ICCV 2025 Papers

With advancements in sensor and display technologies, high-resolution imagery is becoming increasingly prevalent in diverse applications. As a result, optical flow estimation needs to adapt to larger image resolutions, where even moderate movements lead to substantial pixel displacements, making long-range motion estimation more critical than ever. However, existing datasets primarily focus on short-range flow in low-resolution settings, limiting the generalization of models to high-resolution scenarios with large displacements. Additionally, there is a lack of suitable datasets for evaluating model capacity in long-range motion estimation, further hindering progress in this area. To address this, we introduce RelayFlow-4K, high-resolution 4K optical flow dataset designed to capture diverse motion patterns, including long-range intermediate frame flows. While such datasets provide valuable training resources, long-range estimation remains challenging due to increased matching ambiguity. Simply incorporating these datasets does not inherently improve performance. To this end, we propose a novel training framework that integrates matching cost distillation and incremental time-step learning to refine cost volume estimation and stabilize training. Additionally, we leverage the distance map, which measures the distance from unmatched regions to their nearest matched pixels, improving occlusion handling. Our approach significantly enhances long-range optical flow estimation in high-resolution settings. Our datasets and code are available at https://github.com/Chohoonhee/RelayFlow-4K.

Information Pursuit (IP) is a recently introduced learning framework to construct classifiers that are interpretable-by-design. Given a set of task-relevant and interpretable data queries, IP selects a small subset of the most informative queries and makes predictions based on the gathered query-answer pairs. However, a key limitation of IP is its dependency on task-relevant interpretable queries, which typically require considerable data annotation and curation efforts. While previous approaches have explored using general-purpose large language models to generate these query sets, they rely on prompt engineering heuristics and often yield suboptimal query sets, resulting in a performance gap between IP and non-interpretable black-box predictors. In this work, we propose parameterizing IP queries as a learnable dictionary defined in the latent space of vision-language models such as CLIP. We formulate an optimization objective to learn IP queries and propose an alternating optimization algorithm that shares appealing connections with classic sparse dictionary learning algorithms. Our learned dictionary outperforms baseline methods based on handcrafted or prompted dictionaries across several image classification benchmarks.

Diffusion-based virtual try-on aims to synthesize a realistic image that seamlessly integrating the specific garment into a target model. The primary challenge lies in effectively guiding the warping process of the latent diffusion model. However, previous methods either lack direct guidance or explicitly warp the garment image, which highly depends on the performance of the warping module. In this paper, we propose FIA-VTON, which leverages the implicit flow feature as guidance by adopting a Flow Infused Attention module on virtual try-on. The dense warp flow map is projected as indirect guidance to enhance the feature map warping in the generation process implicitly, which is less sensitive to the warping estimation accuracy than an explicit warp of the garment image. To further enhance implicit warp guidance, we incorporate high-level spatial attention to complement the dense warp. Experimental results on the VTON-HD and DressCode dataset significantly outperform state-of-the-art methods, demonstrating that FIA-VTON is effective and robust for virtual try-on.

In recent years, neural networks have achieved significant progress in offline image processing. However, in online scenarios, particularly in on-chip implementations, memory usage emerges as a critical bottleneck due to the limited memory resources of integrated image processors. In this study, we focus on reducing the memory footprint of neural networks for on-chip image processing by optimizing network design for efficient memory utilization. Specifically, we consider a typical scenario in which images outputted from an image sensor are processed sequentially using line buffers in a line-by-line manner. This setting necessitates the modeling of both intra-line and inter-line correlations--capturing dependencies among pixels within a single line group and across different line groups, respectively.To model intra-line correlations, we propose a progressive feature enhancement strategy, where line pixels are processed with expanding strip convolutions in multiple stages.For inter-line correlation modeling, we introduce a hierarchical line buffer formulation, where features extracted from previous lines are incrementally reused and compressed across multiple hierarchical levels.Comprehensive experiments on various image processing tasks, including RAW denoising, Gaussian denoising, and super-resolution, demonstrate that the proposed method achieves a superior trade-off between performance and memory efficiency than previous solutions, e.g., up to 1dB PSNR gain in RAW denoising at one-fifth of peak memory usage.

