ICCV 2025 论文列表

Multimodal Diffusion Transformers (MM-DiTs) have achieved remarkable progress in text-driven visual generation. However, even state-of-the-art MM-DiT models like FLUX struggle with achieving precise alignment between text prompts and generated content. We identify two key issues in the attention mechanism of MM-DiT, namely 1) the suppression of cross-modal attention due to token imbalance between visual and textual modalities and 2) the lack of timestep-aware attention weighting, which hinder the alignment. To address these issues, we propose Temperature-Adjusted Cross-modal Attention (TACA), a parameter-efficient method that dynamically rebalances multimodal interactions through temperature scaling and timestep-dependent adjustment. When combined with LoRA fine-tuning, TACA significantly enhances text-image alignment on the T2I-CompBench benchmark with minimal computational overhead. We tested TACA on state-of-the-art models like FLUX and SD3.5, demonstrating its ability to improve text-image alignment in terms of object appearance, attribute binding, and spatial relationships. Our findings highlight the importance of balancing cross-modal attention in improving semantic fidelity in text-to-image diffusion models. Our codes are publicly available at \href https://github.com/Vchitect/TACA https://github.com/Vchitect/TACA .

Bimanual robotic manipulation is an emerging and critical topic in the robotics community. Previous works primarily rely on integrated control models that take the perceptions and states of both arms as inputs to directly predict their actions. However, we think bimanual manipulation involves not only coordinated tasks but also various uncoordinated tasks that do not require explicit cooperation during execution, such as grasping objects with the closest hand, which integrated control frameworks ignore to consider due to their enforced cooperation in the early inputs. In this paper, we propose a novel decoupled interaction framework that considers the characteristics of different tasks in bimanual manipulation. The key insight of our framework is to assign an independent model to each arm to enhance the learning of uncoordinated tasks, while introducing a selective interaction module that adaptively learns weights from its own arm to improve the learning of coordinated tasks. Extensive experiments on seven tasks in the RoboTwin dataset demonstrate that: (1) Our framework achieves outstanding performance, with a 23.5% boost over the SOTA method. (2) Our framework is flexible and can be seamlessly integrated into existing methods. (3) Our framework can be effectively extended to multi-agent manipulation tasks, achieving a 28% boost over the integrated control SOTA. (4) The performance boost stems from the decoupled design itself, surpassing the SOTA by 16.5% in success rate with only 1/6 of the model size.

Recent works favored dense signals (e.g., depth, DensePose), as an alternative to sparse signals (e.g., OpenPose), to provide detailed spatial guidance for pose-guided text-to-image generation. However, dense representations raised new challenges including editing difficulties and potential inconsistencies with textual prompts. This fact motivates us to revisit sparse signals for pose guidance, owing to their simplicity and shape-agnostic nature, which remains underexplored. This paper proposes a novel Spatial-Pose ControlNet (SP-Ctrl), equipping sparse signals with robust controllability for pose-guided image generation. Specifically, we extend OpenPose to a learnable spatial representation, making keypoint embeddings discriminative and expressive. Additionally, we introduce keypoint concept learning, which encourages keypoint tokens to attend to the spatial positions of each keypoint, thus improving pose alignment. Experiments on animal- and human-centric image generation tasks demonstrate that our method outperforms recent spatially controllable text-to-image generation approaches under the sparse-pose guidance and even matches the performance of dense signal-based methods. Moreover, SP-Ctrl shows promising capabilities in diverse and cross-species generation through sparse signals. Code is available at https://github.com/DREAMXFAR/SP-Ctrl.

Learning to forecast trajectories of intelligent agents has caught much more attention recently. However, it remains a challenge to accurately account for agents' intentions and social behaviors when forecasting, and in particular, to simulate the unique randomness within each of those components in an explainable and decoupled way. Inspired by vibration systems and their resonance properties, we propose the Resonance (short for Re) model to encode and forecast pedestrian trajectories in the form of "co-vibrations". It decomposes trajectory modifications and randomness into multiple vibration portions to simulate agents' reactions to each single cause, and forecasts trajectories as the superposition of these independent vibrations separately. Also, benefiting from such vibrations and their spectral properties, representations of social interactions can be learned by emulating the resonance phenomena, further enhancing its explainability. Experiments on multiple datasets have verified its usefulness both quantitatively and qualitatively.

