ICCV 2025 论文列表

Flat minima, known to enhance generalization and robustness in supervised learning, remain largely unexplored in generative models. In this work, we systematically investigate the role of loss surface flatness in generative models, both theoretically and empirically, with a particular focus on diffusion models. We establish a theoretical claim that flatter minima improve robustness against perturbations in target prior distributions, leading to benefits such as reduced exposure bias---where errors in noise estimation accumulate over iterations---and significantly improved resilience to model quantization, preserving generative performance even under strong quantization constraints. We further observe that Sharpness-Aware Minimization (SAM), which explicitly controls the degree of flatness, effectively enhances flatness in diffusion models even surpassing the indirectly promoting flatness methods---Input Perturbation (IP) which enforces the Lipschitz condition, ensembling-based approach like Stochastic Weight Averaging (SWA) and Exponential Moving Average (EMA)---are less effective. Through extensive experiments on CIFAR-10, LSUN Tower, and FFHQ, we demonstrate that flat minima in diffusion models indeed improve not only generative performance but also robustness.

Co-speech gestures play a vital role in non-verbal communication. In this paper, we introduce a new framework for co-speech gesture understanding in the wild. Specifically, we propose three new tasks and benchmarks to evaluate a model's capability to comprehend gesture-speech-text associations: (i) gesture based retrieval, (ii) gestured word spotting, and (iii) active speaker detection using gestures. We present a new approach that learns a tri-modal video-gesture-speech-text representation to solve these tasks. By leveraging a combination of global phrase contrastive loss and local gesture-word coupling loss, we demonstrate that a strong gesture representation can be learned in a weakly supervised manner from videos in the wild. Our learned representations outperform previous methods, including large vision-language models (VLMs). Further analysis reveals that speech and text modalities capture distinct gesture related signals, underscoring the advantages of learning a shared tri-modal embedding space. The dataset, model, and code are available at: https://www.robots.ox.ac.uk/ vgg/research/jegal.

Joint Photographic Experts Group (JPEG) achieves data compression by quantizing Discrete Cosine Transform (DCT) coefficients, which inevitably introduces compression artifacts. Most existing JPEG quality enhancement methods operate in the pixel domain, suffering from the high computational costs of decoding. Consequently, direct enhancement of JPEG images in the DCT domain has gained increasing attention. However, current DCT-domain methods often exhibit limited performance. To address this challenge, we identify two critical types of correlations within the DCT coefficients of JPEG images. Building on this insight, we propose an Advanced DCT-domain JPEG Quality Enhancement (AJQE) method that fully exploits these correlations. The AJQE method enables the adaptation of numerous well-established pixel-domain models to the DCT domain, achieving superior performance with reduced computational complexity. Compared to the pixel-domain counterparts, the DCT-domain models derived by our method demonstrate a 0.35 dB improvement in PSNR and a 60.5% increase in enhancement throughput on average.

The rapid advancements in Vision Language Models (VLMs) have prompted the development of multi-modal medical assistant systems. Despite this progress, current models still have inherent probabilistic uncertainties, often producing erroneous or unverified responses--an issue with serious implications in medical applications. Existing methods aim to enhance the performance of Medical Vision Language Model (MedVLM) by adjusting model structure, fine-tuning with high-quality data, or through preference fine-tuning. However, these training-dependent strategies are costly and still lack sufficient alignment with clinical expertise. To address these issues, we propose an expert-in-the-loop framework named Expert-Controlled Classifier-Free Guidance (Expert-CFG) to align MedVLM with clinical expertise without additional training. This framework introduces an uncertainty estimation strategy to identify unreliable outputs. It then retrieves relevant references to assist experts in highlighting key terms and applies classifier-free guidance to refine the token embeddings of MedVLM, ensuring that the adjusted outputs are correct and align with expert highlights. Evaluations across three medical visual question answering benchmarks demonstrate that the proposed Expert-CFG, with 4.2B parameters and limited expert annotations, outperforms state-of-the-art models with 13B parameters. The results demonstrate the feasibility of deploying such a system in resource-limited settings for clinical use.

