ECCV 2026 论文列表

Dataset distillation provides an effective approach to reducememory and computational costs by optimizing a compact dataset thatachieves performance comparable to the full original. However, for large-scale datasets and complex deep networks (e.g., ImageNet-1K with ResNet-101), the vast optimization space hinders distillation effectiveness, limit-ing practical applications. Recent methods leverage pre-trained diffusionmodels to directly generate informative images, thereby bypassing pixel-level optimization and achieving promising results. Nonetheless, theseapproaches often suffer from distribution shifts between the pre-traineddiffusion prior and target datasets, as well as the need for multiple dis-tillation steps under varying settings. To overcome these challenges, wepropose a novel framework that is orthogonal to existing diffusion-baseddistillation techniques by utilizing the diffusion prior for patch selectionrather than generation. Our method predicts noise from the diffusionmodel conditioned on input images and optional text prompts (with orwithout label information), and computes the associated loss for eachimage-patch pair. Based on the loss differences, we identify distinctiveregions within the original images. Furthermore, we apply intra-classclustering and ranking on the selected patches to enforce diversity con-straints. This streamlined pipeline enables a one-step distillation process.Extensive experiments demonstrate that our approach consistently out-performs state-of-the-art methods across various metrics and settings.

Segmentation in 3D microscopy is challenging due to hard (tosegment) patches, noisy labels resulting from the use of semi-automatedlabeling methods, image artifacts, off-target fluorescence, and lack of la-beled data because of high labeling effort. In this work, we introducea novel unified method to tackle these issues simultaneously in 3D mi-croscopy. First, we introduce µ3DINO, a 3D model pretrained on anultra-large multimodal dataset of over 2 million microscopy volumes.We then create µDivSeg, a segmentation pipeline that uses pretrainedweights to detect noisy labels and hard patches to guide and correct seg-mentation training. We evaluate our methods on a toy dataset and 4 real-world datasets from light-sheet and two-photon microscopy with a varietyof markers, in comparison to two state-of-the-art (SOTA) pipelines. Ourmethods outperform SOTA techniques on increasing levels of syntheticlabel perturbations and real-world data with diverse distributions.

Streaming 3D reconstruction maintains a persistent latentstate that is updated online from incoming frames, enabling constant-memory inference. A key failure mode is the state update rule: aggressiveoverwrites forget useful history, while conservative updates fail to tracknew evidence, and both behaviors become unstable beyond the traininghorizon. To address this challenge, we propose FILT3R, a training-freelatent filtering layer that casts recurrent state updates as stochastic stateestimation in token space. FILT3R maintains a per-token variance andcomputes a Kalman-style gain that adaptively balances memory reten-tion against new observations. Process noise – governing how much thelatent state is expected to change between frames – is estimated onlinefrom EMA-normalized temporal drift of candidate tokens. Using exten-sive experiments, we demonstrate that FILT3R yields an interpretable,plug-in update rule that generalizes common overwrite and gating poli-cies as special cases. Specifically, we show that gains shrink in stableregimes as uncertainty contracts with accumulated evidence, and risewhen genuine scene change increases process uncertainty, improving long-horizon stability for depth, pose, and 3D reconstruction, compared to theexisting methods.

Autonomous driving relies on LiDAR simulation for gener-ating realistic 3D point clouds, typically achieved using computation-ally expensive raycasting. We introduce GnoSphere360, a scalable Li-DAR data generation method that eliminates raycasting by projectingtriangular meshes onto a unit sphere and interpolating point sampleswith gnomonic projection. This computational approach bypasses tra-ditional ray-mesh intersection, significantly reducing computation timewhile maintaining compatibility for further downstream physics-basedmodeling. Benchmarking against NVIDIA OptiX and native CARLAengines, GnoSphere360 achieves up to 35× speedup over CARLA forphysical LiDAR models and maintains robust performance against Op-tiX, while preserving similar LiDAR point cloud output. However, forextremely dense LiDAR configurations, GnoSphere360 achieves over 3.6xspeedup against OptiX, showcasing the scalability of our proposed method.Ablation studies demonstrate robustness of GnoSphere360 across vary-ing scenes and sensor models, while preserving data fidelity. Our resultsshow GnoSphere360 is an efficient, accurate, and parallelizable alterna-tive to raycasting and rasterization for autonomous driving simulation.Project website: gnosphere360.github.io

