ICCV 2025 论文列表

In this paper, we study task-oriented human grasp synthesis, a new grasp synthesis task that demands both task and context awareness. At the core of our method is the task-aware contact maps. Unlike traditional contact maps that only reason about the manipulated object and its relation with the hand, our enhanced maps take into account scene and task information. This comprehensive map is critical for hand-object interaction, enabling accurate grasping poses that align with the task. We propose a two-stage pipeline that first constructs a task-aware contact map informed by the scene and task. In the subsequent stage, we use this contact map to synthesize task-oriented human grasps. We introduce a new dataset and metric for the proposed task to evaluate our approach. Our experiments validate the importance of modeling both scene and task, demonstrating significant improvements over existing methods in both grasp quality and task performance.

Low-light image enhancement (LLIE) is a fundamental task in computer vision. Its goal is to extract more useful information from dark regions. Many existing methods have made excellent strides in improving image brightness and enhancing texture details. However, these approaches often lead to overexposure in certain regions when dealing with unevenly illuminated images, resulting in the loss of original information within the images. To address this issue, we propose a Bezier surface constraint (BSCNet) method based on task decoupling to enhance low-light images with uneven brightness. Specifically, we design a diffusion model with a branch structure that separates the enhancement process into brightness adjustment and color restoration, enabling independent control over brightness uniformity. Additionally, we use Bezier surfaces as a learning target to impose smoothness constraints on the image, thereby addressing the issue of uneven brightness in the enhanced image. To counteract potential detail loss introduced by Bezier surfaces, we introduce a spatial-frequency reconstruction module based on the Fourier transform to enhance fine-grained texture information. Experimental comparisons of six generalized LLIE datasets show that our proposed method has demonstrated outstanding effectiveness.

Federated Continual Learning (FCL) has recently garnered significant attention due to its ability to continuously learn new tasks while protecting user privacy. However, existing Data-Free Knowledge Transfer (DFKT) methods require training the entire model, leading to high training and communication costs, while prompt pool-based methods with access to other task-specific prompts in the pool may pose a privacy leakage risk. To address these challenges, we propose a novel method: Task-aware Prompt gradient Projection and Replay (TPPR), which leverages visual prompts to build a parameter-efficient tuning architecture. Specifically, we propose the Task-Aware Prompt Gradient Projection (TAPGP) mechanism to balance the learning of task-agnostic and task-specific knowledge. In practice, we make the gradient of the deep prompts orthogonal to the virtual data and prompts of preceding tasks, which prevents the erosion of old task knowledge while allowing the model to learn new information. Additionally, we introduce Dual-Level Prompt Replay (DLPR) based on the exponential moving average to facilitate knowledge review at both inter-task and intra-task levels, effectively inheriting learned knowledge. Extensive experimental results demonstrate that our method effectively reduces model communication overhead and alleviates forgetting while fully protecting privacy. With only 1% of the training parameters, we achieve more than 5% accuracy improvements in all settings than SOTA with the same backbone.

Model merging enables efficient multi-task models by combining task-specific fine-tuned checkpoints. However, storing multiple task-specific checkpoints requires significant memory, limiting scalability and restricting model merging to larger models and diverse tasks. In this paper, we propose quantizing task vectors (i.e., the difference between pre-trained and fine-tuned checkpoints) instead of quantizing fine-tuned checkpoints. We observe that task vectors exhibit a narrow weight range, enabling low-precision quantization (<= 4 bit) within existing task vector merging frameworks. To further mitigate quantization errors within ultra-low bit precision (e.g., 2 bit), we introduce Residual Task Vector Quantization, which decomposes the task vector into a base vector and offset component. We allocate bits based on quantization sensitivity, ensuring precision while minimizing error within a memory budget. Experiments on image classification and dense prediction show our method maintains or improves model merging performance while using only 8% of the memory required for full-precision checkpoints. Our code is available at https://aim-skku.github.io/TVQ/.