Accelerating diffusion model sampling is crucial for efficient AIGC deployment. While diffusion distillation methods -- based on distribution matching and trajectory matching -- reduce sampling to as few as one step, they fall short on complex tasks like text-to-image generation. Few-step generation offers a better balance between speed and quality, but existing approaches face a persistent trade-off: distribution matching lacks flexibility for multi-step sampling, while trajectory matching often yields suboptimal image quality.To bridge this gap, we propose learning few-step diffusion models by Trajectory Distribution Matching (TDM), a unified distillation paradigm that combines the strengths of distribution and trajectory matching. Our method introduces a data-free score distillation objective, aligning the student's trajectory with the teacher's at the distribution level. Further, we develop a sampling-steps-aware objective that decouples learning targets across different steps, enabling more adjustable sampling.This approach supports both deterministic sampling for superior image quality and flexible multi-step adaptation, achieving state-of-the-art performance with remarkable efficiency. Our model, TDM, outperforms existing methods on various backbones, such as SDXL and PixArt-\alpha, delivering superior quality and significantly reduced training costs.In particular, our method distills PixArt-\alpha into a 4-step generator that outperforms its teacher on real user preference at 1024 resolution. This is accomplished with 500 iterations and 2 A800 hours -- a training cost of approximately 1% of the previous state-of-the-art distillation method. In addition, our proposed TDM can be extended to accelerate text-to-video diffusion. Notably, TDM can outperform its teacher model (CogVideoX-2B) by using only 4 NFE on VBench, improving the total score from 80.91 to 81.65.

Modeling animatable human avatars from videos is a long-standing and challenging problem. While conventional methods require per-instance optimization, recent feed-forward methods have been proposed to generate 3D Gaussians with a learnable network.However, these methods predict independent Gaussians for each frame prediction without fully capturing the relations of Gaussians from different frames, which are hard to be animated by novel poses. To address this, we propose Human Gaussian Graph (HGG) to generate generalizable and animatable Gaussian representations. Specifically, we construct a dual-layer graph to model the relations between predicted Gaussians from multiple frames and SMPL mesh. We design an intra-node operation to aggregate various Gaussian information at different timesteps to benefit from video inputs. Furthermore, we propose an inter-node operation to support message passing between SMPL vertices. In this manner, we leverage the human structure prior to recover generalizable and animatable Gaussian representations.Experimental results on novel view synthesis and novel pose animation demonstrate the efficiency and generalization of our method.

Feature matching plays a fundamental role in many computer vision tasks, yet existing methods rely on scarce and clean multi-view image collections, which constrains their generalization to diverse and challenging scenarios. Moreover, conventional feature encoders are typically trained on single-view 2D images, limiting their capacity to capture 3D-aware correspondences. In this paper, we propose a novel two-stage framework that lifts 2D images to 3D space, named as Lift to Match (L2M), taking full advantage of large-scale and diverse single-view images. To be specific, in the first stage, we learn a 3D-aware feature encoder using a combination of multi-view image synthesis and 3D feature Gaussian representation, which injects 3D geometry knowledge into the encoder. In the second stage, a novel-view rendering strategy, combined with large-scale synthetic data generation from single-view images, is employed to learn a feature decoder for robust feature matching, thus achieving generalization across diverse domains. Extensive experiments demonstrate that our method achieves superior generalization across zero-shot evaluation benchmarks, highlighting the effectiveness of the proposed framework for robust feature matching.

Since acquiring large amounts of realistic blurry-sharp image pairs is difficult and expensive, learning blind image deblurring from unpaired data is a more practical and promising solution. Unfortunately, most existing approaches only use adversarial learning to bridge the gap from blurry domains to sharp domains, ignoring the complex and unpredictable nature of real-world blurry patterns. In this paper, we propose a novel diffusion model (DM)-based framework, dubbed TP-Diff, for image deblurring by learning spatially varying texture prior from unpaired sharp data. In particular, TP-Diff performs DM to generate the prior knowledge used to recover the texture of blurry images. To implement it, we propose a Texture Prior Encoder (TPE) that introduces a memory mechanism to encode the texture prior and thereby provide supervision for the DM training. To fully exploit the generated texture priors, we further present the Texture Transfer Transformer layer (TTformer), in which a novel Filter-Modulated Multi-head Self-Attention (FM-MSA) efficiently removes spatially varying blurring through adaptive filtering. In addition, a wavelet-based adversarial loss is used to preserve high-frequency texture details. Extensive evaluations demonstrate that TP-Diff provides a promising unsupervised deblurring solution and outperforms SOTA methods in six widely-used benchmarks.

High-quality image-text data is critical in enhancing Vision-Language Models (VLMs), but traditional image-based pretraining approaches face limitations. These methods are resource-intensive, relying on curated, high-quality interleaved data that is costly and challenging to collect at scale. Additionally, while such datasets improve static image-text understanding, they fail to develop the temporal and motion comprehension needed for video understanding. To address these gaps, we propose incorporating video pretraining into VLMs to improve the model's ability to capture temporal dynamics and general visual perception, which requires reconciling spatial redundancy with strict temporal causality. Therefore, we propose Causal Hierarchical Aggregation to separate computation-heavy spatial encoding from lightweight temporal propagation and construct hierarchical receptive fields at varying granularities. As we scale video context to more than 100 billion tokens, our method excels in high throughput and state-of-the-art performances on both Image and Video understanding, as shown in Figure 1, providing a scalable solution to enhance multimodal learning in dynamic contexts.

