ICCV 2025 论文列表

This paper addresses the challenge of high-fidelity view synthesis of humans with sparse-view videos as input. Previous methods solve the issue of insufficient observation by leveraging 4D diffusion models to generate videos at novel viewpoints. However, the generated videos from these models often lack spatio-temporal consistency, thus degrading view synthesis quality. In this paper, we propose a novel sliding iterative denoising process to enhance the spatio-temporal consistency of the 4D diffusion model. Specifically, we define a latent grid in which each latent encodes the image, camera pose, and human pose for a certain viewpoint and timestamp, then alternately denoise the latent grid along spatial and temporal dimensions with a sliding window, and finally decode the videos at target viewpoints from the corresponding denoised latents. Through the iterative sliding, information flows sufficiently across the latent grid, allowing the diffusion model to obtain a large receptive field and thus enhance the 4D consistency of the output, while making the GPU memory consumption affordable. The experiments on the DNA-Rendering and ActorsHQ datasets demonstrate that our method is able to synthesize high-quality and consistent novel-view videos and significantly outperforms the existing approaches. See our project page for interactive demos and video results: https://diffuman4d.github.io/.

Deep functional maps have recently emerged as a powerful tool for solving non-rigid shape correspondence tasks. Methods that use this approach combine the power and flexibility of the functional map framework, with data-driven learning for improved accuracy and generality. However, most existing methods in this area restrict the learning aspect only to the feature functions and still rely on axiomatic modeling for formulating the training loss or for functional map regularization inside the networks. This limits both the accuracy and the applicability of the resulting approaches only to scenarios where assumptions of the axiomatic models hold. In this work, we show, for the first time, that both in-network regularization and functional map training can be replaced with data-driven methods. For this, we first train a generative model of functional maps in the spectral domain using score-based generative modeling, built from a large collection of high-quality maps. We then exploit the resulting model to promote the structural properties of ground truth functional maps on new shape collections. Remarkably, we demonstrate that the learned models are category-agnostic, and can fully replace commonly used strategies such as enforcing Laplacian commutativity or orthogonality of functional maps. Our key technical contribution is a novel distillation strategy from diffusion models in the spectral domain. Experiments demonstrate that our learned regularization leads to better results than axiomatic approaches for zero-shot non-rigid shape matching. Our code is available at: https://github.com/daidedou/diffumatch/

Generative AI models, particularly diffusion models (DMs), have demonstrated exceptional capabilities in high-quality image synthesis. However, their large memorization capacity raises significant privacy concerns, especially when trained on sensitive datasets. This paper introduces DP-LoRA, a surprisingly simple yet effective framework for differentially private fine-tuning of latent diffusion models (LDMs) using Low-Rank Adaptation (LoRA). By fine-tuning only a small subset of parameters, DP-LoRA achieves state-of-the-art (SoTA) performance in privacy-preserving image generation while significantly improving the privacy-utility trade-off. DP-LoRA leverages pre-trained LDMs and integrates LoRA modules into attention blocks and projection layers, enabling parameter-efficient fine-tuning under Differential Privacy (DP) constraints. Extensive experiments on benchmarks such as CelebA-HQ demonstrate that DP-LoRA outperforms existing methods, achieving competitive Frechet Inception Distance (FID) scores with strict privacy budgets (e.g., epsilon \leq 10). Additionally, we provide a comprehensive analysis of the impact of LoRA rank, noise multiplicity, and trainable components on model performance. Our results highlight the potential of parameter-efficient techniques to scale privacy-preserving generative models to real-world applications, paving the way for safer deployment of diffusion models in sensitive domains. Anonymous codes available at \href https://github.com/EzzzLi/DP-LORA https://github.com/EzzzLi/DP-LORA .

The remarkable success of contrastive-learning-based multimodal models has been greatly driven by training on ever-larger datasets with expensive compute consumption. Sample selection as an alternative efficient paradigm plays an important direction to accelerate the training process. However, recent advances on sample selection either mostly rely on an oracle model to offline select a high-quality coreset, which is limited in the cold-start scenarios, or focus on online selection based on real-time model predictions, which has not sufficiently or efficiently considered the noisy correspondence. To address this dilemma, we propose a novel Differential-Informed Sample Selection (DISSect) method, which accurately and efficiently discriminates the noisy correspondence for training acceleration. Specifically, we rethink the impact of noisy correspondence on contrastive learning and propose that the differential between the predicted correlation of the current model and that of a historical model is more informative to characterize sample quality. Based on this, we construct a robust differential-based sample selection and analyze its theoretical insights. Extensive experiments on three benchmark datasets and various downstream tasks demonstrate the consistent superiority of DISSect over current state-of-the-art methods. Source code is available at: https://github.com/MediaBrain-SJTU/DISSect.