Multi-Task Learning (MTL) enables multiple tasks to be learned within a shared network, but differences in objectives across tasks can cause negative transfer, where the learning of one task degrades another task's performance. While pre-trained transformers significantly improve MTL performance, their fixed network capacity and rigid structure limit adaptability. Previous dynamic network architectures attempt to address this but are inefficient as they directly convert shared parameters into task-specific ones. We propose Dynamic Token Modulation and Expansion (DTME-MTL), a framework applicable to any transformer-based MTL architecture. DTME-MTL enhances adaptability and reduces overfitting by identifying gradient conflicts in token space and applying adaptive solutions based on conflict type. Unlike prior methods that mitigate negative transfer by duplicating network parameters, DTME-MTL operates entirely in token space, enabling efficient adaptation without excessive parameter growth. Extensive experiments demonstrate that DTME-MTL consistently improves multi-task performance with minimal computational overhead, offering a scalable and effective solution for enhancing transformer-based MTL models.

Several video understanding tasks, such as natural language temporal video grounding, temporal activity localization, and audio description generation, require "temporally dense" reasoning over frames sampled at high temporal resolution. However, computing frame-level features for these tasks is computationally expensive given the temporal resolution requirements. In this paper, we make three contributions to reduce the cost of computing features for temporally dense tasks. First, we introduce a vision transformer (ViT) architecture, dubbed ResidualViT, that leverages the large temporal redundancy in videos to efficiently compute temporally dense frame-level features. Our architecture incorporates (i) learnable residual connections that ensure temporal consistency across consecutive frames and (ii) a token reduction module that enhances processing speed by selectively discarding temporally redundant information while reusing weights of a pretrained foundation model. Second, we propose a lightweight distillation strategy to approximate the frame-level features of the original foundation model. Finally, we evaluate our approach across four tasks and five datasets, in both zero-shot and fully supervised settings, demonstrating significant reductions in computational cost (up to 60%) and improvements in inference speed (up to 2.5x faster), all while closely approximating the accuracy of the original foundation model.

Research in quantum machine learning has recently proliferated due to the potential of quantum computing to accelerate machine learning. An area of machine learning that has not yet been explored is neural ordinary differential equation (neural ODE) based residual neural networks (ResNets), which aim to improve the effectiveness of neural networks using the principles of ordinary differential equations. In this work, we present our insights about why analog Rydberg atom quantum computers are especially well-suited for ResNets. We also introduce ResQ, a novel framework to optimize the dynamics of Rydberg atom quantum computers to solve classification problems in machine learning using analog quantum neural ODEs.

Recently, 3D Gaussian Splatting (3D-GS) has prevailed in novel view synthesis, achieving high fidelity and efficiency. However, it often struggles to capture rich details and complete geometry. Our analysis reveals that the 3D-GS densification operation lacks adaptiveness and faces a dilemma between geometry coverage and detail recovery. To address this, we introduce a novel densification operation, residual split, which adds a downscaled Gaussian as a residual. Our approach is capable of adaptively retrieving details and complementing missing geometry. To further support this method, we propose a pipeline named ResGS. Specifically, we integrate a Gaussian image pyramid for progressive supervision and implement a selection scheme that prioritizes the densification of coarse Gaussians over time. Extensive experiments demonstrate that our method achieves SOTA rendering quality. Consistent performance improvements can be achieved by applying our residual split on various 3D-GS variants, underscoring its versatility and potential for broader application in 3D-GS-based applications.

Text-to-image diffusion models have seen significant development recently due to increasing availability of paired 2D data. Although a similar trend is emerging in 3D generation, the limited availability of high-quality 3D data has resulted in less competitive 3D diffusion models compared to their 2D counterparts. In this work, we show how 2D diffusion models, originally trained for text-to-image generation, can be repurposed for 3D object generation. We introduce Gaussian Atlas, a representation of 3D Gaussians with dense 2D grids, which enables the fine-tuning of 2D diffusion models for generating 3D Gaussians. Our approach shows a successful transfer learning from a pretrained 2D diffusion model to 2D manifold flattend from 3D structures. To facilitate model training, a large-scale dataset, Gaussian Atlas, is compiled to comprise 205K high-quality 3D Gaussian fittings of a diverse array of 3D objects. Our experiment results indicate that text-to-image diffusion models can also serve as 3D content generators.

Constructing a compressed latent space through a variational autoencoder (VAE) is the key for efficient 3D diffusion models. This paper introduces COD-VAE that encodes 3D shapes into a COmpact set of 1D latent vectors without sacrificing quality. COD-VAE introduces a two-stage autoencoder scheme to improve compression and decoding efficiency. First, our encoder block progressively compresses point clouds into compact latent vectors via intermediate point patches. Second, our triplane-based decoder reconstructs dense triplanes from latent vectors instead of directly decoding neural fields, significantly reducing computational overhead of neural fields decoding. Finally, we propose uncertainty-guided token pruning, which allocates resources adaptively by skipping computations in simpler regions and improves the decoder efficiency. Experimental results demonstrate that COD-VAE achieves 16x compression compared to the baseline while maintaining quality. This enables 20.8x speedup in generation, highlighting that a large number of latent vectors is not a prerequisite for high-quality reconstruction and generation.