Vision-Language Models (VLMs) like CLIP have shown remarkable zero-shot performance by aligning different modalities in the embedding space, enabling diverse applications from image editing to visual question answering (VQA). However, these models often inherit biases from their training data, resulting in performance disparities across specific subpopulations. Traditional debiasing methods for VLMs primarily focus on specific downstream tasks using labeled datasets, which we argue is insufficient given the broad applicability of VLMs. Specifically, these methods struggle with generalizability, transferability, and feasibility due to overfitting, limited task applicability, and regulatory constraints on the use of sensitive data, making them less practical in real-world scenarios. To address these challenges, we propose a novel task-agnostic method for learning debiased image embeddings in VLMs. Our approach does not require expensive annotated datasets or curated prompts for downstream tasks, while still preserving the inherent zero-shot capabilities of these models. Instead, we leverage easily accessible information: 1) a bias text corpus generated by a large language model, and 2) a generic unsupervised vision dataset. Our method disentangles the image embedding into bias and neutral components by applying centered kernel alignment (CKA) regularization to the text-vision representational similarity, using the bias text corpus over the generic vision dataset. Experimental results validate the effectiveness of our approach across multiple tasks, offering a practical and versatile solution to debiasing VLMs.

The real value of knowledge lies not just in its accumulation, but in its potential to be harnessed effectively to conquer the unknown. Although recent multimodal large language models (MLLMs) exhibit impressing multimodal capabilities, they often fail in rarely encountered domain-specific tasks due to limited relevant knowledge. To explore this, we adopt visual game cognition as a testbed and select "Monster Hunter: World" as the target to construct a multimodal knowledge graph (MH-MMKG), which incorporates multi-modalities and intricate entity relations. We also design a series of challenging queries based on MH-MMKG to evaluate the models' ability for complex knowledge retrieval and reasoning. Furthermore, we propose a multi-agent retriever that enables a model to autonomously search relevant knowledge without additional training. Experimental results show that our approach significantly enhances the performance of MLLMs, providing a new perspective on multimodal knowledge-augmented reasoning and laying a solid foundation for future research.

Pansharpening, a pivotal task in remote sensing for fusing high-resolution panchromatic and multispectral imagery, has garnered significant research interest. Recent advancements employing diffusion models based on stochastic differential equations (SDEs) have demonstrated state-of-the-art performance. However, the inherent multi-step sampling process of SDEs imposes substantial computational overhead, hindering practical deployment. While existing methods adopt efficient samplers, knowledge distillation, or retraining to reduce sampling steps (e.g., from 1,000 to fewer steps), such approaches often compromise fusion quality.In this work, we propose the Optimal Transport Flow Matching (OTFM) framework, which integrates the dual formulation of unbalanced optimal transport (UOT) to achieve one-step, high-quality pansharpening. Unlike conventional OT formulations that enforce rigid distribution alignment, UOT relaxes marginal constraints to enhance modeling flexibility, accommodating the intrinsic spectral and spatial disparities in remote sensing data. Furthermore, we incorporate task-specific regularization into the UOT objective, enhancing the robustness of the flow model.The OTFM framework enables simulation-free training and single-step inference while maintaining strict adherence to pansharpening constraints. Experimental evaluations across multiple datasets demonstrate that OTFM matches or exceeds the performance of previous regression-based models and leading diffusion-based methods while only needing one sampling step.

Open-Vocabulary Segmentation (OVS) aims at segmenting images from free-form textual concepts without predefined training classes. While existing vision-language models such as CLIP can generate segmentation masks by leveraging coarse spatial information from Vision Transformers, they face challenges in spatial localization due to their global alignment of image and text features. Conversely, self-supervised visual models like DINO excel in fine-grained visual encoding but lack integration with language. To bridge this gap, we present Talk2DINO, a novel hybrid approach that combines the spatial accuracy of DINOv2 with the language understanding of CLIP. Our approach aligns the textual embeddings of CLIP to the patch-level features of DINOv2 through a learned mapping function without the need to fine-tune the underlying backbones. At training time, we exploit the attention maps of DINOv2 to selectively align local visual patches with textual embeddings. We show that the powerful semantic and localization abilities of Talk2DINO can enhance the segmentation process, resulting in more natural and less noisy segmentations, and that our approach can also effectively distinguish foreground objects from the background. Experimental results demonstrate that Talk2DINO achieves state-of-the-art performance across several unsupervised OVS benchmarks.