An immersive acoustic experience enabled by spatial audio is just as crucial as the visual aspect in creating realistic virtual environments. However, existing methods for room impulse response estimation rely either on data-demanding learning-based models or computationally expensive physics-based modeling. In this work, we introduce Audio-Visual Differentiable Room Acoustic Rendering (AV-DAR), a framework that leverages visual cues extracted from multi-view images and acoustic beam tracing for physics-based room acoustic rendering. Experiments across six real-world environments from two datasets demonstrate that our multimodal, physics-based approach is efficient, interpretable, and accurate, significantly outperforming a series of prior methods. Notably, on the Real Acoustic Field dataset, AV-DAR achieves comparable performance to models trained on 10 times more data while delivering relative gains ranging from 16.6% to 50.9% when trained at the same scale.

Diffusion models have demonstrated exceptional capabilities in image restoration, yet their application to video super-resolution (VSR) faces significant challenges in balancing fidelity with temporal consistency. Our evaluation reveals a critical gap: existing approaches consistently fail on severely degraded videos--precisely where diffusion models' generative capabilities are most needed. We identify that existing diffusion-based VSR methods struggle primarily because they face an overwhelming learning burden: simultaneously modeling complex degradation distributions, content representations, and temporal relationships with limited high-quality training data. To address this fundamental challenge, we present DiffVSR, featuring a Progressive Learning Strategy (PLS) that systematically decomposes this learning burden through staged training, enabling superior performance on complex degradations. Our framework additionally incorporates an Interweaved Latent Transition (ILT) technique that maintains competitive temporal consistency without additional training overhead. Experiments demonstrate that our approach excels in scenarios where competing methods struggle, particularly on severely degraded videos. Our work reveals that addressing the learning strategy, rather than focusing solely on architectural complexity, is the critical path toward robust real-world video super-resolution with diffusion models.

The image difference captioning (IDC) task is to describe the distinctions between two images. However, existing datasets do not offer comprehensive coverage across all image-difference categories. In this work, we introduce a high-quality dataset, DiffTell with various types of image manipulations, including global image alterations, object-level changes, and text manipulations. The data quality is controlled by careful human filtering. Additionally, to scale up the data collection without prohibitive human labor costs, we explore the possibility of automatically filtering for quality control. We demonstrate that both traditional methods and recent multimodal large language models (MLLMs) exhibit performance improvements on the IDC task after training on the DiffTell dataset. Through extensive ablation studies, we provide a detailed analysis of the performance gains attributed to DiffTell. Experiments show DiffTell significantly enhances the availability of resources for IDC research, offering a more comprehensive foundation and benchmark for future investigations.

Diffusion models have fundamentally transformed the field of generative models, making the assessment of similarity between customized model outputs and reference inputs critically important. However, traditional perceptual similarity metrics operate primarily at the pixel and patch levels, comparing low-level colors and textures but failing to capture mid-level similarities and differences in image layout, object pose, and semantic content. Contrastive learning-based CLIP and self-supervised learning-based DINO are often used to measure semantic similarity, but they highly compress image features, inadequately assessing appearance details. This paper is the first to discover that pretrained diffusion models can be utilized for measuring visual similarity and introduces the DiffSim method, addressing the limitations of traditional metrics in capturing perceptual consistency in custom generation tasks. By aligning features in the attention layers of the denoising U-Net, DiffSim evaluates both appearance and style similarity, showing superior alignment with human visual preferences. Additionally, we introduce the Sref and IP benchmarks to evaluate visual similarity at the level of style and instance, respectively. Comprehensive evaluations across multiple benchmarks demonstrate that DiffSim achieves state-of-the-art performance, providing a robust tool for measuring visual coherence in generative models.

The robustness of 3D object detection in large-scale outdoor point clouds degrades significantly when deployed in an unseen environment due to domain shifts. To minimize the domain gap, existing works on domain adaptive detection focuses on several factors, including point density, object shape and sizes, to reduce the false negative detections. However, the adaptation results indicate that there are still remaining challenges. We argue that this is due to the challenge in recognizing comparably less distinctive region on object surface due to sparsity, occlusion, etc. In this work, we aim to reinforce those features by generating points on object surface to make them straightforwardly recognizable. We draw our motivation from a common observation that detection proposals already contain the accurate bounding boxes, but with relatively low objectness score predictions, which lead to false negatives. Given these box proposals, we densify sparse object points with a diffusion approach. As a result, our model DiffRefine can act as a simple additional module before second-stage refinement, where most existing detection models for two-stage detection can use. Experimental results on domain adaptive detection show competitive performance, especially on vanishing points due to distance on various detection architectures.