Face super-resolution (FSR) aims to enhance low-resolution facial im-ages into high-resolution versions. Existing FSR methods face significant chal-lenges, including the lack of publicly available and racially diverse datasets,which hinders reproducibility and subgroup-aware benchmarking. Additionally,many methods struggle to balance the restoration of fine-grained local detailswith the preservation of global facial structures, often producing results that areeither geometrically inconsistent or lack realistic details. To address these is-sues, we introduce SFHQFSR, a new FSR dataset offering high-quality, raciallydiverse facial images generated from generative models, addressing challengesin reproducibility, fairness, and dataset construction. Building on this, we pro-pose GLASNet, a hybrid state-space network that integrates global-local adaptivescanning with multi-domain refinement for high-quality facial image reconstruc-tion. GLASNet combines global semantic reasoning with region-specific localmodeling through a Region-Adaptive Scan Module and enhances restoration witha Multi-domain Refiner using channel, spatial, and frequency attention. Extensiveexperiments demonstrate that GLASNet achieves state-of-the-art performance.

End-to-end autonomous driving models based on Vision-Language-Action (VLA) architectures have shown promising results bylearning driving policies through behavior cloning on expert demonstra-tions. However, imitation learning inherently limits the model to repli-cating observed behaviors without exploring diverse driving strategies,leaving it brittle in novel or out-of-distribution scenarios. Reinforcementlearning (RL) offers a natural remedy by enabling policy explorationbeyond the expert distribution. Yet VLA models, typically trained onoffline datasets, lack directly observable state transitions, necessitating alearned world model to anticipate action consequences. In this work, wepropose a unified understanding-and-generation framework that lever-ages world modeling to simultaneously enable meaningful explorationand provide dense supervision. Specifically, we augment trajectory pre-diction with future RGB and depth image generation as dense worldmodeling objectives, requiring the model to learn fine-grained visual andgeometric representations that substantially enrich the planning back-bone. Beyond serving as a supervisory signal, the world model furtheracts as a source of intrinsic reward for policy exploration: its imageprediction uncertainty naturally measures a trajectory’s novelty rela-tive to the training distribution, where high uncertainty indicates out-of-distribution scenarios that, if safe, represent valuable learning op-portunities. We incorporate this exploration signal into a safety-gatedreward and optimize the policy via Group Relative Policy Optimiza-tion (GRPO). Experiments on the NAVSIM and nuScenes benchmarksdemonstrate the effectiveness of our approach, achieving a state-of-the-art PDMS score of 93.7 and an EPDMS of 88.8 on NAVSIM. The codeis available at https://zihaosheng.github.io/ExploreVLA/.

Diffusion models typically suffer from error accumulationduring iterative sampling, commonly referred to as exposure bias. Wereveal systematic frequency-dependent discrepancies between trainingand inference, which can be interpreted as frequency-dependent SNRerror. Crucially, the direction of this mismatch varies across modelsand timesteps, indicating that fixed correction rules do not generalize.We propose Spectral Alignment (SPA), a lightweight, guidance-basedmethod that calibrates the power spectrum of intermediate predictionsto a pre-computed prior. Our approach consists of two stages: (1) of-fline fitting of a parametric spectrum model from training data, and (2)inference-time guidance via efficient FFT-based gradient computation.SPA introduces minimal computational overhead (3-4%) and is com-plementary to Classifier-Free Guidance (CFG). We demonstrate consis-tent improvements across diverse architectures, from pixel-space mod-els (DDPM, ADM) to latent diffusion models (SD2.0, SDXL) and flow-matching models (SD3.5, FLUX).

We propose D2PO (Dynamic Direct Preference Optimiza-tion), a principled framework for optimizing diffusion sampling policieswith respect to timestep schedules and classifier-free guidance (CFG)weights. Our work is motivated by a fundamental limitation of existingstudent-teacher regression frameworks; low-NFE student samplers aretrained to mimic high-NFE teachers, often sacrificing high-frequency tex-ture fidelity while preserving coarse global structures, thereby misalign-ing the sampler with perceptual quality. D2PO addresses this challengeby reformulating sampler optimization as a preference-based alignmentproblem, leveraging the Direct Preference Optimization (DPO) frame-work. To make DPO applicable to diffusion samplers, we model thesampling policy as an energy-based model (EBM), transforming pref-erence comparisons into tractable energy differences. We further intro-duce a novel energy formulation derived directly from the pretrainedscore network, enabling preference evaluation in perturbed spaces thatjointly capture structural consistency and fine-grained details. More-over, we introduce dynamic preferences, where the preferred samplesused for alignment progressively improve as the sampling policies arelearned. This self-improving mechanism replaces rigid static teacher su-pervision with an iterative, preference-guided refinement process, pro-viding progressively stronger alignment signals. Extensive experimentsdemonstrate that D2PO aligns diffusion samplers with perceptual qual-ity more faithfully, unlocking the full potential of high-quality teachersand consistently outperforming conventional regression-based schedulersunder low-NFE constraints.

