ICCV 2025 Papers

Scene flow provides crucial motion information for autonomous driving. Recent LiDAR scene flow models utilize the rigid-motion assumption at the instance level, assuming objects are rigid bodies. However, these instance-level methods are not suitable for sparse radar point clouds. In this work, we present a novel Traffic-Aware Radar Scene-Flow (TARS) estimation method, which utilizes motion rigidity at the traffic level. To address the challenges in radar scene flow, we perform object detection and scene flow jointly and boost the latter. We incorporate the feature map from the object detector, trained with detection losses, to make radar scene flow aware of the environment and road users. From this, we construct a Traffic Vector Field (TVF) in the feature space to achieve holistic traffic-level scene understanding in our scene flow branch. When estimating the scene flow, we consider both point-level motion cues from point neighbors and traffic-level consistency of rigid motion within the space. TARS outperforms the state of the art on a proprietary dataset and the View-of-Delft dataset, improving the benchmarks by 23% and 15%, respectively.

This paper introduces Timestep-Adaptive Representation Alignment with Onset-Aware Conditioning (TARO), a novel framework for high-fidelity and temporally coherent video-to-audio synthesis. Built upon flow-based transformers, which offer stable training and continuous transformations for enhanced synchronization and audio quality, TARO introduces two key innovations: (1) Timestep-Adaptive Representation Alignment (TRA), which dynamically aligns latent representations by adjusting alignment strength based on the noise schedule, ensuring smooth evolution and improved fidelity, and (2) Onset-Aware Conditioning (OAC), which integrates onset cues that serve as sharp event-driven markers of audio-relevant visual moments to enhance synchronization with dynamic visual events. Extensive experiments on the VGGSound and Landscape datasets demonstrate that TARO outperforms prior methods, achieving relatively 53% lower Frechet Distance (FD), 29% lower Frechet Audio Distance (FAD), and a 97.19% Alignment Accuracy, highlighting its superior audio quality and synchronization precision.

We present TAR3D, a novel framework that consists of a 3D-aware Vector Quantized-Variational AutoEncoder (VQVAE) and a Generative Pre-trained Transformer (GPT) to generate high-quality 3D assets. The core insight of this work is to migrate the multimodal unification and promising learning capabilities of the next-token prediction paradigm to conditional 3D object generation. To achieve this, the3D VQ-VAE first encodes a wide range of 3D shapes into a compact triplane latent space and utilizes a set of discrete representations from a trainable codebook to reconstruct fine-grained geometries under the supervision of query point occupancy. Then, the 3D GPT, equipped with a custom triplane position embedding called TriPE, predicts the codebook index sequence with prefilling prompt tokensin an autoregressive manner so that the composition of 3D geometries can be modeled part by part. Extensive experiments on ShapeNet and Objaverse demonstrate that TAR3D can achieve superior generation quality over existing methods in text-to-3D and image-to-3D tasks.

Tracking Any Point (TAP) in a video is a challenging computer vision problem with many demonstrated applications in robotics, video editing, and 3D reconstruction. Existing methods for TAP rely heavily on complex tracking-specific inductive biases and heuristics, limiting their generality and potential for scaling. To address these challenges, we present TAPNext, a new approach that casts TAP as sequential masked token decoding. Our model is causal, tracks in a purely online fashion, and removes tracking-specific inductive biases. This enables TAPNext to run with minimal latency, and removes the temporal windowing required by many existing state of art trackers. Despite its simplicity, TAPNext achieves a new state-of-the-art tracking performance among both online and offline trackers. Finally, we present evidence that many widely used tracking heuristics emerge naturally in TAPNext through end-to-end training.

AI-generated content (AIGC) enables efficient visual creation but raises copyright and authenticity risks. As a common technique for integrity verification and source tracing, digital image watermarking is regarded as a potential solution to above issues. However, the widespread adoption and advancing capabilities of generative image editing tools have amplified malicious tampering risks, while simultaneously posing new challenges to passive tampering detection and watermark robustness. To address these challenges, this paper proposes a Tamper-Aware Generative image WaterMarking method named TAG-WM. The proposed method comprises four key modules: a dual-mark joint sampling (DMJS) algorithm for embedding copyright and localization watermarks into the latent space while preserving generative quality, the watermark latent reconstruction (WLR) utilizing reversed DMJS, a dense variation region detector (DVRD) leveraging diffusion inversion sensitivity to identify tampered areas via statistical deviation analysis, and the tamper-aware decoding (TAD) guided by localization results. The experimental results demonstrate that TAG-WM achieves state-of-the-art performance in both tampering robustness and localization capability even under distortion, while preserving lossless generation quality and maintaining a watermark capacity of 256 bits.