Intellectual property (IP) protection for diffusion models is a critical concern, given the significant resources and time required for their development. To effectively safeguard the IP of diffusion models, a key step is enabling the comparison of unique identifiers (fingerprints) between suspect and victim models. However, performing robust and effective fingerprint comparisons among diffusion models remains an under-explored challenge, particularly for diffusion models that have already been released. To address this, in this work, we propose DiffIP, a novel framework for robust and effective fingerprint comparison between suspect and victim diffusion models. Extensive experiments demonstrate the efficacy of our framework.

In spite of recent progress, image diffusion models still produce artifacts. A common solution is to leverage the feedback provided by quality assessment systems or human annotators to optimize the model, where images are generally rated in their entirety. In this work, we believe problem-solving starts with identification, yielding the request that the model should be aware of not just the presence of defects in an image, but their specific locations. Motivated by this, we propose DiffDoctor, a two-stage pipeline to assist image diffusion models in generating fewer artifacts. Concretely, the first stage targets developing a robust artifact detector, for which we collect a dataset of over 1M flawed synthesized images and set up an efficient human-in-the-loop annotation process, incorporating a carefully designed class-balance strategy. The learned artifact detector is then involved in the second stage to optimize the diffusion model by providing pixel-level feedback. Extensive experiments on text-to-image diffusion models demonstrate the effectiveness of our artifact detector as well as the soundness of our diagnose-then-treat design.

Learning cross-modal correspondences is essential for image-to-point cloud (I2P) registration. Existing methods achieve this mostly by utilizing metric learning to enforce feature alignment across modalities, disregarding the inherent modality gap between image and point data. Consequently, this paradigm struggles to ensure accurate cross-modal correspondences. To this end, inspired by the cross-modal generation success of recent large diffusion models, we propose **Diff^2I2P**, a fully **Diff**erentiable **I2P** registration framework, leveraging a novel and effective **Diff**usion prior for bridging the modality gap. Specifically, we propose a Control-Side Score Distillation (CSD) technique to distill knowledge from a depth-conditioned diffusion model to directly optimize the predicted transformation. However, the gradients on the transformation fail to backpropagate onto the cross-modal features due to the non-differentiability of correspondence retrieval and PnP solver. To this end, we further propose a Deformable Correspondence Tuning (DCT) module to estimate the correspondences in a differentiable way, followed by the transformation estimation using a differentiable PnP solver. With these two designs, the Diffusion model serves as a strong prior to guide the cross-modal feature learning of image and point cloud for forming robust correspondences, which significantly improves the registration. Extensive experimental results demonstrate that **Diff^2I2P** consistently outperforms state-of-the-art I2P registration methods, achieving over 7 % improvement in registration recall on the 7-Scenes benchmark. Moreover, **Diff^2I2P** exhibits robust and superior scene-agnostic registration performance.

Recent vision-language models (e.g., CLIP) have demonstrated remarkable class-generalizable ability to unseen classes in few-shot anomaly segmentation (FSAS), leveraging supervised prompt learning or fine-tuning on seen classes. However, their cross-category generalization largely depends on prior knowledge of real seen anomaly samples. In this paper, we propose a novel framework, namely DictAS, which enables a unified model to detect visual anomalies in unseen object categories without any retraining on the target data, only employing a few normal reference images as visual prompts. The insight behind DictAS is to transfer dictionary lookup capabilities to the FSAS task for unseen classes via self-supervised learning, instead of merely memorizing the normal and abnormal feature patterns from the training set. Specifically, DictAS mainly consists of three components: (1) **Dictionary Construction** - to simulate the index and content of a real dictionary using features from normal reference images. (2) **Dictionary Lookup** - to retrieve queried region features from the dictionary via a sparse lookup strategy. When a query feature cannot be retrieved, it is classified as an anomaly. (3) **Query Discrimination Regularization**- to enhance anomaly discrimination by making abnormal features harder to retrieve from the dictionary. To achieve this, Contrastive Query Constraint and Text Alignment Constraint are further proposed. Extensive experiments on seven public industrial and medical datasets demonstrate that DictAS consistently outperforms state-of-the-art FSAS methods.

