ECCV 2026 Papers

Generative models based on the Flow Matching objective,particularly Rectified Flow, have emerged as a dominant paradigm forefficient, high-fidelity image synthesis. However, while existing researchheavily prioritizes generation quality and architectural scaling, the un-derlying dynamics of how RF models memorize training data remainlargely underexplored. In this paper, we systematically investigate thememorization behaviors of RF through the test statistics of MembershipInference Attacks (MIA). We progressively formulate three test statistics,culminating in a complexity-calibrated metric (Tmc_cal ) that successfullydecouples intrinsic image spatial complexity from genuine memorizationsignals. This calibration yields a significant performance surge—boostingattack AUC by up to 15% and the privacy-critical TPR@1%FPR metricby up to 45%—establishing the first non-trivial MIA specifically tailoredfor RF. Leveraging these refined metrics, we uncover a distinct temporalpattern: under standard uniform temporal training, a model’s suscepti-bility to MIA strictly peaks at the integration midpoint, a phenomenonwe justify via the network’s forced deviation from linear approximations.Finally, we demonstrate that substituting uniform timestep samplingwith a Symmetric Exponential (U-shaped) distribution effectively min-imizes exposure to vulnerable intermediate timesteps. Extensive evalu-ations across three datasets confirm that this temporal regularizationsuppresses memorization while preserving generative fidelity.

Multimodal Large Language Models (MLLMs) demonstrateexceptional semantic reasoning but struggle with 3D spatial perceptionwhen restricted to pure RGB inputs. Despite leveraging implicit geo-metric priors from 3D reconstruction models, image-based methods stillexhibit a notable performance gap compared to methods using explicit3D data. We argue that this gap does not arise from insufficient geo-metric priors, but from a misalignment in the training paradigm: text-dominated fine-tuning fails to activate geometric representations withinMLLMs. Existing approaches typically resort to naive feature concate-nation and optimize directly for downstream tasks without geometry-specific supervision, leading to suboptimal structural utilization. To ad-dress this limitation, we propose GAP-MLLM, a Geometry-AlignedPre-training paradigm that explicitly activates structural perception be-fore downstream adaptation. Specifically, we introduce a visual-promptedjoint task that compels the MLLMs to predict sparse pointmaps along-side semantic labels, thereby enforcing geometric awareness. Further-more, we design a multi-level progressive fusion module with a token-level gating mechanism, enabling adaptive integration of geometric priorswithout suppressing semantic reasoning. Extensive experiments demon-strate that GAP-MLLM significantly enhances geometric feature fusionand consistently enhances performance across 3D visual grounding, 3Ddense captioning, and 3D video object detection tasks.

Existing multispectral detectors are limited by discrete spec-tral processing, a scale-dependent shift in the relative reliability of spec-tral and spatial cues across pyramid levels, and the lack of explicitrotation-equivariant geometric priors for arbitrarily oriented objects. Totackle these limitations, we propose FressDet, a fully rotation-equivariantspectral_x0015_spatial learning framework for multispectral object detection,capable of capturing the continuous, ordered nature of spectral struc-ture and enabling reliable spectral_x0015_spatial fusion across pyramid levelsunder arbitrary in-plane rotations. FressDet integrates three complemen-tary components. Spectral Implicit Warp (SpeIW) enables query-basedspectral resampling via a coordinate-conditioned implicit _x001C_eld, yieldinga monotone, order-preserving warp. Rotation-Equivariant ConsistencyWeighting (ReCoW) adaptively fuses spectral and spatial branches basedon branch reliability, reinforcing informative cues while suppressing noiseacross pyramid levels. The oriented-aware head exploits group-indexedfeatures to stably predict oriented objects without parameter replication.Taken together, FressDet learns more discriminative and robust spectral_x0015_spatial representations even under rotational perturbations. By achievingstate-of-the-art performance with 93% fewer parameters on _x001C_ve publicbenchmarks, FressDet demonstrates its e_x001B_ectiveness and generalizability.Code is available at https://github.com/Riiluo/FressDet.