Despite advances in generic object detection, there remains a performance gap in detecting small objects compared to normal-scale objects. We reveal that conventional object localization methods suffer from gradient instability in small objects due to sharper loss curvature, leading to a convergence challenge. To address the issue, we propose Uncertainty-Aware Gradient Stabilization (UGS), a framework that reformulates object localization as a classification task to stabilize gradients. UGS quantizes continuous labels into interval non-uniform discrete representations. Under a classification-based objective, the localization branch generates bounded and confidence-driven gradients, mitigating instability. Furthermore, UGS integrates an uncertainty minimization (UM) loss that reduces prediction variance and an uncertainty-guided refinement (UR) module that identifies and refines high-uncertainty regions via perturbations. Evaluated on four benchmarks, UGS consistently improves anchor-based, anchor-free, and leading small object detectors. Notably, UGS enhances DINO-5scale by 2.6 AP on VisDrone, surpassing prior state-of-the-art performance.

Reconstructing 3D scenes from a single image is a fundamentally ill-posed task due to the severely under-constrained nature of the problem. Consequently, when the scene is rendered from novel camera views, particularly in unseen regions far away from the input camera, existing single image to 3D reconstruction methods render incoherent and blurry views. In this work, we address these inherent limitations in existing single image-to-3D scene feedforward networks. To alleviate the poor performance due to insufficient information beyond the input image's view, we leverage a strong generative prior in the form of a pre-trained latent video diffusion model, for iterative refinement of a coarse scene represented by optimizable Gaussian parameters. To ensure that the style and texture of the generated images align with that of the input image, we incorporate on-the-fly Fourier-style transfer between the generated images and the input image. Additionally, we design a semantic uncertainty quantification module which calculates the per-pixel entropy and yields uncertainty maps which are used to guide the refinement process from the most confident pixels while discarding the remaining highly uncertain ones. We conduct extensive experiments on real-world scene datasets, including in-domain RealEstate-10K and out-of-domain KITTI-v2, showing that our approach can provide more realistic and high-fidelity novel view synthesis results compared to existing state-of-the-art methods.

Spherical motion is a special case of camera motion where the camera moves on the imaginary surface of a sphere with the optical axis normal to the surface. Common sources of spherical motion are a person capturing a stereo panorama with a phone held in an outstretched hand, or a hemi-spherical camera rig used for multi-view scene capture. However, traditional structure-from-motion pipelines tend to fail on spherical camera motion sequences, especially when the camera is facing outward. Building upon prior work addressing the calibrated case, we explore uncalibrated reconstruction from spherical motion, assuming a fixed but unknown focal length parameter. We show that, although two-view spherical motion is always a critical case, self-calibration is possible from three or more views. Through analysis of the relationship between focal length and spherical relative pose, we devise a global structure-from-motion approach for uncalibrated reconstruction. We demonstrate the effectiveness of our approach on real-world captures in various settings, even when the camera motion deviates from perfect spherical motion. Code and data for our method are available at https://github.com/jonathanventura/spherical-sfm.

Pre-trained vision-language models (VLMs), such as CLIP, have demonstrated impressive zero-shot recognition capability, but still underperform in dense prediction tasks. Self-distillation recently is emerging as a promising approach for fine-tuning VLMs to better adapt to local regions without requiring extensive annotations. However, previous state-of-the-art approaches often suffer from significant `foreground bias', where models tend to wrongly identify background regions as foreground objects. To alleviate this issue, we propose DenseVLM, a framework designed to learn unbiased region-language alignment from powerful pre-trained VLM representations. To alleviate this issue, we propose DenseVLM, a framework designed to learn unbiased region-language alignment from powerful pre-trained VLM representations. DenseVLM leverages the pre-trained VLM to retrieve categories for unlabeled regions and then decouples the interference between foreground and background features. We show that DenseVLM can directly replace the original VLM in open-vocabulary object detection and image segmentation methods, leading to notable performance improvements. Furthermore, it exhibits promising zero-shot scalability when training on more extensive and diverse datasets. Our code is publicly available https://github.com/HVision-NKU/DenseVLM.

