Multi Image VQA (MVQA). However, the increased number of images in MVQA inevitably introduces substantial visual redundancy that is irrelevant to question answering, negatively impacting both accuracy and efficiency. To address this issue, existing methods lack flexibility in controlling the number of compressed visual tokens and tend to produce discrete visual fragments, which hinder MLLMs' ability to comprehend images holistically. In this paper, we propose a straightforward yet universal Adaptive Visual Anchoring strategy, which can be seamlessly integrated into existing MLLMs, offering significant accuracy improvements through adaptive compression. Meanwhile, to balance the results derived from both global and compressed visual input, we further introduce a novel collaborative decoding mechanism, enabling optimal performance. Extensive experiments validate the effectiveness of our method, demonstrating consistent performance improvements across various MLLMs. The code will be publicly available.

We propose AV-Link, a unified framework for Video-to-Audio (A2V) and Audio-to-Video (A2V) generation that leverages the activations of frozen video and audio diffusion models for temporally-aligned cross-modal conditioning. The key to our framework is a Fusion Block that facilitates bidirectional information exchange between video and audio diffusion models through temporally-aligned self-attention operations. Unlike prior work that uses dedicated models for A2V and V2A tasks and relies on pretrained feature extractors, AV-Link achieves both tasks in a single self-contained framework, directly leveraging features obtained by the complementary modality (i.e. video features to generate audio, or audio features to generate video). Extensive evaluations demonstrate that AV-Link achieves substantial improvements in audio-video synchronization, outperforming more expensive baselines such as MovieGen V2A model.

We introduce AV-Flow, an audio-visual generative model that animates photo-realistic 4D talking avatars given only text input. In contrast to prior work that assumes an existing speech signal, we synthesize speech and vision jointly. We demonstrate human-like speech synthesis, synchronized lip motion, lively facial expressions and head pose; all generated from just text characters. The core premise of our approach lies in the architecture of our two parallel diffusion transformers. Intermediate highway connections ensure communication between the audio and visual modalities, and thus, synchronized speech intonation and facial dynamics (e.g., eyebrow motion). Our model is trained with flow matching, leading to expressive results and fast inference. In case of dyadic conversations, AV-Flow produces an always-on avatar, that actively listens and reacts to the audio-visual input of a user. Through extensive experiments, we show that our method outperforms prior work, synthesizing natural-looking 4D talking avatars.

Recent advancements in reasoning optimization have greatly enhanced the performance of large language models (LLMs). However, existing work fails to address the complexities of audio-visual scenarios, underscoring the need for further research. In this paper, we introduce AURELIA, a novel actor-critic based audio-visual (AV) reasoning framework that distils structured, step-by-step reasoning into AVLLMs at test time, improving their ability to process complex multi-modal inputs without additional training or fine-tuning. To further advance AVLLM reasoning skills, we present AVReasonBench, a challenging benchmark comprising 4500 audio-visual questions, each paired with detailed step-by-step reasoning. Our benchmark spans six distinct tasks, including AV-GeoIQ, which evaluates AV reasoning combined with geographical and cultural knowledge. Evaluating 18 AVLLMs on AVReasonBench reveals significant limitations in their multi-modal reasoning capabilities. Using AURELIA, we achieve up to a 100% relative improvement, demonstrating its effectiveness. This performance gain highlights the potential of reasoning-enhanced data generation for advancing AVLLMs in real-world applications.

While 3D facial animation has made impressive progress, challenges still exist in realizing fine-grained stylized 3D facial expression manipulation due to the lack of appropriate datasets. In this paper, we introduce the AUBlendSet, a 3D facial dataset based on AU-Blendshape representation for fine-grained facial expression manipulation across identities. AUBlendSet is a blendshape data collection based on 32 standard facial action units (AUs) across 500 identities, along with an additional set of facial postures annotated with detailed AUs. Based on AUBlendSet, we propose AUBlendNet to learn AU-Blendshape basis vectors for different character styles. AUBlendNet predicts, in parallel, the AU-Blendshape basis vectors of the corresponding style for a given identity mesh, thereby achieving stylized 3D emotional facial manipulation. We comprehensively validate the effectiveness of AUBlendSet and AUBlendNet through tasks such as stylized facial expression manipulation, speech-driven emotional facial animation, and emotion recognition data augmentation. Through a series of qualitative and quantitative experiments, we demonstrate the potential and importance of AUBlendSet and AUBlendNet in 3D facial animation tasks. To the best of our knowledge, AUBlendSet is the first dataset, and AUBlendNet is the first network for continuous 3D facial expression manipulation for any identity through facial AUs. Our source code is available at https://github.com/wslh852/AUBlendNet.git.