Transformers have demonstrated remarkable success across vision, language, and video. Yet, increasing task complexity has led to larger models and more tokens, raising the quadratic cost of self-attention and the overhead of GPU memory access. To reduce the computation cost of self-attention, prior work has proposed token compression techniques that drop redundant or less informative tokens. Meanwhile, fused attention kernels such as FlashAttention have been developed to alleviate memory overhead by avoiding attention map construction and its associated I/O to HBM. This, however, makes it incompatible with most training-free token compression methods, which rely on attention maps to determine token importance. Here, we propose Representation Shift, a training-free, model-agnostic metric that measures the degree of change in each token's representation. This seamlessly integrates token compression with FlashAttention, without attention maps or retraining. Our method further generalizes beyond Transformers to CNNs and state space models. Extensive experiments show that Representation Shift enables effective token compression compatible with FlashAttention, yielding significant speedups of up to 5.5x and 4.4x in video-text retrieval and video QA, respectively. Code is available at https://github.com/mlvlab/Representation-Shift.

Despite the promise of Multi-Task Learning (MTL) in leveraging complementary knowledge across tasks, existing multi-task optimization (MTO) techniques remain fixated on resolving conflicts through optimizer-centric loss scaling and gradient manipulation, yet fail to deliver consistent gains. In this paper, we argue that the shared representation space, where task interactions naturally occur, offers rich information and potential for operations complementary to existing optimizer designs, especially for facilitating the inter-task complementarity, which is rarely explored in MTO. This intuition leads to Rep-MTL, which exploits the representation-level task saliency to quantify interactions between task-specific optimization and shared representation learning. By steering these saliencies through entropybased penalization and sample-wise cross-task alignment, Rep-MTL aims to mitigate negative transfer by maintaining the effective training of individual tasks instead pure conflict-solving, while explicitly promoting complementary information sharing. Experiments are conducted on four challenging MTL benchmarks covering both task-shift and domain-shift scenarios. The results show that Rep-MTL, even paired with the basic equal weighting (EW) policy, achieves competitive performance gains with favorable efficiency. Beyond standard performance metrics, Power Law (PL) exponent analysis demonstrates Rep-MTL's efficacy in balancing task-specific learning and cross-task sharing.

Reflective flares are common artifacts in photography that degrade image quality, introducing in-focus flares, which appear as bright, regular spot patterns, and out-of-focus flares, which are diffuse and semi-transparent, obscuring the underlying scene. While previous methods have achieved some success in removing in-focus flares, they struggle with the diffuse nature of out-of-focus flares. The lack of an out-of-focus flare dataset has further hindered the development of effective flare removal models. In this work, we construct a large-scale out-of-focus flare dataset generated with Blender. We propose to use a color alignment approach on diffusion models to address the challenges of out-of-focus reflective flare removal. Rather than fully reconstructing flare-affected regions, our method adjusts the color distribution to reduce artifact visibility while preserving image content. Specifically, we apply a differentiable histogram loss on the diffusion model, which is derived from the Earth Mover's Distance (EMD), to effectively align color distributions. The proposed approach, trained exclusively on synthetic data, outperforms existing methods on both synthetic and real-world data, demonstrating improved performance in flare removal.

Cost volumes are used in every modern optical flow estimator, but due to their computational and space complexity, they are often a limiting factor regarding both processing speed and the resolution of input frames. Motivated by our empirical observation that cost volumes lose their importance once all other network parts of, e.g., a RAFT-based pipeline have been sufficiently trained, we introduce a training strategy that allows removing the cost volume from optical flow estimators throughout training. This leads to significantly improved inference speed and reduced memory requirements. Using our training strategy, we create three different models covering different compute budgets. Our most accurate model reaches state-of-the-art accuracy while being 1.2xfaster and having a 6xlower memory footprint than comparable models; our fastest model is capable of processing Full HD frames at 20\,\mathrm FPS using only 500\,\mathrm MB of GPU memory.