Deep neural networks often rely on spurious features to makepredictions, which makes them brittle under distribution shift and onsamples where the spurious correlation does not hold (e.g., minority-group examples). Recent studies have shown that, even in such settings,the feature extractor of an Empirical Risk Minimization (ERM)-trainedmodel can learn rich and informative representations, and that muchof the failure may be attributed to the classifier head. In particular,retraining a lightweight head while keeping the backbone frozen cansubstantially improve performance on shifted distributions and minor-ity groups. Motivated by this observation, we propose a bilevel meta-learning method that performs augmentation directly in feature space toimprove spurious correlation handling in the classifier head. Our methodlearns support-side feature edits such that, after a small number of inner-loop updates on the edited features, the classifier achieves lower loss onhard examples and improved worst-group performance. By operating atthe backbone output rather than in pixel space or through end-to-endoptimization, the method is highly efficient and stable, requiring only afew minutes of training on a single GPU. We further validate our methodwith CLIP-based visualizations, showing that the learned feature-spaceupdates induce semantically meaningful shifts aligned with spurious at-tributes. Our code is available at: https://github.com/ArianYp/HSFM.

While 3D Gaussian Splatting (3DGS) has revolutionized 3Dreconstruction and novel-view synthesis, scenarios with limited inputviews often lead to poor reconstruction quality and artifacts in ren-dered novel views. Recent efforts attempt to utilize powerful diffusionpriors, yet they typically process rendered and reference views concate-nated along an additional dimension in a single network. These methodsoverlook an inherent nature that different views should maintain ap-pearance similarity but differ in structure due to view shifts, leadingto blur caused by conflicts between the two properties. In this paper,we propose DualDiff, a novel pipeline that leverages dual diffusion pri-ors with a Structure-Appearance Attention (SAA) module to introduce∗ Equal contributionB Corresponding author: zhangjian.sz@pku.edu.cnreference guidance for refining low-quality novel views rendered fromflawed 3D representations. Specifically, we retain one diffusion branchto focus on extracting structural information from the low-quality novelviews, while introducing another branch to ensure appearance consis-tency with reference views. Furthermore, we present a 3D reconstructionframework named DualDiff3D, which integrates a reliability-enhancedRender-Refine-Optimize (RRO) loop to progressively and robustly incor-porate the refined novel views, yielding more accurate 3DGS. Extensiveexperiments demonstrate that our approach outperforms state-of-the-art methods even in the inference-only setting, with further performancegains achievable through training. Our code and pre-trained weights willbe publicly released.

Distribution Matching Distillation (DMD) facilitates efficientinference by distilling multi-step diffusion models into few-step variants.Concurrently, Reinforcement Learning (RL) has emerged as a vital toolfor aligning generative models with human preferences. While both repre-sent critical post-training stages for large-scale diffusion models, existingstudies typically treat them as independent, sequential processes, leav-ing a systematic framework for their unification largely unexplored. Inthis work, we demonstrate that jointly optimizing these two objectivesyields mutual benefits: RL enables more preference-aware and control-lable distillation rather than uniformly compressing the full data dis-tribution, while DMD serves as an effective regularizer to mitigate re-ward hacking during RL training. Building on these insights, we proposeDMDR, a unified framework that incorporates Reward-Tilted Distribu-tion Matching optimization alongside two dynamic distillation trainingstrategies in the initial stage, followed by the joint DMD and RL opti-mization in the second stage. Extensive experiments demonstrate thatDMDR achieves state-of-the-art visual quality and prompt adherenceamong few-step generation methods, even surpassing the performance ofits multi-step teacher model.

Conventional RGB cameras suffer from lighting artifacts suchas flare, glare, flicker, and overexposure, leading to irrecoverable informa-tion loss that necessitates computational restoration. However, existingapproaches treat these problems in isolation, failing to recover structuraldetails completely obscured by complex spatially discrete image degrada-tions. In this paper, we propose a novel cross-modal restoration paradigmand present DeLux, a modular proof-of-concept pipeline that leveragesneuromorphic event streams as a structural prior to guide the targeteddetection and inpainting of lighting artifacts in RGB video. Validation onsynthetic benchmarks and real-world automotive footage demonstratesthat DeLux effectively suppresses local artifacts and restores affected re-gions. The proposed approach outperforms existing RGB-only baselinesand event-guided HDR models, achieving an average MS-SSIM of over0.99 across all artifact types and demonstrating up to an 88% reductionin artifact severity in real-world automotive footage. The synthetic arti-fact generation tools and curated real-world evaluation datasets are madepublicly available to foster future research on cross-modal restoration.