End-to-end autonomous driving technology has recently become a focal point of research and application in autonomous driving. State-of-the-art (SOTA) methods are often trained and evaluated on the NuScenes dataset. However, the NuScenes dataset, introduced in 2019 for 3D perception tasks, faces several limitations--such as insufficient scale, simple scenes, and homogeneous driving behaviors--that restrict the upper-bound development of end-to-end autonomous driving algorithms. In light of these issues, we propose a novel, large-scale real-world dataset specifically designed for end-to-end autonomous driving tasks, named TAD-E2E, which is 25x larger, 1.7x scene complexity over NuScenes, and features a highly diverse range of driving behaviors. We replicated SOTA methods on the TAD-E2E dataset and observed that these methods no longer performed well, as expected. Additionally, in response to the challenging scenarios presented in the TAD-E2E dataset, we devised a multimodal sparse end-to-end method that significantly outperforms SOTA methods. Ablation studies demonstrate the effectiveness of our method, and we analyze the contributions of each module. The dataset will be released in the near future.

Humans can infer complete shapes and appearances of objects from limited visual cues, relying on extensive prior knowledge of the physical world. However, completing partially observable objects while ensuring consistency across video frames remains challenging for existing models, especially for unstructured, in-the-wild videos. This paper tackles the task of Video Amodal Completion (VAC), which aims to generate the complete object consistently throughout the video given a visual prompt specifying the object of interest. Leveraging the rich, consistent manifolds learned by pre-trained video diffusion models, we propose a conditional diffusion model, TACO, that repurposes these manifolds for VAC. To enable its effective and robust generalization to challenging in-the-wild scenarios, we curate a large-scale synthetic dataset with multiple difficulty levels by systematically imposing occlusions onto un-occluded videos. Building on this, we devise a progressive fine-tuning paradigm that starts with simpler recovery tasks and gradually advances to more complex ones. We demonstrate TACO's versatility on a wide range of in-the-wild videos from Internet, as well as on diverse, unseen datasets commonly used in autonomous driving, robotic manipulation, and scene understanding. Moreover, we show that TACO can be effectively applied to various downstream tasks like object reconstruction and pose estimation, highlighting its potential to facilitate physical world understanding and reasoning.

Multi-head self-attention (MHSA) is a key component of Transformers, a widely popular architecture in both language and vision. Multiple heads intuitively enable different parallel processes over the same input. Yet, they also obscure the attribution of each input patch to the output of a model. We propose a novel 1-head Transformer Attention Bottleneck (TAB) layer, inserted after the traditional MHSA architecture, to serve as an attention bottleneck for interpretability and intervention. Unlike standard self-attention, TAB constrains the total attention over all patches to \in [0, 1]. That is, when the total attention is 0, no visual information is propagated further into the network, and the vision-language model (VLM) would default to a generic, image-independent response. To demonstrate the advantages of TAB, we train VLMs with TAB to perform image-difference captioning. Over three datasets, our models perform similarly to baseline VLMs in captioning but the bottleneck is superior in localizing changes and in identifying when no changes occur. TAB is the first architecture to enable users to debug by editing attention, which often produces expected outputs by VLMs.

Text-to-Image (T2I) generative models have revolutionized content creation but remain highly sensitive to prompt phrasing, often requiring users to repeatedly refine prompts multiple times without clear feedback. While techniques such as automatic prompt engineering, controlled text embeddings, denoising, and multi-turn generation mitigate these issues, they offer limited controllability, or often necessitate additional training, restricting the generalization abilities. Thus, we introduce T2I-Copilot, a training-free multi-agent system that leverages collaboration between (Multimodal) Large Language Models to automate prompt phrasing, model selection, and iterative refinement. This approach significantly simplifies prompt engineering while enhancing generation quality and text-image alignment compared to direct generation. Specifically, T2I-Copilot consists of three agents: (1) Input Interpreter, which parses the input prompt, resolves ambiguities, and generates a standardized report; (2) Generation Engine, which selects the appropriate model from different types of T2I models and organizes visual and textual prompts to initiate generation; and (3) Quality Evaluator, which assesses aesthetic quality and text-image alignment, providing scores and feedback for potential regeneration. T2I-Copilot can operate fully autonomously while also supporting human-in-the-loop intervention for fine-grained control. On GenAI-Bench, using open-source generation models, T2I-Copilot achieves a VQA score comparable to commercial models RecraftV3 and Imagen 3, surpasses FLUX1.1-pro by 6.17% at only 16.59% of its cost, and outperforms FLUX.1-dev and SD 3.5 Large by 9.11% and 6.36%. Code is released at: https://github.com/SHI-Labs/T2I-Copilot.

