ECCV 2026 Papers

Collaborative perception extends single-agent perception by enabling multiple vehicles to exchange complementary perceptual information. However, it introduces an inherent trade-off between perception gain and communication overhead, which is particularly severe for 3D semantic occupancy prediction that relies on fine-grained spatial structures. Existing methods typically compress 3D features into 2D, causing severe spatial information loss, or transmit dense 3D representations, hindering real-world deployment. To overcome these limitations, we propose a bandwidth-efficient collaborative Vector Quantization Semantic Occupancy Prediction (VQSOP) framework. VQSOP employs a SparseAware Vector Quantization (SAVQ) mechanism that exploits 3D scene sparsity to compactly encode informative regions, drastically reducing communication overhead while preserving complete geometric context. Furthermore, to enhance structural consistency and feature continuity, we design a Dual-Branch Adaptive Spatial Refinement (ASR) module that dynamically fuses local high-frequency details with broad contextual semantics. Extensive experiments demonstrate that our approach achieves state-of-the-art performance while reducing communication volume by up to 82×.

Multimodal large language models (MLLMs) achieve strongperformance on benchmarks that evaluate text, image, or video un-derstanding separately. However, these settings do not assess a criticalreal-world requirement, which involves retrieving relevant evidence fromlarge, heterogeneous multimodal corpora prior to reasoning. Most ex-isting benchmarks restrict retrieval to small, single-modality candidatesets, substantially simplifying the search space and overstating end-to-end reliability. To address this gap, we introduce MultiHaystack, thefirst benchmark designed to evaluate both retrieval and reasoning underlarge-scale, cross-modal conditions. MultiHaystack comprises over 46,000multimodal retrieval candidates across documents, images, and videos,along with 747 open yet verifiable questions. Each question is groundedin a unique validated evidence item within the retrieval pool, requiringevidence localization across modalities and fine-grained reasoning. In ourstudy, we find that models perform competitively with evidence, but per-formance drops sharply when retrieving evidence from the full corpus.Additionally, even the strongest retriever, E5-V, achieves only 40.8%Recall@1, while state-of-the-art MLLMs such as GPT-5 experience asignificant drop in reasoning accuracy from 80.86% when provided withthe corresponding evidence to 51.4% under top-5 retrieval. These resultsindicate that multimodal retrieval over heterogeneous pools remains aprimary bottleneck for MLLMs, positioning MultiHaystack as a valuabletestbed that highlights underlying limitations obscured by small-scaleevaluations and promotes retrieval-centric advances in multimodal sys-tems. Our code and benchmark are available at link.

Most advances in human mesh recovery (HMR) have focusedon pelvis-centered recovery, overlooking metric 3D localization and detec-tion accuracy in the camera coordinate system – two key factors for real-world applications such as human–robot interaction and social scene un-derstanding. Current evaluation protocols often ignore these aspects, em-phasizing per-person, root-centered recovery rather than camera-spaceperception. As a result, existing approaches rely on fixed camera as-sumptions or handcrafted post-processing, limiting their robustness andpractical deployment. We introduce Multi-HMR 2, a simple yet ro-bust DETR-based framework for Multi-person camera-centric Humandetection, Mesh Recovery, and tracking. Multi-HMR 2 predicts a scene-consistent camera together with human meshes, enabling metric 3D lo-calization without ground-truth intrinsics. Moreover, by distilling image-based memory features from SAM2, Multi-HMR 2 extends to tracking,achieving consistent identity association without video supervision. De-spite its conceptual simplicity – no handcrafted components, no videoinput, and no ground-truth cameras – Multi-HMR 2 achieves state-of-the-art pelvis-centered performance while substantially improving detec-tion accuracy and metric 3D localization.

Scaling has driven major advances in vision foundation mod-els, yet extending it to metric depth estimation remains challenging dueto heterogeneous sensor noise, camera-dependent biases, and metric am-biguity in noisy cross-source 3D data. We introduce MetricAnything,a simple and scalable pretraining framework that learns metric depthfrom diverse, noisy 3D sources without manual prompts, camera-specificmodeling, or task-specific architectures. At its core is the Sparse Met-ric Prompt, formed by randomly masking depth maps, which acts asa universal interface to decouple spatial reasoning from sensor and cam-era biases. Pretrained on ∼20M image–depth pairs from reconstructed,captured, and rendered 3D data across 10,000+ camera models, Met-ricAnything demonstrates—for the first time—a clear scaling trend inmetric depth. The model excels at prompt-driven tasks such as depthcompletion, super-resolution, and Radar-camera fusion, while its distilledprompt-free student achieves state-of-the-art results in monocular depthestimation, camera intrinsics recovery, single/multi-view metric 3D re-construction, and VLA planning. Using its pretrained ViT as a visualencoder also significantly enhances Multimodal Large Language Modelsin spatial intelligence. These results show that metric depth estimationcan follow the same scaling laws as modern foundation models, paving ascalable and efficient path toward real-world metric perception. Projectpage: https://metric-anything.github.io/metric-anything-io/

This study introduces a masked coordinate pooling-basedattention network (MCPNet), which is a novel framework designed toimprove medical landmark detection. Coordinate features are essentialin accurately localizing medical landmarks. Recently, many studies haveused coordinate information in final regression heads; however, its directincorporation into the attention mechanism remains largely unexplored.To address this issue, we propose MCPNet, which directly extracts posi-tional features from CNN feature maps and analyzes them in the coordi-nate domain. Specifically, MCPNet employs learnable masks and coordi-nate maps to effectively encode diverse and richer anatomical positionalcues of the landmarks. Moreover, the proposed mask maps are utilizedwhen transforming the analyzed coordinate features back into the origi-nal CNN feature space, ensuring smooth reconstruction within the net-work. Extensive experiments on four public benchmark datasets demon-strate that the proposed MCPNet outperforms state-of-the-art methods,offering improved accuracy and efficiency.