We introduce DialNav, a novel collaborative embodied dialog task, where a navigation agent (Navigator) and a remote guide (Guide) engage in multi-turn dialog to reach a goal location. Unlike prior work, DialNav aims for holistic evaluation and requires the Guide to infer the Navigator's location, making communication essential for task success. To support this task, we collect and release Remote Assistance in Navigation (RAIN) dataset, human-human dialog paired with navigation trajectories in photorealistic environments. We design a comprehensive benchmark to evaluate both navigation and dialog, and conduct extensive experiments analyzing the impact of different Navigator and Guide models. We highlight key challenges and publicly release the dataset, code, and evaluation framework to foster future research in embodied dialog. Our code and dataset are available at: https://happilee12.github.io/DialNav.

This work questions a common assumption of OOD detection, that models with higher in-distribution (ID) accuracy tend to have better OOD performance. Recent findings show this assumption doesn't always hold. A direct observation is that the later version of torchvision models improves ID accuracy but suffers from a significant drop in OOD performance. We systematically diagnose torchvision training recipes and explain this effect by analyzing the maximal logits of ID and OOD samples. We then propose post-hoc and training-time solutions to mitigate the OOD decrease by fixing problematic augmentations in torchvision recipes. Both solutions enhance OOD detection and maintain strong ID performance.

Masked Diffusion Models (MDMs) have emerged as a powerful generative modeling technique. Despite their remarkable results, they typically suffer from slow inference with several steps. In this paper, we propose Di\mathtt [M] O, a novel approach that distills masked diffusion models into a one-step generator.Di\mathtt [M] O addresses two key challenges: (1) the intractability of using intermediate-step information for one-step generation, which we solve through token-level distribution matching that optimizes model output logits by an `on-policy framework' with the help of an auxiliary model; and (2) the lack of entropy in the initial distribution, which we address through a token initialization strategy that injects randomness while maintaining similarity to teacher training distribution. We show Di\mathtt [M] O's effectiveness on both class-conditional and text-conditional image generation, impressively achieving performance competitive to multi-step teacher outputs while drastically reducing inference time. To our knowledge, we are the first to successfully achieve one-step distillation of masked diffusion models and the first to apply discrete distillation to text-to-image generation, opening new paths for efficient generative modeling.

Generating naturalistic and nuanced listener motions for extended interactions remains an open problem. Existing methods often rely on low-dimensional motion codes for facial behavior generation followed by photorealistic rendering, limiting both visual fidelity and expressive richness. To address these challenges, we introduce DiTaiListener, powered by a video diffusion model with multimodal conditions. Our approach first generates short segments of listener responses conditioned on the speaker's speech and facial motions with DiTaiListener-Gen. It then refines the transitional frames via DiTaiListener-Edit for a seamless transition. Specifically, DiTaiListener-Gen adapts a Diffusion Transformer (DiT) for the task of listener head portrait generation by introducing a Causal Temporal Multimodal Adapter (CTM-Adapter) to process speakers' auditory and visual cues. CTM-Adapter integrates speakers' input in a causal manner into the video generation process to ensure temporally coherent listener responses. For long-form video generation, we introduce DiTaiListener-Edit, a transition refinement video-to-video diffusion model. The model fuses video segments into smooth and continuous videos, ensuring temporal consistency in facial expressions and image quality when merging short video segments produced by DiTaiListener-Gen. Quantitatively, DiTaiListener achieves the state-of-the-art performance on benchmark datasets in both photorealism (+73.8% in FID on RealTalk) and motion representation (+6.1% in FD metric on VICO) spaces. User studies confirm the superior performance of DiTaiListener, with the model being the clear preference in terms of feedback, diversity, and smoothness, outperforming competitors by a significant margin. See the project page for more results: https://ihp-lab.github.io/DiTaiListener/

Text-to-image generation models, especially Multimodal Diffusion Transformers (MMDiT), have shown remarkable progress in generating high-quality images. However, these models often face significant computational bottlenecks, particularly in attention mechanisms, which hinder their scalability and efficiency. In this paper, we introduce DiTFastAttnV2, a post-training compression method designed to accelerate attention in MMDiT. Through an in-depth analysis of MMDiT's attention patterns, we identify key differences from prior DiT-based methods and propose head-wise arrow attention and caching mechanisms to dynamically adjust attention heads, effectively bridging this gap. We also design an Efficient Fused Kernel for further acceleration. By leveraging local metric methods and optimization techniques, our approach significantly reduces the search time for optimal compression schemes to just minutes while maintaining generation quality. Furthermore, with the customized kernel, DiTFastAttnV2 achieves a 68% reduction in attention FLOPs on 2K image generation without compromising visual fidelity.