Parametric body models offer expressive 3D representation of humans across a wide range of poses, shapes, and facial expressions, typically derived by learning a basis over registered 3D meshes. However, existing human mesh modeling approaches struggle to capture detailed variations across diverse body poses and shapes, largely due to limited training data diversity and restrictive modeling assumptions. Moreover, the common paradigm first optimizes the external body surface using a linear basis, then regresses internal skeletal joints from surface vertices. This approach introduces problematic dependencies between internal skeleton and outer soft tissue, limiting direct control over body height and bone lengths. To address these issues, we present ATLAS, a high-fidelity body model learned from 600k high-resolution scans captured using 240 synchronized cameras. Unlike previous methods, we explicitly decouple the shape and skeleton bases by grounding our mesh representation in the human skeleton. This decoupling enables enhanced shape expressivity, fine-grained customization of body attributes, and keypoint fitting independent of external soft-tissue characteristics. ATLAS outperforms existing methods by fitting unseen subjects in diverse poses more accurately, and quantitative evaluations show that our non-linear pose correctives more effectively capture complex poses compared to linear models.

Vision-language tracking aims to locate the target object in the video sequence using a template patch and a language description provided in the initial frame. To achieve robust tracking, especially in complex long-term scenarios that reflect real-world conditions as recently highlighted by MGIT, it is essential not only to characterize the target features but also to utilize the context features related to the target. However, the visual and textual target-context cues derived from the initial prompts generally align only with the initial target state. Due to their dynamic nature, target states are constantly changing, particularly in complex long-term sequences. It is intractable for these cues to continuously guide Vision-Language Trackers (VLTs). Furthermore, for the text prompts with diverse expressions, our experiments reveal that existing VLTs struggle to discern which words pertain to the target or the context, complicating the utilization of textual cues. In this work, we present a novel tracker named ATCTrack, which can obtain multimodal cues Aligned with the dynamic target states through comprehensive Target-Context feature modeling, thereby achieving robust tracking. Specifically, (1) for the visual modality, we propose an effective temporal visual target-context modeling approach that provides the tracker with timely visual cues. (2) For the textual modality, we achieve precise target words identification solely based on textual content, and design an innovative context words calibration method to adaptively utilize auxiliary context words. (3) We conduct extensive experiments on mainstream benchmarks and ATCTrack achieves a new SOTA performance. The code and models will be released at: https://github.com/XiaokunFeng/ATCTrack.

Vision-language models such as CLIP have recently propelled open-vocabulary dense prediction tasks by enabling recognition of a broad range of visual concepts. However, CLIP still struggles with fine-grained, region-level understanding, hindering its effectiveness on these dense prediction tasks. We identify two pivotal factors required to address this limitation: semantic coherence and fine-grained vision-language alignment. Current adaptation methods often improve fine-grained alignment at the expense of semantic coherence, and often rely on extra modules or supervised fine-tuning. To overcome these issues, we propose Any-to-Any Self-Distillation (ATAS), a novel approach that simultaneously enhances semantic coherence and fine-grained alignment by leveraging a model's own knowledge across all representation levels. Unlike prior methods, ATAS uses only unlabeled images and an internal self-distillation process to refine CLIP's representations, preserving local semantic consistency while sharpening local detail recognition. On open-vocabulary object detection and semantic segmentation benchmarks, ATAS achieves substantial performance gains, outperforming baseline CLIP models. These results validate the effectiveness of our approach and underscore the importance of jointly maintaining semantic coherence and fine-grained alignment for advanced open-vocabulary dense prediction.

Albeit existing Single-Domain Generalized Object Detection (Single-DGOD) methods enable models to generalize to unseen domains, most assume that the training and testing data share the same label space. In real-world scenarios, unseen domains often introduce previously unknown objects, a challenge that has been largely overlooked. In this paper, we tackle the practical problem of Single-domain Generalizable Open-Set Object Detection (SG-OSOD), which addresses both unseen domains and unknown classes. We identify two key challenges: (1) detecting unknown classes with only known-class data, and (2) learning robust features to mitigate domain shift. To address these challenges, we propose the framework termed \texttt ASGS , which leverages adaptive subgraph structures to enhance the understanding of unknown scenes and classes. \texttt ASGS consists of Subgraph-wise Unknown-class Learning (SUL) and Class-wise Embedding Compaction (CEC). SUL employs non-parametric methods to detect unknown samples and performs Adaptive Subgraph Searching (ASS) for high-order structural feature extraction, enabling domain-robust unknown class learning. Moreover, the CEC module enhances class discrimination robustness through contrastive learning, which results in more compact class clusters in unknown scenarios. Experimental results demonstrate the effectiveness of the proposed \texttt ASGS .

