ICCV 2025 论文列表

Recent methods for point cloud surface normal estimation predominantly use the generalized winding number field induced by the normals. Optimizing the field towards satisfying desired properties, such as the input points being on the surface defined by the field, provides a principled way to obtain globally consistent surface normals. However, we show that the existing winding number formulation for point clouds is a poor approximation near the input surface points, diverging as the query point approaches a surface point. This is problematic for methods that rely on the accuracy and stability of this approximation, requiring heuristics to compensate. Instead, we derive a more accurate approximation that is properly bounded and converges to the correct value. We then examine two distinct approaches that optimize for globally consistent normals using point cloud winding numbers. We show how the original unbounded formulation influences key design choices in both methods and demonstrate that substituting our formulation yields substantive improvements with respect to normal estimation and surface reconstruction accuracy.

Recent advances in Latent Video Diffusion Models (LVDMs) have revolutionized video generation by leveraging Video Variational Autoencoders (Video VAEs) to compress intricate video data into a compact latent space. However, as LVDM training scales, the computational overhead of Video VAEs becomes a critical bottleneck, particularly for encoding high-resolution videos. To address this, we propose LeanVAE, a novel and ultra-efficient Video VAE framework that introduce two key innovations: (1) a lightweight architecture based on a Neighborhood-Aware Feedforward (NAF) module and non-overlapping patch operations, drastically reducing computational cost, and (2) the integration of wavelet transforms and compressed sensing techniques to enhance reconstruction quality. Extensive experiments validate LeanVAE's superiority in video reconstruction and generation, particularly in enhancing efficiency over existing Video VAEs. Our model offers up to 50x fewer FLOPs and 44x faster inference speed while maintaining competitive reconstruction quality, providing insights for scalable, efficient video generation. Our models and code are available at https://github.com/westlake-repl/LeanVAE.

Vision Transformers (ViTs) have become a standard architecture in computer vision. However, because of their modeling of long-range dependencies through self-attention mechanisms, the explainability of these models remains a challenge. To address this, we propose LeGrad, an explainability method specifically designed for ViTs. LeGrad computes the gradient with respect to the attention maps of single ViT layers, considering the gradient itself as the explainability signal. We aggregate the signal over all layers, combining the activations of the last as well as intermediate tokens to produce the merged explainability map. This makes LeGrad a conceptually simple and an easy-to-implement method to enhance the transparency of ViTs. We evaluate LeGrad in various setups, including segmentation, perturbation, and open-vocabulary settings, showcasing its improved spatial fidelity and its versatility compared to other SotA explainability methods. A demo and the code is available at https://walidbousselham.com/LeGrad

Masked Autoregressive (MAR) models have emerged as a promising approach in image generation, expected to surpass traditional autoregressive models in computational efficiency by leveraging the capability of parallel decoding. However, their dependence on bidirectional self-attention inherently conflicts with conventional KV caching mechanisms, creating unexpected computational bottlenecks that undermine their expected efficiency. To address this problem, this paper studies the caching mechanism for MAR by leveraging two types of redundancy: Token Redundancy indicates that a large portion of tokens have very similar representations in the adjacent decoding steps, which allows us to first cache them in previous steps and then reuse them in the later steps. Condition Redundancy indicates that the difference between conditional and unconditional output in classifier-free guidance exhibits very similar values in adjacent steps. Based on these two redundancies, we propose LazyMAR, which introduces two caching mechanisms to handle them one by one. LazyMAR is training-free and plug-and-play for all MAR models. Experimental results demonstrate that our method achieves 2.83 times acceleration with almost no drop in generation quality. Our codes will be released in https://github.com/feihongyan1/LazyMAR.

Generating cognitive-aligned layered SVGs remains challenging due to existing methods' tendencies toward either oversimplified single-layer outputs or optimization-induced shape redundancies. We propose LayerTracer, a DiT based framework that bridges this gap by learning designers' layered SVG creation processes from a novel dataset of sequential design operations. Our approach operates in two phases: First, a text-conditioned DiT generates multi-phase rasterized construction blueprints that simulate human design workflows. Second, layer-wise vectorization with path deduplication produces clean, editable SVGs. For image vectorization, we introduce a conditional diffusion mechanism that encodes reference images into latent tokens, guiding hierarchical reconstruction while preserving structural integrity. Extensive experiments show that LayerTracer surpasses optimization-based and neural baselines in generation quality and editability.