We present T2Bs, a framework for generating high-quality, animatable character head morphable models from text by combining static text-to-3D generation with video diffusion. Text-to-3D models produce detailed static geometry but lack motion synthesis, while video diffusion models generate motion with temporal and multi-view geometric inconsistencies. T2Bs bridges this gap by leveraging deformable 3D Gaussian splatting to align static 3D assets with video outputs. By constraining motion with static geometry and employing a view-dependent deformation MLP, T2Bs (i) outperforms existing 4D generation methods in accuracy and expressiveness while reducing video artifacts and view inconsistencies, and (ii) reconstructs smooth, coherent, fully registered 3D geometries designed to scale for building morphable models with diverse, realistic facial motions. This enables synthesizing expressive, animatable character heads that surpass current 4D generation techniques.

We investigate how to enhance the physical fidelity of video generation models by leveraging synthetic videos generated via standard computer graphics techniques. These rendered videos respect real-world physics -- such as maintaining 3D consistency -- thereby serving as a valuable resource that can potentially improve video generation models. To harness this potential, we propose a solution that curates and integrates synthetic data while introducing a method to transfer its physical realism to the model, minimizing unwanted artifacts. Through experiments on three representative tasks emphasizing physical consistency, we demonstrate its effectiveness in enhancing physical fidelity. While our model still lacks a deep understanding of physics, our work offers one of the first empirical demonstrations that synthetic video enhances physical fidelity in video synthesis.

Pre-trained vision-language models have exhibited remarkable abilities in detecting out-of-distribution (OOD) samples. However, some challenging OOD samples, which lie close to in-distribution (InD) data in image feature space, can still lead to misclassification. The emergence of foundation models like diffusion models and multimodal large language models (MLLMs) offers a potential solution to this issue. In this work, we propose SynOOD, a novel approach that harnesses foundation models to generate synthetic, challenging OOD data for fine-tuning CLIP models, thereby enhancing boundary-level discrimination between InD and OOD samples. Our method uses an iterative in-painting process guided by contextual prompts from MLLMs to produce nuanced, boundary-aligned OOD samples. These samples are refined through noise adjustments based on gradients from OOD scores like the energy score, effectively sampling from the InD/OOD boundary. With these carefully synthesized images, we fine-tune the CLIP image encoder and negative label features derived from the text encoder to strengthen connections between near-boundary OOD samples and a set of negative labels. Finally, SynOOD achieves state-of-the-art performance on the large-scale ImageNet benchmark, with minimal increases in parameters and runtime. Our approach significantly surpasses existing methods, improving AUROC by 2.80% and reducing FPR95 by 11.13%. Codes are available in https://github.com/Jarvisgivemeasuit/SynOOD.

Large-scale multi-modal models have demonstrated remarkable performance across various visual recognition tasks by leveraging extensive paired multi-modal training data. However, in real-world applications, the presence of missing or incomplete modality inputs often leads to significant performance degradation. Recent research has focused on prompt-based strategies to tackle this issue; however, existing methods are hindered by two major limitations: (1) static prompts lack the flexibility to adapt to varying missing-data conditions, and (2) basic prompt-tuning methods struggle to ensure reliable performance when critical modalities are missing. To address these challenges, we propose a novel Synergistic Prompting (SyP) framework for robust visual recognition with missing modalities. The proposed SyP introduces two key innovations: (I) a Dynamic Adapter, which computes adaptive scaling factors to dynamically generate prompts, replacing static parameters for flexible multi-modal adaptation, and (II) a Synergistic Prompting Strategy, which combines static and dynamic prompts to balance information across modalities, ensuring robust reasoning even when key modalities are missing. The proposed SyP achieves significant performance improvements over existing approaches across three widely-used visual recognition datasets, demonstrating robustness under diverse missing rates and conditions. Extensive experiments and ablation studies validate its effectiveness in handling missing modalities, highlighting its superior adaptability and reliability. The source code will be released.