"What I cannot create, I do not understand." Human wisdom reveals that creation is one of the highest forms of learning. For example, Diffusion Models have demonstrated remarkable semantic structural and memory capabilities in image generation, denoising, and restoration, which intuitively benefits representation learning. However, current gait networks rarely embrace this perspective, relying primarily on learning by contrasting gait samples under varying complex conditions, leading to semantic inconsistency and uniformity issues. To address these issues, we propose Origins with generative capabilities whose underlying philosophy is that different entities are generated from a unified template, inherently regularizing gait representations within a consistent and diverse semantic space to capture differences accurately. Admittedly, learning this unified template is exceedingly challenging, as it requires the comprehensiveness of the template to encompass gait representations with various conditions. Inspired by Diffusion Models, Origins diffuses the unified template into timestep templates for gait generative modeling, and meanwhile transfers the unified template for gait representation learning. Especially, gait generative modeling and representation learning serve as a unified framework to end-to-end joint training. Extensive experiments on CASIA-B, CCPG, SUSTech1K, Gait3D, GREW and CCGR-MINI demonstrate that Origins performs representation learning within a unified template, achieving superior performance.

This paper presents an effective approach for learning novel 4D embodied world models, which predict the dynamic evolution of 3D scenes over time in response to an embodied agent's actions, providing both spatial and temporal consistency. We propose to learn a 4D world model by training on RGB-DN (RGB, Depth, and Normal) videos.This not only surpasses traditional 2D models by incorporating detailed shape, configuration, and temporal changes into their predictions, but also allows us to effectively learn accurate inverse dynamic models for an embodied agent. Specifically, we first extend existing robotic manipulation video datasets with depth and normal information leveraging off-the-shelf models. Next, we fine-tune a video generation model on this annotated dataset, which jointly predicts RGB-DN (RGB, Depth, and Normal) for each frame. We then present an algorithm to directly convert generated RGB, Depth, and Normal videos into a high-quality 4D scene of the world. Our method ensures temporal and spatial coherence in 4D scene predictions from embodied scenarios, enables novel view synthesis for embodied environments, and facilitates policy learning that significantly outperforms those derived from prior video-based world models.

Understanding scene contexts is crucial for machines to perform tasks and adapt prior knowledge in unseen or noisy 3D environments. As data-driven learning is intractable to comprehensively encapsulate diverse ranges of layouts and open spaces, we propose teaching machines to identify relational commonalities in 3D spaces. Instead of focusing on point-wise or object-wise representations, we introduce 3D scene analogies, which are smooth maps between 3D scene regions that align spatial relationships. Unlike well-studied single instance-level maps, these scene-level maps smoothly link large scene regions, potentially enabling unique applications in trajectory transfer in AR/VR, long demonstration transfer for imitation learning, and context-aware object rearrangement. To find 3D scene analogies, we propose neural contextual scene maps, which extract descriptor fields summarizing semantic and geometric contexts, and holistically align them in a coarse-to-fine manner for map estimation. This approach reduces reliance on individual feature points, making it robust to input noise or shape variations. Experiments demonstrate the effectiveness of our approach in identifying scene analogies and transferring trajectories or object placements in diverse indoor scenes, indicating its potential for robotics and AR/VR applications. Project page including the code is available through this link: https://82magnolia.github.io/3d_scene_analogies/.

We present a method for learning 3D spatial relationships between object pairs, referred to as object-object spatial relationships (OOR), by leveraging synthetically generated 3D samples from pre-trained 2D diffusion models. We hypothesize that images synthesized by 2D diffusion models inherently capture realistic OOR cues, enabling efficient collection of a 3D dataset to learn OOR for various unbounded object categories. Our approach synthesizes diverse images that capture plausible OOR cues, which we then uplift into 3D samples. Leveraging our diverse collection of 3D samples for the object pairs, we train a score-based OOR diffusion model to learn the distribution of their relative spatial relationships. Additionally, we extend our pairwise OORtomulti-object OOR by enforcing consistency across pairwise relations and preventing object collisions. Extensive experiments demonstrate the robustness of our method across various object-object spatial relationships, along with its applicability to 3D scene arrangement tasks and human motion synthesis using our OOR diffusion model.

Context modeling is essential in learned image compression for accurately estimating the distribution of latents. While recent advanced methods have expanded context modeling capacity, they still struggle to efficiently exploit long-range dependency and diverse context information across different coding steps. In this paper, we introduce a novel Hierarchical Progressive Context Model (HPCM) for more efficient context information acquisition. Specifically, HPCM employs a hierarchical coding schedule to sequentially model the contextual dependencies among latents at multiple scales, which enables more efficient long-range context modeling. Furthermore, we propose a progressive context fusion mechanism that incorporates contextual information from previous coding steps into the current step, effectively exploiting diverse contextual information. Experimental results demonstrate that our method achieves state-of-the-art rate-distortion performance and strikes a better balance between compression performance and computational complexity. The code is available at https://github.com/lyq133/LIC-HPCM.