We introduce LayerLock, a simple yet effective approach for self-supervised visual representation learning, that gradually transitions throughout training from predicting shallow features to deeper ones through progressive layer freezing. First, we make the observation that during training of video masked-autoencoding (MAE) models, ViT layers converge in the order of their depth: shallower layers converge early, deeper layers converge late. We then show that this observation can be exploited to accelerate standard MAE by progressively freezing the model according to an explicit schedule, throughout training. Furthermore, this same schedule can be used in a simple and scalable approach to latent prediction that does not suffer from "representation collapse". We apply our proposed approach, LayerLock, to both pixel and latent prediction approaches with large models of up to 4B parameters and show improvements on both semantic (action classification) and low level (depth estimation) vision tasks.

Designers craft and edit graphic designs in a layer representation, but layer-based editing becomes impossible once composited into a raster image. In this work, we propose LayerD, a method to decompose raster graphic designs into layers for re-editable creative workflow. LayerD addresses the decomposition task by iteratively extracting unoccluded foreground layers. We propose a simple yet effective refinement approach taking advantage of the assumption that layers often exhibit uniform appearance in graphic designs. As decomposition is ill-posed and the ground-truth layer structure may not be reliable, we develop a quality metric that addresses the difficulty. In experiments, we show that LayerD successfully achieves high-quality decomposition and outperforms baselines. We also demonstrate the use of LayerD with state-of-the-art image generators and layer-based editing. Code and models are publicly available.

Traditional animation production decomposes visual elements into discrete layers to enable independent processing for sketching, refining, coloring, and in-betweening. Existing anime generation video methods typically treat animation as a distinct data domain different from real-world videos, lacking fine-grained control at the layer level. To bridge this gap, we introduce LayerAnimate, a novel video diffusion framework with layer-aware architecture that empowers the manipulation of layers through layer-level controls. The development of a layer-aware framework faces a significant data scarcity challenge due to the commercial sensitivity of professional animation assets. To address the limitation, we propose a data curation pipeline featuring Automated Element Segmentation and Motion-based Hierarchical Merging. Through quantitative and qualitative comparisons and user study, we demonstrate that LayerAnimate outperforms current methods in terms of animation quality, control precision, and usability, making it an effective tool for both professional animators and amateur enthusiasts. This framework opens up new possibilities for layer-level animation applications and creative flexibility. Our code is available at https://layeranimate.github.io.

Storytelling tasks involving generating consistent subjects have gained significant attention recently. However, existing methods, whether training-free or training-based, continue to face challenges in maintaining subject consistency due to the lack of fine-grained guidance and inter-frame interaction. Additionally, the scarcity of high-quality data in this field makes it difficult to precisely control storytelling tasks, including the subject's position, appearance, clothing, expression, and posture, thereby hindering further advancements. In this paper, we demonstrate that layout conditions, such as the subject's position and detailed attributes, effectively facilitate fine-grained interactions between frames. This not only strengthens the consistency of the generated frame sequence but also allows for precise control over the subject's position, appearance, and other key details. Building on this, we introduce an advanced storytelling task: Layout-Toggable Storytelling, which enables precise subject control by incorporating layout conditions. To address the lack of high-quality datasets with layout annotations for this task, we develop Lay2Story-1M, which contains over 1 million 720p and higher-resolution images, processed from approximately 11,300 hours of cartoon videos. Building on Lay2Story-1M, we create Lay2Story-Bench, a benchmark with 3,000 prompts designed to evaluate the performance of different methods on this task. Furthermore, we propose Lay2Story, a robust framework based on the Diffusion Transformers (DiTs) architecture for Layout-Togglable Storytelling tasks. Through both qualitative and quantitative experiments, we find that our method outperforms the previous state-of-the-art (SOTA) techniques, achieving the best results in terms of consistency, semantic correlation, and aesthetic quality.

We present Lay-Your-Scene (shorthand LayouSyn), a novel text-to-layout generation pipeline for natural scenes. Prior scene layout generation methods are either closed-vocabulary or use proprietary large language models for open-vocabulary generation, limiting their modeling capabilities and broader applicability in controllable image generation. In this work, we propose to use lightweight open-source language models to obtain scene elements from text prompts and a novel aspect-aware diffusion Transformer architecture trained in an open-vocabulary manner for conditional layout generation. Extensive experiments demonstrate that LayouSyn outperforms existing methods and achieves state-of-the-art performance on challenging spatial and numerical reasoning benchmarks. Additionally, we present two applications of LayouSyn: First, we show that coarse initialization from large language models can be seamlessly combined with our method to achieve better results. Second, we present a pipeline for adding objects to images, demonstrating the potential of LayouSyn in image editing applications.