Generalizing neural networks to unseen target domains is a significant challenge in real-world deployments. Test-time training (TTT) addresses this by using an auxiliary self-supervised task to reduce the domain gap caused by distribution shifts between the source and target. However, we find that when models are required to perform multiple tasks under domain shifts, conventional TTT methods suffer from unsynchronized task behavior, where the adaptation steps needed for optimal performance in one task may not align with the requirements of other tasks. To address this, we propose a novel TTT approach called Synchronizing Tasks for Test-time Training (S4T), which enables the concurrent handling of multiple tasks. The core idea behind S4T is that predicting task relations across domain shifts is key to synchronizing tasks during test time. To validate our approach, we apply S4T to conventional multi-task benchmarks, integrating it with traditional TTT protocols. Our empirical results show that S4T outperforms state-of-the-art TTT methods across various benchmarks.

Synchronizing videos captured simultaneously from multiple cameras in the same scene is often easy and typically requires only simple time shifts. However, synchronizing videos from different scenes or, more recently, generative AI videos, poses a far more complex challenge due to diverse subjects, backgrounds, and nonlinear temporal misalignment. We propose Temporal Prototype Learning (TPL), a prototype-based framework that constructs a shared, compact 1D representation from high-dimensional embeddings extracted by any of various pretrained models. TPL robustly aligns videos by learning a unified prototype sequence that anchors key action phases, thereby avoiding exhaustive pairwise matching. Our experiments show that TPL improves synchronization accuracy, efficiency, and robustness across diverse datasets, including fine-grained frame retrieval and phase classification tasks. Importantly, TPL is the first approach to mitigate synchronization issues in multiple generative AI videos depicting the same action. Our code and a new multiple video synchronization dataset are available at https://bgu-cs-vil.github.io/TPL/

Synthesizing realistic human-object interaction motions is a critical problem in VR/AR and human animation. Unlike the commonly studied scenarios involving a single human or hand interacting with one object, we address a more generic multi-body setting with arbitrary numbers of humans, hands, and objects. The high correlations and mutual influences among bodies leads to two major challenges, for which we propose solutions. First, to satisfy the high demands for synchronization of different body motions, we mathematically derive a new set of alignment scores during the training process, and use maximum likelihood sampling on a dynamic graphical model for explicit synchronization during inference. Second, the high-frequency interactions between objects are often overshadowed by the large-scale low-frequency movements. To address this, we introduce frequency decomposition and explicitly represent high-frequency components in the frequency domain. Extensive experiments across five datasets with various multi-body configurations demonstrate the superiority of SyncDiff over existing state-of-the-art motion synthesis methods.

Robustness is significant for generative image watermarking, typically achieved by injecting distortion-invariant watermark features. The leading paradigm, i.e., inversion-based framework, excels against non-geometric distortions but struggles with geometric ones. To address this, we propose SynTag, a synchronization tag injection-based method that enhances geometric robustness in inversion-based schemes. Due to the complexity of geometric distortions, finding universally geometric-invariant features is challenging, and it is not clear whether such an invariant representation exists. Therefore, instead of seeking invariant representations, we embed a sensitive template feature alongside the watermarking features. This template evolves with geometric distortions, allowing us to reconstruct the distortion trajectory for correction before extraction. Focusing on latent diffusion models, we fine-tune the VAE decoder to inject the invisible SynTag feature, pairing it with a prediction network for extraction and correction. Additionally, we introduce a dither compensation mechanism to further improve correction accuracy. SynTag is highly compatible with existing inversion-based methods. Extensive experiments demonstrate a significant boost in geometric distortion robustness while maintaining resilience against non-geometric distortions.

Facial expression datasets remain limited in scale due to privacy concerns, the subjectivity of annotations, and the labor-intensive nature of data collection. This limitation poses a significant challenge for developing modern deep learning-based facial expression analysis models, particularly foundation models, that rely on large-scale data for optimal performance. To tackle the overarching and complex challenge, we introduce SynFER (Synthesis of Facial Expressions with Refined Control), a novel framework for synthesizing facial expression image data based on high-level textual descriptions as well as more fine-grained and precise control through facial action units. To ensure the quality and reliability of the synthetic data, we propose a semantic guidance technique to steer the generation process and a pseudo-label generator to help rectify the facial expression labels for the synthetic images. To demonstrate the generation fidelity and the effectiveness of the synthetic data from SynFER, we conduct extensive experiments on representation learning using both synthetic data and real-world data. Results validate the efficacy of our approach and the synthetic data. Notably, our approach achieves a 67.23% classification accuracy on AffectNet when training solely with synthetic data equivalent to the AffectNet training set size, which increases to 69.84% when scaling up to five times the original size.