Video polyp segmentation plays a pivotal role in early col-orectal cancer diagnosis, yet its clinical deployment is heavily bottle-necked by the prohibitive cost of dense pixel-level annotations. To addressthis challenge, we propose a foundation-guided spatiotemporal distilla-tion framework named FSD-Net for sparsely annotated video polyp seg-mentation. Our approach introduces a Semantic Flow Distillation (SFD)module that leverages a frozen foundation model as a teacher to extractrobust semantic features and generate continuous supervision signals fora lightweight student network. To tackle the frequent target disappear-ance caused by complex intestinal folds, we design an Occlusion-AwareCycle Consistency (OACC) mechanism that dynamically blocks erro-neous gradient propagation based on visibility estimation. Furthermore,we introduce a Prototype Re-identification Module (PRM) to maintainlong-term memory and successfully recapture target identities upon reap-pearance. Extensive experiments on three standardized datasets, includ-ing SUN-SEG, CVC-612, and CVC-300, demonstrate that our frame-work significantly outperforms existing weakly supervised methods andeven surpasses state-of-the-art fully supervised models. Notably, on thehighly challenging SUN-SEG-Hard dataset, our method achieves a mDice87.04%, mIoU of 79.20%, and mHD of 20.64 mm, proving its exceptionalrobustness and label efficiency in complex endoscopic environments.

Personalized segmentation and personalized retrieval bothaim to identify the same physical object across different images. Whilethe former localizes the object within a target image, the latter retrievesimages where it appears. Despite this shared instance-level objective, thetwo tasks have largely evolved separately and are addressed with distinctsolutions. In this work, we introduce FoundYou, a unified framework builton the observation that Segment Anything 2 (SAM 2), trained to pre-serve object identity across video frames, inherently captures instance-level cues. We leverage this property to match objects across independentimages, enabling segmentation and retrieval to emerge as two outcomesof the same instance alignment process. This unified view unlocks newcapabilities beyond traditional benchmarks, including few-shot person-alized retrieval and promptable personalized segmentation with flexibleprompts. Extensive experiments show consistent gains over unified andtask-specific methods, including +18.4 mIoU on PerMIS and +17.8 mAPon ILIAS. Performance scales with additional references and remainsrobust to weaker prompts. Beyond personalization, FoundYou achievesstate-of-the-art results on category-level retrieval benchmarks. Notably,our approach keeps the SAM 2-small model entirely frozen and addsonly 5.9 M trainable parameters, yielding a 52 M-parameter model thatis over 75× faster and 20× smaller than the only prior unified solution.

Deciphering human visual intent and manipulating visual con-tent via brain signals represents a significant milestone in neuro-generativeAI. Recent studies have demonstrated impressive image reconstructionfrom brain activity, yet they often rely on slow iterative diffusion inversionprocesses, makes it difficult to achieve real-time editing and fails to unifygeneration with precise context modification. Rectified Flow has recentlyemerged as a powerful solution for fast, deterministic generation. Inspiredby this, we propose MinD-Flow, a pioneering inversion-free framework forunified brain-driven synthesis and editing. It incorporates a Neuro-awareEncoder to achieve unified encoding of heterogeneous brain signals. Com-bined with the Neuro-Perceiver Bridge (NPB), it aligns brain semanticswith visual anchors, mapping the ‘Mind Space’ to the generative latentspace. Furthermore, the Decoupled Flow Guidance (DFG) and AdaptiveSoft Masking overcome the conflict between structural preservation andsemantic changes, enabling precise local editing without noise inversion.We introduce Brain-Gen Benchmark to standardize the evaluation ofNeuro-context/creation/editing tasks. Comprehensive experiments showSOTA performance, providing a versatile tool for BCI applications.