We propose ASCENT, a novel framework for tracking neurons in 3D fluorescence microscopy recordings without relying on manual track annotations. ASCENT leverages self-supervised contrastive learning to learn robust, discriminative embeddings from detected neuron candidates. At its core is a volume compression module that transforms full 3D volumetric data into an efficient 2D representation by iteratively projecting along the z-axis and integrating positional information. This compressed representation is processed by a deep encoder (e.g., ResNet or Vision Transformer) to yield robust feature vectors that capture both appearance and spatial relationships among neurons. Extensive experiments on both in-house and public datasets demonstrate that ASCENT achieves state-of-the-art tracking performance with fast inference speed while removing the need for costly manual labeling and heavy pre- and post-processing. Our results suggest that this approach provides a scalable solution for 3D neuron tracking and holds promise for applications such as inter-individual neuron identity matching and demixing overlapping cells.

Visual relation detection (VRD) is the task of identifying the relationships between objects in a scene. VRD models trained solely on relation detection data struggle to generalize beyond the relations on which they are trained. While prompt tuning has been used to adapt vision-language models (VLMs) for VRD, it uses handcrafted prompts and struggles with novel or complex relations. We argue that instruction tuning offers a more effective solution by fine-tuning VLMs on diverse instructional data. We thus introduce ART, an Adaptive Relation Tuning framework that adapts VLMs for VRD through instruction tuning and strategic instance selection. By converting VRD datasets into an instruction-tuning format and employing an adaptive sampling algorithm, ART directs the VLM to focus on informative relations while maintaining generalizability. Specifically, we focus on the relation classification, where subject-object boxes are given and the model predicts the predicate between them. We tune on a held-in set and evaluate across multiple held-out datasets of varying complexity. Our approach strongly improves over its baselines and can infer unseen relation concepts, a capability absent in mainstream VRD methods. We demonstrate ART's practical value by using the predicted relations for segmenting complex scenes.

Recent advancements in large-scale generative models have significantly improved the quality and diversity of 3D shape generation. However, most existing methods focus primarily on generating static 3D models, overlooking the potential dynamic nature of certain shapes, such as humanoids, animals, and insects. To address this gap, we focus on rigging, a fundamental task in animation that establishes skeletal structures and skinning for 3D models. In this paper, we introduce OmniRig, the first large-scale rigging dataset, comprising 79,499 meshes with detailed skeleton and skinning information. Unlike traditional benchmarks that rely on predefined standard poses (e.g., A-pose, T-pose), our dataset embraces diverse shape categories, styles, and poses. Leveraging this rich dataset, we propose ARMO, a novel rigging framework that utilizes an autoregressive model to predict both joint positions and connectivity relationships in a unified manner. By treating the skeletal structure as a complete graph and discretizing it into tokens, we encode the joints using an auto-encoder to obtain a latent embedding and an autoregressive model to predict the tokens. A mesh-conditioned latent diffusion model is used to predict the latent embedding for conditional skeleton generation. Our method addresses the limitations of regression-based approaches, which often suffer from error accumulation and suboptimal connectivity estimation. Through extensive experiments on the OmniRig dataset, our approach achieves state-of-the-art performance in skeleton prediction, demonstrating improved generalization across diverse object categories. The code and dataset will be made public for academic use upon acceptance.

Face-to-face communication, as a common human activity, motivates the research on interactive head generation. A virtual agent can generate motion responses with both listening and speaking capabilities based on the audio or motion signals of the other user and itself. However, previous clip-wise generation paradigm or explicit listener/speaker generator-switching methods have limitations in future signal acquisition, contextual behavioral understanding, and switching smoothness, making it challenging to be real-time and realistic.In this paper, we propose an autoregressive (AR) based frame-wise framework called ARIG to realize the real-time generation with better interaction realism. To achieve real-time generation, we model motion prediction as a non-vector-quantized AR process. Unlike discrete codebook-index prediction, we represent motion distribution using diffusion procedure, achieving more accurate predictions in continuous space. To improve interaction realism, we emphasize interactive behavior understanding (IBU) and detailed conversational state understanding (CSU). In IBU, based on dual-track dual-modal signals, we summarize short-range behaviors through bidirectional-integrated learning and perform contextual understanding over long ranges. In CSU, we use voice activity signals and context features of IBU to understand the various states (interruption, feedback, pause, etc.) that exist in actual conversations. These serve as conditions for the final progressive motion prediction. Extensive experiments have verified the effectiveness of our model.

Video large language models have not yet been widely deployed, largely due to their tendency to hallucinate. Typical benchmarks for Video-LLMs rely simply on multiple choice questions. Unfortunately, VideoLLMs hallucinate far more aggressively on freeform text generation tasks like video captioning than they do on multiple choice verification tasks. To address this weakness, we propose ARGUS, a VideoLLM benchmark that measures freeform video captioning performance. By comparing VideoLLM outputs to human ground truth captions, ARGUS quantifies dual metrics. First, we measure the rate of hallucinations in the form of incorrect statements about video content or temporal relationships. Second, we measure the rate at which the model omits important descriptive details. Together, these metrics form a comprehensive view of video captioning.