Flow Matching (FM) has achieved remarkable generativeperformance, yet it suffers from exposure bias due to discrepancies be-tween training and inference. Existing mitigation strategies typicallyrely on static constraints or external heuristics. In this work, we pro-pose that exposure bias itself inherently contains dynamic signals thatcan guide its own rectification. To leverage this, we introduce DEFAR(DirEctional-Frequency Adaptive Rectification). This framework simu-lates the single-step inference process during training to identify exposurebias. It utilizes the directional and frequency adaptive feedback signalswithin bias itself to enhance the bias tolerance of the model. It consistsof two key components: (1) Anti-Drift Rectification (ADR). ADRtreats inference-time drift as a signal to learn the direction to steer devi-ated states back toward the target. ADR endows the model with intrin-sic active self-rectification capabilities; (2) Frequency Compensation(FC). Empirically, we observe that accumulated bias often stems from alack of low-frequency components in high-noise stages and exposure biascarries the missing frequency information. FC leverages the bias itself asa self-feedback weighting factor to reinforce the missing frequency com-ponents. Experiments on CIFAR-10, CelebA-64, and ImageNet-256/512show that DEFAR outperforms prior baselines and further demonstratesfavorable scalability, compatibility, and inference robustness. Code willbe made available in https://github.com/wuliwuliy/DEFAR.

Detecting anomalies in industrial objects is an importanttopic for increasing production efficiency. More complex objects oftenrequire the analysis of several view points, which has led to the fieldof multi-view anomaly detection. We present MATCH , the first multi-view anomaly detection method based on Flow Matching (FM). Withthe ODE formulation of Flow Matching, we can estimate likelihoods andthereby derive an anomaly score to detect anomalies in multi-view imagedata at object, image, and pixel-level. The architectural flexibility of FMmodels allows us to efficiently transform features of different spatial sizesto the normal distribution. We evaluate thoroughly on the already estab-lished Real-IAD data set and are also the first to provide a comprehen-sive evaluation of popular anomaly detection methods for the MANTA-Tiny data set. MATCH achieves state-of-the-art performance in bothanomaly detection and segmentation, all while running on consumer-levelhardware. By omitting the costly divergence term needed for likelihoodestimation, we ensure that MATCH is usable in real-time productionscenarios. Lastly, several ablation studies are conducted to validate themethodological choices. Code: https://github.com/m-kruse98/MATCH

Diffusion-based methods demonstrate significant potentialfor remote sensing image super-resolution at large scaling factors, partic-ularly in reference-based super-resolution (RefSR), where high-resolutionreference images provide critical fine-grained texture priors. However, ex-isting methods often suffer from a trade-off between over-reliance on ref-erence information, which leads to texture artifacts, and under-utilizationof such information, which results in insufficient detail recovery. To ad-dress these issues, we propose DS-DiT, a Decoupled Siamese DiffusionTransformer that decouples the interaction between low-resolution (LR)and reference (Ref) conditions within the attention mechanism. By al-lowing LR structural priors and Ref texture information to independentlyinteract with the noisy latent, the framework effectively mitigates com-petition between the two conditional sources. To further compensatefor the limited local modeling ability of global attention, we introduce aPatch-Level Weighting (PLW) module that adaptively modulates the fu-sion of conditional sources. In addition, the siamese architecture enablesan inference-time autoguidance strategy that exploits the prediction dis-crepancy between strong and weak Ref conditions to improve generationquality without additional training. Experimental results across multi-ple datasets and scaling factors show that DS-DiT outperforms existingmethods in both quantitative metrics and visual fidelity. The source codeis available at https://github.com/B1nary-L/DS-DiT.

Metric scale monocular geometry estimation has seen signif-icant progress through large-scale data aggregation, yet current founda-tion models suffer from a persistent “scale-collapse” phenomenon: dis-tant landmarks and vast landscapes are metrically underestimated. Thisperformance gap stems from a training data bottleneck, where exist-ing metric-scale datasets are hardware-constrained to unvaried street-level LiDAR or short-range indoor scans, or consist of synthetic datathat lacks the semantic complexity of the physical world. To bridge thisgap, we curate a new metrically-grounded, in-the-wild dataset that wecall MetricScenes, gathered from a variety of sources including Inter-net photo collections and stereo imagery. We estimate camera poses andinitial depth maps for each scene using off-the-shelf methods, and re-cover absolute scale from geo-tagged metadata as well as known stereocamera baselines. We also improve the quality of depth maps derivedfrom MetricScenes via a new two-stage Poisson completion method. Fine-tuning MoGe-2 on our dataset significantly mitigates scale-collapse andachieves superior metric accuracy in unconstrained, open-domain sceneswhile maintaining state-of-the-art performance on standard benchmarks.Project page: https://metricscenes.github.io/.

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.