In autonomous driving perception, the fusion of LiDAR andcamera modalities has become the dominant paradigm for 3D object de-tection. However, current multi-modal frameworks heavily rely on mas-sive visual backbones pretrained on 2D semantic tasks. This relianceintroduces substantial parameter redundancy and a structural misalign-ment, as 2D priors are ill-equipped to handle the extreme sparsity of Li-DAR projections required for Bird’s-Eye-View geometry. To address this,we present DeGuNet, an ultra-compact and plug-and-play image back-bone explicitly designed for depth-guided representation learning. By in-corporating sparsity-aware feature extraction mechanisms, DeGuNet ef-fectively aligns multi-view images with unstructured LiDAR depth whilestrictly preventing invalid-region contamination. Extensive experimentson the nuScenes dataset demonstrate DeGuNet’s broad plug-and-playapplicability and superior efficiency. When integrated into establishedbaselines, it fundamentally eliminates architectural redundancy, reduc-ing GPU memory consumption by up to 66.5% and achieving a 1.16×inference speedup. Concurrently, DeGuNet delivers up to a 6.20 absolutemAP gain, establishing a new paradigm for parameter-efficient multi-modal 3D perception.

Omnimodal large language models (OmniLLMs) jointly process audio and visual streams, but the resulting long multimodal token sequences make inference prohibitively expensive. Existing compression methods typically rely on fixed window partitioning and attention-based pruning, which overlook the piecewise semantic structure of audio-visual signals and become fragile under aggressive token reduction. We propose Dynamic Audio-driven Semantic cHunking (DASH), a training-free framework that aligns token compression with semantic structure. DASH treats audio embeddings as a semantic anchor and detects boundary candidates via cosine-similarity discontinuities, inducing dynamic, variablelength segments that approximate the underlying piecewise-coherent organization of the sequence. These boundaries are projected onto video tokens as a soft temporally co-registered segmentation prior. Within each segment, token retention is determined by a tri-signal importance estimator that fuses structural boundary cues, representational distinctiveness, and attention-based salience, mitigating the sparsity bias of attentiononly selection. This structure-aware allocation preserves transition-critical tokens while reducing redundant regions. Extensive experiments on AVUT, VideoMME, and WorldSense demonstrate that DASH maintains competitive or superior accuracy while achieving higher compression ratios compared to prior methods. Code is available at: https://github.com/ laychou666/DASH.

Although highly effective in vision and language domains, applying in-context learning to robotics remains challenging. Existing autoregressive in-context imitation methods discretize continuous actions and exacerbate the accumulation of early prediction errors through next-token prediction, limiting their generalization on unseen task configurations. Meanwhile, flow-matching policies have been explored for continuous robot control and can help mitigate compounding errors; however, in-context imitation learning within a flow-matching framework remains underexplored. To address these limitations, we introduce ContextFlow, a conditional flow-matching model that learns continuous action distributions for in-context imitation learning. ContextFlow conditions flow-based action prediction on demonstrations and observations, enabling robust generation from noisy action distributions. To better encode multimodal in-context demonstrations, we adapt perceiver-style multimodal context compressors that distill visual, proprioceptive, and action sequences into compact, task-relevant latent representations. On LIBERO, ContextFlow outperforms ICRT by 35 percentage points in average success rate on unseen task configurations, while matching the performance of the task-specific fine-tuned VLA model π with- 0 out any fine-tuning on unseen tasks. On real robots, it generalizes to unseen configurations of both single-arm and bimanual tasks, achieving 40% success on a new pen-uncapping configuration. Project Page: https://dingjiansw101.github.io/contextflow-page/.