Recovering missing modalities in multimodal learning has recently been approached using diffusion models to synthesize absent data conditioned on available modalities. However, existing methods often suffer from modality generation bias: while certain modalities are generated with high fidelity, others--such as video--remain challenging due to intrinsic modality gaps, leading to imbalanced training. To address this issue, we propose MD^2N (Multi-stage Duplex Diffusion Network), a novel framework for unbiased missing-modality recovery. MD^2N introduces a modality transfer module within a duplex diffusion architecture, enabling bidirectional generation between available and missing modalities through three stages: (1) global structure generation, (2) modality transfer, and (3) local cross-modal refinement. By training with duplex diffusion, both available and missing modalities generate each other in an intersecting manner, effectively achieving a balanced generation state.Extensive experiments demonstrate that MD^2N significantly outperforms existing state-of-the-art methods, achieving up to 4% improvement over IMDer on the CMU-MOSEI dataset. Project page: https://crystal-punk.github.io/.

This paper introduces UnZipLoRA, a method for decomposing an image into its constituent subject and style, represented as two distinct LoRAs (Low-Rank Adaptations). Unlike existing personalization techniques that focus on either subject or style in isolation, or require separate training sets for each, UnZipLoRA disentangles these elements from a single image by training both the LoRAs simultaneously. UnZipLoRA ensures that the resulting LoRAs are compatible, i.e., they can be seamlessly combined using direct addition. UnZipLoRA enables independent manipulation and recontextualization of subject and style, including generating variations of each, applying the extracted style to new subjects, and recombining them to reconstruct the original image or create novel variations. To address the challenge of subject and style entanglement, UnZipLoRA employs a novel prompt separation technique, as well as column and block separation strategies to accurately preserve the characteristics of subject and style, and ensure compatibility between the learned LoRAs. Evaluation with human studies and quantitative metrics demonstrates UnZipLoRA's effectiveness compared to other state-of-the-art methods, including DreamBooth-LoRA, Inspiration Tree, and B-LoRA.

Neural Radiance Field (NeRF)-based segmentation methods focus on object semantics and rely solely on RGB data, lacking intrinsic material properties. This limitation restricts accurate material perception, which is crucial for robotics, augmented reality, simulation, and other applications. We introduce UnMix-NeRF, a framework that integrates spectral unmixing into NeRF, enabling joint hyperspectral novel view synthesis and unsupervised material segmentation. Our method models spectral reflectance via diffuse and specular components, where a learned dictionary of global endmembers represents pure material signatures, and per-point abundances capture their distribution. For material segmentation, we use spectral signature predictions along learned endmembers, allowing unsupervised material clustering. Additionally, UnMix-NeRF enables scene editing by modifying learned endmember dictionaries for flexible material-based appearance manipulation. Extensive experiments validate our approach, demonstrating superior spectral reconstruction and material segmentation to existing methods.

Ultra-precision estimation of 6DoF pose is essential in applications such as semiconductor manufacturing and nanoscale manipulation. Conventional vision-based techniques are often hampered by sensitivity to defocus and limited estimation accuracy. In this paper, we propose a novel two-dimensional interpolated Discrete Fourier Transform (2D-IpDFT) method for robust 6DoF pose estimation using periodic patterns. We further develop a mathematical framework that links image parameters--phase and frequency--to 6DoF pose, which is applicable to both orthographic and quasi-orthographic imaging systems. Extensive experiments on a low-cost setup, featuring an industrial camera and an etched checkerboard pattern, demonstrate translation estimation accuracy at the nanometer level and rotation estimation accuracy at the microradian level.

In this work, we present Patch-Adapter, an effective framework for high-resolution text-guided image inpainting. Unlike existing methods limited to lower resolutions, our approach achieves 4K+ resolution while maintaining precise content consistency and prompt alignment--two critical challenges in image inpainting that intensify with increasing resolution and texture complexity.Patch-Adapter leverages a two-stage adapter architecture to scale the Diffusion models's resolution from 1K to 4K+ without requiring structural overhauls:(1)Dual Context Adapter: Learns coherence between masked and unmasked regions at reduced resolutions to establish global structural consistency.(2)Reference Patch Adapter: Implements a patch-level attention mechanism for full-resolution inpainting, preserving local detail fidelity through adaptive feature fusion.This dual-stage architecture uniquely addresses the scalability gap in high-resolution inpainting by decoupling global semantics from localized refinement. Experiments demonstrate that Patch-Adapter not only resolves artifacts common in large-scale inpainting but also achieves state-of-the-art performance on the OpenImages and photo-concept-bucket datasets, outperforming existing methods in both perceptual quality and text-prompt adherence. The code is available at: https://github.com/Roveer/Patch-Based-Adapter