Diffusion-based face generation commonly combines iden-tity cues and edit conditions through Euclidean mixing at the condi-tioning interface. This becomes fragile when facial attributes are cou-pled: naively adding feature vectors ignores their interactions and canlead to identity drift and artifacts. We present FlowFace, a geometry-aware identity-conditioning framework that rectifies the Stable Diffusionconditioning geometry during training without modifying the inferencesampler. FlowFace includes (i) a dual-stream manifold encoder inspiredby fiber bundles to disentangle geometry-related variation from seman-tic identity, (ii) a Lie-algebraic composition module based on a trun-cated BCH expansion to introduce an explicit interaction term betweencoupled edits, and (iii) a geodesic-consistency regularizer that learns alocal SPD metric that encourages Euclidean operations to better cor-relate with a geodesic-consistent surrogate under the learned geome-try. Experiments show that FlowFace outperforms strong baselines inidentity preservation, text alignment, and perceptual quality, while re-taining adapter-level inference efficiency. The code will be available athttps://github.com/SunFly0/Flowface.

Video generative models show emerging reasoning behaviors.It is essential to ensure that generated events remain causally consistentacross frames for reliable deployment, a property we define as reasoningcoherence. To bridge the gap in literature for missing reasoning coher-ence evaluation, we propose MME-CoF-Pro, a comprehensive videoreasoning benchmark to assess reasoning coherence in video models.Specifically, MME-CoF-Pro contains 303 samples across 16 categories,ranging from visual logical to scientific reasoning. It introduces Rea-soning Score as evaluation metric for assessing process-level necessaryintermediate reasoning steps, and includes three evaluation settings, (a)no hint, (b) text hint, and (c) visual hint, enabling a controlled investi-gation into the underlying mechanisms of reasoning hint guidance. Eval-uation results in 7 open and closed-source video models reveals insightsincluding: (1) Video generative models exhibit weak reasoning coherence,decoupled from generation quality. (2) Text hints boost apparent correct-ness but often cause inconsistency and hallucinated reasoning (3) Visualhints benefit structured perceptual tasks but struggle with fine-grainedperception. We provide our project page along with data and code at:https://video-reasoning-coherence.github.io/.

We introduce FEEL (Force-Enhanced Egocentric Learning), the first large-scale dataset pairing force measurements gathered from custom piezoresistive gloves with egocentric video. Our gloves enable scalable data collection, and FEEL contains approximately 2 million force-synchronized frames of natural unscripted manipulation in kitchen environments, with ∼45% of frames involving hand-object contact. Because force is the underlying cause that drives physical interaction, it is a critical primitive for physical action understanding. We demonstrate the utility of force for physical action understanding through application of FEEL to two families of tasks: (1) contact understanding, where we jointly perform temporal contact segmentation and pixel-level contacted object segmentation; and, (2) action representation learning, where force prediction serves as a self-supervised pretraining objective for video backbones. We achieve state-of-the-art temporal contact segmentation results and competitive pixel-level segmentation results without any need for manual contacted object segmentation annotations. Furthermore we demonstrate that action representation learning with FEEL improves transfer performance on action understanding tasks without any manual labels over EPIC-Kitchens, SomethingSomething-V2, EgoExo4D and Meccano.

Next-scale visual autoregressive (VAR) models offer stronggenerative fidelity but remain computationally prohibitive for resource-constrained edge devices. We introduce EVAR, a structured-pruningframework tailored to next-scale VAR models. We find that standardpruning paradigms are poorly matched to multi-scale VAR architec-tures: rapid token growth at later scales dominates Hessian accumu-lation, marginalizing critical coarse scales and amplifying autoregres-sive error cascading. To address this issue, EVAR introduces Pyramid-Aware Optimal Brain Surgeon. Through pyramid-aware Hessian accu-mulation, Pyramid-Aware OBS rebalances scale-wise activation statisticsand biases the solver toward preserving coarse-scale representations whilemaintaining an invertible Hessian estimate for closed-form weight com-pensation. Additionally, we propose Progressive Scale-Aware Distillation(PSAD) to counteract scale-wise gradient imbalance during fine-tuning.On ImageNet benchmarks, EVAR substantially reduces parameters andmodel footprint while retaining competitive quality. On an iOS deploy-ment, EVAR further cuts single-image latency from 494 ms to 277 ms(1.8× speedup), with FID changing only marginally.