We present LawDIS, a language-window-based controllable dichotomous image segmentation (DIS) framework that produces high-quality object masks. Our framework recasts DIS as an image-conditioned mask generation task within a latent diffusion model, enabling seamless integration of user controls. LawDIS is enhanced with macro-to-micro control modes. Specifically, in macro mode, we introduce a language-controlled segmentation strategy (LS) to generate an initial mask based on user-provided language prompts. In micro mode, a window-controlled refinement strategy (WR) allows flexible refinement of user-defined regions (i.e., size-adjustable windows) within the initial mask. Coordinated by a mode switcher, these modes can operate independently or jointly, making the framework well-suited for high-accuracy, personalised applications. Extensive experiments on the DIS5K benchmark reveal that our LawDIS significantly outperforms 11 cutting-edge methods across all metrics. Notably, compared to the second-best model MVANet, we achieve weighted F-measure gains of 4.6% with both the LS and WR strategies and 3.6% gains with only the LS strategy on DIS-TE. Codes will be made available at https://github.com/XinyuYanTJU/LawDIS.

Test-time adaptation with pre-trained vision-language models has gained increasing attention for addressing distribution shifts during testing. Among these approaches, memory-based algorithms stand out due to their training-free nature and ability to leverage historical test data. However, existing test-time adaptation methods are typically designed for a single domain with abundant data. In decentralized settings such as federated learning, applying these methods individually to each client suffers from limited test data, while directly sharing a single global memory via the server prevents proper personalization to each client's unique distribution. To address this, we propose Latte, a novel framework where each client maintains a local memory to store embeddings from its own historical test data and an external memory to store class prototypes from other relevant clients. During communication, each client retrieves prototypes from similar clients under the server's coordination to expand its memory. For local adaptation, Latte utilizes both embedding similarity and uncertainty to enhance model performance. Our theoretical analysis shows that Latte effectively leverages in-distribution clients while remaining robust to out-of-distribution clients. Extensive experiments on domain adaptation and corruption benchmarks validate that Latte achieves superior performance in decentralized settings, while introducing only negligible communication and computation costs. Our code is available at https://github.com/baowenxuan/Latte.

Precise camera pose control is crucial for video generation with diffusion models. Existing methods require fine-tuning with additional datasets containing paired videos and camera pose annotations, which are both data-intensive and computationally costly, and may disrupt the model's distribution learned from the training data. We introduce Latent-Reframe, which enables camera control in a pre-trained video diffusion model without fine-tuning. Unlike existing methods, Latent-Reframe operates during the sampling stage, maintaining efficiency while preserving the distribution learned during pretraining. Our approach reframes the latent code of video frames to align with the input camera trajectory through time-aware point clouds. Latent code inpainting and harmonization then refine the model's latent space, ensuring high-quality video generation. Latent-Reframe can be applied to both DiT- and UNet-based video diffusion models. Experimental results demonstrate that Latent-Reframe can achieve comparable or superior camera control precision and video quality to training-based methods, without the need for fine-tuning on additional datasets.

This paper introduces Swap Forward (SaFa), a modality-agnostic and efficient method to generate seamless and coherent long spectrum and panorama using a latent swap joint diffusion process across multi-views. We first investigate spectrum aliasing problem in spectrum-based audio generation caused by existing joint diffusion methods. Through a comparative analysis of the VAE latent representation of spectra and RGB images, we identify that the failure arises from excessive suppression of high-frequency components due to the step-wise averaging operator. To address this issue, we propose Self-Loop Latent Swap, a frame-level bidirectional swap operator, applied to the overlapping region of adjacent views. Leveraging step-wise differentiated trajectories, this swap operator avoids spectrum distortion and adaptively enhances high-frequency components. Furthermore, to improve global cross-view consistency in non-overlapping regions, we introduce Reference-Guided Latent Swap, a unidirectional latent swap operator that provides a centralized reference trajectory to synchronize subview diffusions. By refining swap timing and intervals, we canachieve a balance between cross-view similarity and diversity in a feed-forward manner. Quantitative and qualitative experiments demonstrate that SaFa significantly outperforms existing joint diffusion methods and even training-based methods in audio generation using both U-Net and DiT models. It also adapts well to panorama generation, achieving comparable performance with a 2 xto 20 xspeedup. The project website is available at https://swapforward.github.io.