Spatial awareness is key to enable embodied multimodal AI systems. Yet, without vast amounts of spatial supervision, current Multimodal Large Language Models (MLLMs) struggle at this task. In this paper, we introduce TWIST & SCOUT, a framework that equips pre-trained MLLMs with visual grounding ability without forgetting their existing image and language understanding skills. To this end, we propose TWIST, a twin-expert stepwise tuning module that modifies the decoder of the language model using one frozen module pre-trained on image understanding tasks and another learnable one for visual grounding tasks. This allows the MLLM to retain previously learned knowledge and skills, while acquiring what is missing. To fine-tune the model effectively, we generate a high-quality synthetic dataset we call SCOUT, which mimics human reasoning in visual grounding. This dataset provides rich supervision signals, describing a step-by-step multimodal reasoning process, thereby simplifying the task of visual grounding. We evaluate our approach on several standard benchmark datasets, encompassing grounded image captioning, zero-shot localization, and visual grounding tasks. Our method consistently delivers strong performance across all tasks, while retaining the pre-trained image understanding capabilities.

Deep Neural Networks (DNNs) have been successfully applied in various computer tasks. However, they remain vulnerable to adversarial attacks, which could lead to severe security risks. In recent years, robust neural architecture search (NAS) has gradually become an emerging direction for designing adversarially robust architectures. However, existing robust NAS methods rely on repeatedly training numerous DNNs to evaluate robustness, which makes the search process extremely expensive. In this paper, we propose a training-free robust NAS method (TRNAS) that significantly reduces search costs. First, we design a zero-cost proxy model (R-Score) that formalizes adversarial robustness evaluation by exploring the theory of DNN's linear activation capability and feature consistency. This proxy only requires initialized weights for evaluation, thereby avoiding expensive adversarial training costs. Secondly, we introduce a multi-objective selection (MOS) strategy to save candidate architectures with robustness and compactness. Experimental results show that TRNAS only requires 0.02 GPU days to find a promising robust architecture in a vast search space including approximately 10^ 20 networks. TRNAS surpasses other state-of-the-art robust NAS methods under both white-box and black-box attacks. Finally, we summarize a few meaningful conclusions for designing the robust architecture and promoting the development of robust NAS field.

Dynamic Scene Graph Generation (DSGG) aims to create a scene graph for each video frame by detecting objects and predicting their relationships. Weakly Supervised DSGG (WS-DSGG) reduces annotation workload by using an un- localized scene graph from a single frame per video for training. Existing WS-DSGG methods depend on an off-the- shelf external object detector to generate pseudo labels for subsequent DSGG training. However, detectors trained on static, object-centric images struggle in dynamic, relation- aware scenarios required for DSGG, leading to inaccurate localization and low-confidence proposals. To address the challenges posed by external object detectors in WS-DSGG, we propose a Temporal-enhanced Relation-aware Knowl- edge Transferring (TRKT) method, which leverages knowl- edge to enhance detection in relation-aware dynamic sce- narios. TRKT is built on two key components: (1) Relation-aware knowledge mining: we first employ object and re- lation class decoders that generate category-specific atten- tion maps to highlight both object regions and interactive areas. Then we propose an Inter-frame Attention Augmen- tation strategy that exploits optical flow for neighboring frames to enhance the attention maps, making them motion- aware and robust to motion blur. This step yields relation- and motion-aware knowledge mining for WS-DSGG. (2) we introduce a Dual-stream Fusion Module that integrates category-specific attention maps into external detections to refine object localization and boost confidence scores for object proposals. Extensive experiments demonstrate that TRKT achieves state-of-the-art performance on Ac- tion Genome dataset. Our code is avaliable at https: //github.com/XZPKU/TRKT.git.

Diffusion models have emerged as the mainstream approach for visual generation. However, these models typically suffer from sample inefficiency and high training costs. Consequently, methods for efficient finetuning, inference and personalization were quickly adopted by the community. However, training these models in the first place still remains very costly. While several recent approaches--including masking, distillation, and architectural modifications--have been proposed to improve training efficiency, each of these methods comes with its own tradeoffs: some achieve enhanced performance at the expense of increased computational cost. In contrast, this work aims to improve training efficiency as well as generative performance at the same time through routes that act as transport mechanism for randomly selected tokens from early layers to deeper layers of the model. Our method is not limited to the common transformer-based model - it can also be applied to state-space models and achieves this without architectural modifications or additional parameters. Finally, we show that TREAD reduces the computational cost and simultaneously boosts model performance on the standard benchmark ImageNet-256 in class-conditional synthesis. Both of these benefits multiply to a convergence speedup of 14x at 400K training iterations compared to DiT and 37x compared to the best benchmark performance of DiT at 7M training iterations. Further, we achieve a competitive FID of 2.09 in a guided and 3.93 in an unguided setting which improves upon the DiT, without architectural changes. We will release our code.