Recent studies have highlighted the interplay between diffusion models and representation learning. Intermediate representations from diffusion models can be leveraged for downstream visual tasks, while self-supervised vision models can enhance the convergence and generation quality of diffusion models. However, transferring pretrained weights from vision models to diffusion models is challenging due to input mismatches and the use of latent spaces. To address these challenges, we propose Unified Self-supervised Pretraining (USP), a framework that initializes diffusion models via masked latent modeling in a Variational Autoencoder (VAE) latent space. USP achieves comparable performance in understanding tasks while significantly improving the convergence speed and generation quality of diffusion models. Our code will be publicly available.

Pre-trained point cloud analysis models have shown promising advancements in various downstream tasks, yet their effectiveness is typically suffering from low-quality point cloud (i.e., noise and incompleteness), which is a common issue in real scenarios due to casual object occlusions and unsatisfactory data collected by 3D sensors. To this end, existing methods focus on enhancing point cloud quality by developing dedicated denoising and completion models. However, due to the isolation between the point cloud enhancement and downstream tasks, these methods fail to work in various real-world domains. In addition, the conflicting objectives between denoising and completing tasks further limit the ensemble paradigm to preserve critical geometric features. To tackle the above challenges, we propose a unified point-level prompting method that reformulates point cloud denoising and completion as a prompting mechanism, enabling robust analysis in a parameter-efficient manner. We start by introducing a Rectification Prompter to adapt to noisy points through the predicted rectification vector prompts, effectively filtering noise while preserving intricate geometric features essential for accurate analysis. Sequentially, we further incorporate a Completion Prompter to generate auxiliary point prompts based on the rectified point clouds, facilitating their robustness and adaptability. Finally, a Shape-Aware Unit module is exploited to efficiently unify and capture the filtered geometric features for the downstream point cloud analysis. Extensive experiments on four datasets demonstrate the superiority and robustness of our method when handling noisy and incomplete point cloud data against existing state-of-the-art methods.

Reconstruction from multi-view images is a fundamental challenge in computer vision that has been extensively studied over the past decades. Recently, neural radiance fields have driven significant advancements, especially through methods using implicit functions and volume rendering, achieving high levels of accuracy. A core component of these methods is the mapping that transforms an implicit function's output into corresponding volume densities. Despite its critical role, this mapping has received limited attention in existing literature. In this paper, we provide a comprehensive and systematic study of mapping functions, examining their properties and representations. We first outline the necessary conditions for the mapping function and propose a family of functions that meet these criteria, to ensure first-order unbiasedness. We further demonstrate that the mappings employed by NeuS and VolSDF, two representative neural implicit surface techniques, are special cases within this broader family. Building on our theoretical framework, we introduce several new mapping functions and evaluate their effectiveness through numerical experiments. Our approach offers a fresh perspective on this well-established problem, opening avenues for the development of new techniques in the field.

Visual object tracking has gained promising progress in past decades. Most of the existing approaches focus on learning target representation in well-conditioned daytime data, while for the unconstrained real-world scenarios with adverse weather conditions, e.g. nighttime or foggy environment, the tremendous domain shift leads to significant performance degradation. In this paper, we propose UMDATrack, which is capable of maintaining high-quality target state prediction under various adverse weather conditions within a unified domain adaptation framework. Specifically, we first use a controllable scenario generator to synthesize a small amount of unlabeled videos (less than 2% frames in source daytime datasets) in multiple weather conditions under the guidance of different text prompts. Afterwards, we design a simple yet effective domain-customized adapter (DCA), allowing the target objects' representation to rapidly adapt to various weather conditions without redundant model updating. Furthermore, to enhance the localization consistency between source and target domains, we propose a target-aware confidence alignment module (TCA) following optimal transport theorem. Extensive experiments demonstrate that UMDATrack can surpass existing advanced visual trackers and lead new state-of-the-art performance by a significant margin.