Machine unlearning enables the removal of specific data from ML models to uphold the *right to be forgotten*. While approximate unlearning algorithms offer efficient alternatives to full retraining, this work reveals that they fail to adequately protect the privacy of unlearned data. In particular, these algorithms introduce implicit residuals which facilitate privacy attacks targeting at unlearned data. We observe that these residuals persist regardless of model architectures, parameters, and unlearning algorithms, exposing a new attack surface beyond conventional output-based leakage. Based on this insight, we propose the *Reminiscence Attack (ReA)*, which amplifies the correlation between residuals and membership privacy through targeted fine-tuning processes. ReA achieves up to 1.90x and 1.12x higher accuracy than prior attacks when inferring class-wise and sample-wise membership, respectively. To mitigate such residual-induced privacy risk, we develop a dual-phase approximate unlearning framework that first eliminates deep-layer unlearned data traces and then enforces convergence stability to prevent models from "pseudo-convergence", where their outputs are similar to retrained models but still preserve unlearned residuals. Our framework works for both classification and generation tasks. Experimental evaluations confirm that our approach maintains high unlearning efficacy, while reducing the adaptive privacy attack accuracy to nearly random guess, at the computational cost of 2-12% of full retraining.

We address the challenge of constructing a consistent and photorealistic Neural Radiance Field in inhomogeneously illuminated, scattering environments with unknown, co-moving light sources. While most existing works on underwater scene representation focus on a static homogeneous illumination, limited attention has been paid to scenarios such as when a robot explores water deeper than a few tens of meters, where sunlight becomes insufficient. To address this, we propose a novel illumination field locally attached to the camera, enabling the capture of uneven lighting effects within the viewing frustum. We combine this with a volumetric medium representation to an overall method that effectively handles interaction between dynamic illumination field and static scattering medium. Evaluation results demonstrate the effectiveness and flexibility of our approach.

Modern deep architectures often rely on large-scale datasets, but training on these datasets incurs high computational and storage overhead. Real-world datasets often contain substantial redundancies, prompting the need for more data-efficient training paradigms. Data selection has shown promise to mitigate redundancy by identifying the most representative samples, thereby reducing training costs without compromising performance. Existing methods typically rely on static scoring metrics or pretrained models, overlooking the combined effect of selected samples and their evolving dynamics during training. To address this, we introduce the concept of epsilon-sample cover, which quantifies sample redundancy based on inter-sample relationships, capturing the intrinsic structure of the dataset. Based on this, we reformulate data selection as a reinforcement learning (RL) process, where a lightweight RL agent optimizes the selection policy by leveraging epsilon-sample cover derived from evolving dataset distribution as a reward signal. Extensive experiments across benchmark datasets and diverse architectures demonstrate that our method consistently outperforms existing state-of-the-art baselines. Models trained with our selected datasets show enhanced generalization performance with improved training efficiency.

Affine Grassmannian has been favored for expressing proximity between lines and planes due to its theoretical exactness in measuring distances among features. Despite this advantage, the existing method can only measure the proximity without yielding the distance as an explicit function of rigid body transformation. Thus, an optimizable distance function on the manifold has remained underdeveloped, stifling its application in registration problems. This paper is the first to explicitly derive an optimizable cost function between two Grassmannian features with respect to rigid body transformation (R and t). Specifically, we present a rigorous mathematical proof demonstrating that the bases of high-dimensional linear subspaces can serve as an explicit representation of the cost. Finally, we propose an optimizable cost function based on the transformed bases that can be applied to the registration problem of any affine subspace. Compared to vector parameter-based approaches, our method is able to find a globally optimal solution by directly minimizing the geodesic distance which is agnostic to representation ambiguity. The resulting cost function and its extension to the inlier-set maximizing Branch-and-Bound (BnB) solver have been demonstrated to improve the convergence of existing solutions or outperform them in various computer vision tasks. The code is available on https://github.com/joomeok/GrassmannRegistration.

Learning visual representations is foundational for a broad spectrum of downstream tasks. Although recent vision-language contrastive models, such as CLIP and SigLIP, have achieved impressive zero-shot performance via large-scale vision-language alignment, their reliance on global representations constrains their effectiveness for dense prediction tasks, such as grounding, OCR, and segmentation. To address this gap, we introduce Region-Aware Cluster Discrimination (RICE), a novel method that enhances region-level visual and OCR capabilities. We first construct a billion-scale candidate region dataset and propose a Region Transformer layer to extract rich regional semantics. We further design a unified region cluster discrimination loss that jointly supports object and OCR learning within a single classification framework, enabling efficient and scalable distributed training on large-scale data. Extensive experiments show that RICE consistently outperforms previous methods on tasks, including segmentation, dense detection, and visual perception for Multimodal Large Language Models (MLLMs). The pre-trained models have been released at https://github.com/deepglint/MVT.