Recent advances in text-to-image diffusion models enable photorealistic image generation, but they also risk producing malicious content, such as NSFW images. To mitigate risk, concept erasure methods are studied to facilitate the model to unlearn specific concepts. However, current studies struggle to fully erase malicious concepts implicitly embedded in prompts (e.g., metaphorical expressions or adversarial prompts) while preserving the model's normal generation capability. To address this challenge, our study proposes TRCE, using a two-stage concept erasure strategy to achieve an effective trade-off between reliable erasure and knowledge preservation. Firstly, TRCE starts by erasing the malicious semantics implicitly embedded in textual prompts. By identifying a critical mapping objective(i.e., the [EoT] embedding), we optimize the cross-attention layers to map malicious prompts to contextually similar prompts but with safe concepts. This step prevents the model from being overly influenced by malicious semantics during the denoising process. Following this, considering the deterministic properties of the sampling trajectory of the diffusion model, TRCE further steers the early denoising prediction toward the safe direction and away from the unsafe one through contrastive learning, thus further avoiding the generation of malicious content. Finally, we conduct comprehensive evaluations of TRCE on multiple malicious concept erasure benchmarks, and the results demonstrate its effectiveness in erasing malicious concepts while better preserving the model's original generation ability.

In this paper, we aim to model 3D scene geometry, appearance, and physical information just from dynamic multi-view videos in the absence of any human labels. By leveraging physics-informed losses as soft constraints or integrating simple physics models into neural networks, existing works often fail to learn complex motion physics, or doing so requires additional labels such as object types or masks. In this paper, we propose a new framework named **TRACE** to model the motion physics of complex dynamic 3D scenes. The key novelty of our approach is that, by formulating each 3D point as a rigid particle with size and orientation in space, we choose to directly learn a translation rotation dynamics system for each particle, explicitly estimating a complete set of physical parameters to govern the particle's motion over time. Extensive experiments on three existing dynamic datasets and one newly created challenging synthetic datasets demonstrate the extraordinary performance of our method over baselines in the task of future frame extrapolation. A nice property of our framework is that multiple objects or parts can be easily segmented just by clustering the learned physical parameters. Our datasets and code will be released at https://github.com/.

Large pre-trained models achieve remarkable performance in vision tasks but are impractical for fine-tuning due to high computational and storage costs. Parameter-Efficient Fine-Tuning (PEFT) methods mitigate this issue by updating only a subset of parameters; however, most existing approaches are task-agnostic, failing to fully exploit task-specific adaptations, which leads to suboptimal efficiency and performance. To address this limitation, we propose Task-Relevant Parameter and Token Selection (TR-PTS), a task-driven framework that enhances both computational efficiency and accuracy. Specifically, we introduce Task-Relevant Parameter Selection, which utilizes the Fisher Information Matrix (FIM) to identify and fine-tune only the most informative parameters in a layer-wise manner, while keeping the remaining parameters frozen. Simultaneously, Task-Relevant Token Selection dynamically preserves the most informative tokens and merges redundant ones, reducing computational overhead. By jointly optimizing parameters and tokens, TR-PTS enables the model to concentrate on task-discriminative information. We evaluate TR-PTS on benchmark, including FGVC and VTAB-1k, where it achieves state-of-the-art performance, surpassing full fine-tuning by 3.40% and 10.35%, respectively. The code are available at https://github.com/synbol/TR-PTS.