Multimodal Large Language Models (MLLMs) have demon-strated remarkable potential in medical image analysis. However, theirapplication in gastrointestinal endoscopy is currently hindered by twocritical limitations: the misalignment between general model reasoningand standardized clinical cognitive pathways, and the lack of causal as-sociation between visual features and diagnostic outcomes. In this paper,we propose a novel Clinical-Cognitive-Aligned (CogAlign) framework toaddress these challenges. First, we endow the model with rigorous clinicalanalytical capabilities by constructing the hierarchical clinical cognitiondataset and employing Supervised Fine-Tuning (SFT). Unlike conven-tional approaches, this strategy internalizes the hierarchical diagnosticlogic of experts, ranging from anatomical localization and morphologicalevaluation to microvascular analysis, directly into the model. Second, toeliminate visual bias, we provide a theoretical analysis demonstratingthat standard supervised tuning inevitably converges to spurious back-ground correlations. Guided by this insight, we propose a counterfactual-driven reinforcement learning strategy to enforce causal rectification. Bygenerating counterfactual normal samples via lesion masking and opti-mizing through clinical-cognition-centric rewards, we constrain the modelto strictly ground its diagnosis in causal lesion features. Extensive exper-iments demonstrate that our approach achieves State-of-the-Art (SoTA)performance across multiple benchmarks, significantly enhancing diag-nostic accuracy in complex clinical scenarios.

Traditional Visual Grounding (VG) predominantly relies ontextual descriptions to localize objects, a paradigm that inherently strug-gles with linguistic ambiguity and often ignores non-verbal deictic cuesprevalent in real-world interactions. In natural egocentric engagements,hand-pointing combined with speech forms the most intuitive referringmechanism. To bridge this gap, we introduce EgoPoint-Ground, thefirst large-scale multimodal dataset dedicated to egocentric deictic visualgrounding. Comprising over 15k interactive samples in complex scenes,the dataset provides rich, multi-grained annotations including hand-targetbounding box pairs and dense semantic captions. We establish a com-prehensive benchmark for hand-pointing referring expression resolution,evaluating a wide spectrum of mainstream Multimodal Large LanguageModels (MLLMs) and state-of-the-art VG architectures. Furthermore, wepropose SV-CoT, a novel baseline framework that reformulates groundingas a structured inference process, synergizing gestural and linguistic cuesthrough a Visual Chain-of-Thought paradigm. Extensive experimentsdemonstrate that SV-CoT achieves an 11.7% absolute improvement overexisting methods, effectively mitigating semantic ambiguity and advanc-ing the capability of agents to comprehend multimodal physical intents.Code: https://github.com/lingli1724/Egopoint

Diffusion models have achieved state-of-the-art performanceacross 2D, video, and 3D generation, but their sampling quality remainsheavily dependent on guidance methods. Existing training-free guidancetechniques construct a weakened auxiliary branch through static per-turbations, yet these methods fundamentally disregard the stage-wisestructure of the denoising process: early timesteps govern magnitude-driven global formation, whereas later timesteps refine directional ge-ometry. We propose a principled training-free diffusion guidance withStage-Aware Noise Decomposition (SAND) that considers this temporalasymmetry by decomposing Gaussian noise into orthogonal magnitude-and direction-oriented components and modulating them with a time-adaptive weighting schedule aligned to diffusion dynamics. The decom-posed perturbations are injected directly into the self-attention queries,keys, and values, forming a dynamically degraded branch that requires noretraining while faithfully mirroring the evolving geometry of the sample.This representation-level perturbation provides a temporally coherentguidance signal that better balances global structure and fine detail. Ex-tensive experiments on Stable Diffusion XL (SDXL) and Stable Diffusion3 (SD3) demonstrate that our method consistently improves global co-herence, sharpens fine-grained texture, and yields superior sample qual-ity in both unconditional and conditional image generation. We furthershow that SAND can be extended to the guidance-distilled image gener-ation model FLUX.1 and the text-to-video model Wan2.2, highlightingits broad applicability.

Recently, visual token pruning has been studied to handlethe vast number of visual tokens in Multimodal Large Language Models.However, we observe that while existing pruning methods perform reliablyon simple visual understanding, they struggle to effectively generalizeto complex visual reasoning tasks, a critical gap underexplored in previ-ous studies. Through a systematic analysis, we identify Relevant VisualInformation Shift (RVIS) during decoding as the primary failure driver.To address this, we propose Decoding-stage Shift-aware Token Pruning(DSTP), a training-free add-on framework that enables existing pruningmethods to align visual tokens with shifting reasoning requirements duringthe decoding stage. Extensive experiments demonstrate that DSTP signifi-cantly mitigates performance degradation of pruning methods in complexreasoning tasks, while consistently yielding performance gains even acrossvisual understanding benchmarks. Furthermore, DSTP demonstrates ef-fectiveness across diverse state-of-the-art architectures, highlighting itsgeneralizability and efficiency with minimal computational overhead. Oursource code is available here.