Visual grounding tasks, such as referring image segmentation (RIS) and referring expression comprehension (REC), aim to localize a target object based on a given textual description. The target object in an image can be described in multiple ways, reflecting diverse attributes such as color, position, and more. However, most existing methods rely on a single textual input, which captures only a fraction of the rich information available in the visual domain. This mismatch between rich visual details and sparse textual cues can lead to the misidentification of similar objects. To address this, we propose a novel visual grounding framework that leverages multiple latent expressions generated from a single textual input by incorporating complementary visual details absent from the original description. Specifically, we introduce subject distributor and visual concept injector modules to embed both shared-subject and distinct-attributes concepts into the latent representations, thereby capturing unique and target-specific visual cues. We also propose a positive-margin contrastive learning strategy to align all latent expressions with the original text while preserving subtle variations. Experimental results show that our method not only outperforms state-of-the-art RIS and REC approaches on multiple benchmarks but also achieves outstanding performance on the generalized referring expression segmentation (GRES) benchmark.

In spite of the remarkable potential of Latent Diffusion Models (LDMs) in image generation, the desired properties and optimal design of the autoencoders have been underexplored. In this work, we analyze the role of autoencoders in LDMs and identify three key properties: latent smoothness, perceptual compression quality, and reconstruction quality. We demonstrate that existing autoencoders fail to simultaneously satisfy all three properties, and propose Variational Masked AutoEncoders (VMAEs), taking advantage of the hierarchical features maintained by Masked AutoEncoder. We integrate VMAEs into the LDM framework, introducing Latent Diffusion Models with Masked AutoEncoders (LDMAEs). Through comprehensive experiments, we demonstrate significantly enhanced image generation quality and computational efficiency.

For Few-Shot Class-Incremental Learning (FSCIL), direct fine-tuning causes significant parameter shifts, resulting in catastrophic forgetting and increased resource consumption. While, freezing the pre-trained backbone exacerbates the inconsistency between the backbone and the evolving classifier. To overcome these challenges, we introduce a method called Low-Rank updates after knowledge localization (Lark). In the knowledge localization phase, the Fisher Information Matrix is calculated to measure the sensitivity of parameters in different layers to previously acquired knowledge. This phase ultimately identifies the parameters within the model that are most suitable for learning new knowledge. In the subsequent incremental editing phase, a low-rank incremental update strategy is applied. This strategy ensures that the model parameter updates adhere to a Rank-One matrix structure. By doing so, it minimizes alterations to the original parameters, thereby enabling the model to integrate new knowledge while retaining as much of the previous knowledge as possible. Extensive experimental results demonstrate that the Lark method achieves significant performance improvements on the CIFAR100, mini-ImageNet, and CUB200 datasets, surpassing current state-of-the-art methods.

We propose a novel approach for captioning and object grounding in video, where the objects in the caption are grounded in the video via temporally dense bounding boxes. We introduce the following contributions. First, we present a large-scale automatic annotation method that aggregates frame-level captions grounded with bounding boxes into temporally dense and consistent annotations. We apply this approach on the HowTo100M dataset to construct a large-scale pre-training dataset, named HowToGround1M. We also introduce a Grounded Video Caption Generation model, dubbed GROVE, and pre-train the model on HowToGround1M. Second, we introduce iGround-a dataset of 3513 videos with manually annotated captions and dense spatio-temporally grounded bounding boxes. This allows us to measure progress on this challenging problem, as well as to fine-tune our model on this small-scale but high-quality data. Third, we demonstrate that our approach achieves state-of-the-art results on the proposed iGround dataset, as well as on the VidSTG, ActivityNet-Entities, GroundingYouTube, and YouCook-Interactions datasets. Our ablations demonstrate the importance of pre-training on our automatically annotated HowToGround1M dataset followed by fine-tuning on the manually annotated iGround dataset and validate the key technical contributions of our model. The dataset and code are available at https://ekazakos.github.io/grounded_video_caption_generation/.