Automated radiology report generation is essential for improving diagnostic efficiency and reducing the workload of medical professionals. However, existing methods face significant challenges, such as disease class imbalance and insufficient cross-modal fusion. To address these issues, we propose the learnable Retrieval Enhanced Visual-Text Alignment and Fusion (REVTAF) framework, which effectively tackles both class imbalance and visual-text fusion in report generation. REVTAF incorporates two core components: (1) a Learnable Retrieval Enhancer (LRE) that utilizes semantic hierarchies from hyperbolic space and intra-batch context through a ranking-based metric. LRE adaptively retrieves the most relevant reference reports, enhancing image representations, particularly for underrepresented (tail) class inputs; and (2) a fine-grained visual-text alignment and fusion strategy that ensures consistency across multi-source cross-attention maps for precise alignment. This component further employs an optimal transport-based cross-attention mechanism to dynamically integrate task-relevant textual knowledge for improved report generation. By combining adaptive retrieval with multi-source alignment and fusion, REVTAF achieves fine-grained visual-text integration under weak image-report level supervision while effectively mitigating data imbalance issues. The experiments demonstrate that REVTAF outperforms state-of-the-art methods, achieving an average improvement of 7.4% on the MIMIC-CXR dataset and 2.9% on the IU X-Ray dataset. Comparisons with mainstream multimodal LLMs (e.g., GPT-series models), further highlight its superiority in radiology report generation.

In class-imbalanced learning (CIL), post-hoc logit adjustment (LA) effectively mitigates class imbalance by adjusting biased logits according to label frequencies. Given the success of LA in CIL, recent class-imbalanced semi-supervised learning (CISSL) algorithms incorporated LA, leading to improved performance when labeled and unlabeled datasets share the same class distribution. However, a common real-world scenario involves the unknown class distribution of the unlabeled set, which may mismatch that of the labeled set. In this case, LA may result in an inappropriate degree of logit adjustments, potentially degrading classification performance due to its inability to incorporate the unknown class distribution of the unlabeled set. To address this problem, we propose a novel CISSL algorithm named learnable logit adjustment (LLA). Unlike the original LA, LLA learns the appropriate degree of logit adjustment by minimizing the class-averaged loss computed for both the labeled and unlabeled sets. Based on the learned degree, LLA refines the biased pseudo-labels of base semi-supervised learning algorithms and adjusts the biased class predictions on the test set by adjusting the logits. Experimental results on benchmark datasets demonstrate that LLA achieves state-of-the-art performance in CISSL.

Under-display ToF imaging aims to achieve accurate depth sensing through a ToF camera placed beneath a screen panel. However, transparent OLED (TOLED) layers introduce severe degradations--such as signal attenuation, multi-path interference (MPI), and temporal noise--that significantly compromise depth quality. To alleviate this drawback, we propose Learnable Fractional Reaction-Diffusion Dynamics (LFRD^2), a hybrid framework that combines the expressive power of neural networks with the interpretability of physical modeling. Specifically, we implement a time-fractional reaction-diffusion module that enables iterative depth refinement with dynamically generated differential orders, capturing long-term dependencies. In addition, we introduce an efficient continuous convolution operator via coefficient prediction and repeated differentiation to further improve restoration quality. Experiments on four benchmark datasets demonstrate the effectiveness of our approach. The code is publicly available at https://github.com/wudiqx106/LFRD2.

A major challenge in Low-Light Image Enhancement (LLIE) is its ill-posed nature: low-light images often lack sufficient information to align with normal-light ones (e.g., not all training data can be fully fitted to the ground truth). Numerous studies have attempted to bridge the gap between low- and normal-light data by introducing effective additional information, which is called "references" in this paper. However, existing methods overlook the valuable references hidden within the training dataset itself. In this work, we propose a novel LLIE strategy that simultaneously learns image-specific features by neural networks while formulating effective common features from the training data as the reference. These common features are correlated with the samples that are not fully fitted by the LLIE network itself, and they are represented as a set of Learnable Feature Patches and Vectors (LFPVs) in the hidden feature space. LFPVs are updated through two mechanisms: the sample-updater, which extracts useful features from training samples to refine LFPVs, and the mutual-updater, which propagates information across LFPVs to mutually update them. LFPVs can be adaptively aligned with image-specific features via our designed query-and-fusion procedure, boosting the LLIE performance. Our proposed method can be integrated into any LLIE framework, improving both enhancement quality and downstream task performance. Extensive experiments on various benchmarks demonstrate the effectiveness of our approach.