X-ray imaging, based on penetration, enables detailed visualization of internal structures. Building on this capability, existing implicit 3D reconstruction methods have adapted the NeRF model and its variants for internal CT reconstruction. However, these approaches often neglect the significance of objects' anatomical priors for implicit learning, limiting both reconstruction precision and learning efficiency, particularly in ultra-sparse view scenarios. To address these challenges, we propose a novel 3D CT reconstruction framework that employs a 'target prior' derived from the object's projection data to enhance implicit learning. Our approach integrates positional and structural encoding to facilitate voxel-wise implicit reconstruction, utilizing the target prior to guide voxel sampling and enrich structural encoding. This dual strategy significantly boosts both learning efficiency and reconstruction quality. Additionally, we introduce a CUDA-based algorithm for rapid estimation of high-quality 3D target priors from sparse-view projections. Experiments utilizing projection data from a complex abdominal dataset demonstrate that the proposed model substantially enhances learning efficiency, outperforming the current leading model, NAF, by a factor of ten. In terms of reconstruction quality, it also exceeds the most accurate model, NeRP, achieving PSNR improvements of 3.57 dB, 5.42 dB, and 5.70 dB with 10, 20, and 30 projections, respectively. The code is available upon request.

Pedestrian trajectory prediction is crucial for many intelligent tasks. While existing methods predict future trajectories from fixed-frame historical observations, they are limited by the observational perspective and the need for extensive historical information, resulting in prediction delays and inflexible generalization in real-time systems. In this paper, we propose a novel task called Transferable Online Pedestrian Trajectory Prediction (TOTP), which synchronously predicts future trajectories with variable observations and enables effective task transfer under different observation constraints. To advance TOTP modeling, we propose a Temporal-Adaptive Mamba Latent Diffusion (TAMLD) model. It utilizes the Social-Implicit Mamba Synthesizer to extract motion states with social interaction and refine temporal representations through Temporal-Aware Distillation. A Trend-Conditional Mamba Decomposer generates the motion latent distribution of the future motion trends and predicts future motion trajectories through sampling decomposition. We utilize Motion-Latent Mamba Diffusion to reconstruct the latent space disturbed by imbalanced temporal noise. Our method achieves state-of-the-art results on multiple datasets and tasks, showcasing temporal adaptability and generalization ability.

We address the problem of video question answering (video QA) with temporal grounding in a weakly supervised setup, without any temporal annotations. Given a video and a question, we generate an open-ended answer grounded with the start and end time. For this task, we propose TOGA: a vision-language model for Temporally Grounded Open-Ended Video QA with Weak Supervision. We instruct-tune TOGA to jointly generate the answer and the temporal grounding. We operate in a weakly supervised setup where the temporal grounding annotations are not available.We generate pseudo labels for temporal grounding and ensure the validity of these labels by imposing a consistency constraint between the question of a grounding response and the response generated by a question referring to the same temporal segment. We notice that jointly generating the answers with the grounding improves performance on question answering as well as grounding.We evaluate TOGA on grounded QA and open-ended QA tasks. For grounded QA, we consider the NExT-GQA benchmark which is designed to evaluate weakly supervised grounded open-ended question answering.For open-ended QA, we consider the MSVD-QA and ActivityNet-QA benchmarks. We achieve state-of-the-art performance for both tasks on these benchmarks.

Video Frame Interpolation (VFI) aims to predict the intermediate frame I_n (we use n to denote time in videos to avoid notation overload with the timestep t in diffusion models) based on two consecutive neighboring frames I_0 and I_1. Recent approaches apply diffusion models (both image-based and video-based) in this task and achieve strong performance. However, image-based diffusion models are unable to extract temporal information and are relatively inefficient compared to non-diffusion methods. Video-based diffusion models can extract temporal information, but they are too large in terms of training scale, model size, and inference time. To mitigate the above issues, we propose Temporal-Aware Latent Brownian Bridge Diffusion for Video Frame Interpolation (TLB-VFI), an efficient video-based diffusion model. By extracting rich temporal information from video inputs through our proposed 3D-wavelet gating and temporal-aware autoencoder, our method achieves 20% improvement in FID on the most challenging datasets over recent SOTA of image-based diffusion models. Meanwhile, due to the existence of rich temporal information, our method achieves strong performance while having 3xfewer parameters. Such a parameter reduction results in 2.3xspeed up. By incorporating optical flow guidance, our method requires 9000xless training data and achieves over 20xfewer parameters than video-based diffusion models.