Despite rapid progress, pretrained vision–language models still struggle when answers depend on tiny visual details or on combining clues spread across multiple regions, as in documents and compositional queries. We address this by framing grounding as test-time evidence retrieval: given a query, the model should actively identify where to look next to resolve ambiguity. To this end, we propose a training-free, model-intrinsic grounding method that uses uncertainty as supervision. Specifically, we compute the entropy of the model’s next-token distribution and backpropagate it to the visual token embeddings to obtain an entropy-gradient relevance map, without auxiliary detectors or attentionmap heuristics. We then extract and rank multiple coherent regions to support multi-evidence queries, and introduce an iterative zoom-andreground procedure with a spatial-entropy stopping rule to avoid over-refinement. Experiments on seven benchmarks across four VLM architectures demonstrate consistent improvements over existing methods, with the largest gains on detail-critical and high-resolution settings, while also producing more interpretable evidence localizations.

We address the problem of recovering the full-body meshfrom only the head pose. This task has become essential for variousapplications based on head-mounted devices or smart glasses. The chal-lenge of this task lies in estimating the pose information of unobservedbody parts based solely on a single joint (i.e., head) trajectory. Severalstudies have begun to adopt head-conditioned generative models, how-ever, such previous methods are costly and time-consuming due to thediffusion-based iterative process. As an alternative, we propose a simpleyet novel method that leverages the latent space of the guidance network,which is designed as a variational autoencoder taking full-body poses asinputs. By enforcing latent distributions of this guidance network andour head-to-motion network to be similar, latent features sampled fromthe ‘guided’ distribution, i.e., distribution learned in our head-to-motionnetwork, can be reliably decoded for natural representations of full-bodyposes even only with the head pose. One important advantage of theproposed method is that one-step sampling scheme achieves remark-ably fast inference (more than 50 times faster) compared to diffusion-based approaches. Experimental results on benchmark datasets showthat the proposed method efficiently improves the performance of ego-body mesh reconstruction. The code and model are publicly available at:https://github.com/DCVL-3D/EgoGVAE_release.

Recent advancements in LiDAR-only 3D object detectionhave demonstrated improved detection accuracy over benchmark datasets.However, the adversarial robustness of these models remains untested.Very few adversarial robustness studies exist for LiDAR-only 3D ob-ject detection and unfortunately, even they are limited to legacy models.Moreover, there is a systemic gap in the existing evaluation frameworksthat rely simply on mAP ignoring other structural and predictive factors.To fill this gap, we propose a holistic framework that evaluates adversar-ial robustness using two structural factors (point cloud density and pointcloud localization) and three predictive factors (misclassification, local-ization error, distance from ego). Using this framework, we perform anempirical study and critical analysis on recent and legacy state-of-the-art models using adversarial attacks specifically designed for LiDAR-based models. Our key finding is that high-capacity, voxel-based detec-tors are more susceptible to structured coordinate perturbations thanpillar-based detectors. Additionally, non-anchor-based detectors demon-strate poor adversarial robustness, which necessitates rethinking modeltraining techniques. Overall, our results demonstrate that recent modelsare as vulnerable to adversarial attacks as their predecessors. Therefore,we argue that there is a need to improve the evaluation benchmarksfor 3D object detection that not only reward architectural modificationsfor improving detection accuracy, but also evaluate whether the designchoices improve adversarial robustness.

V2X collaborative object detection features overcoming thelimitations of single-vehicle systems by aggregating environmental fea-tures from multiple collaborative agents. However, existing mainstreamV2X perception methods mainly focus on 2D BEV object detection.When 3D detection task is concerned, inferior results are obtained be-cause they ignore the 3D spatial misalignment caused by differing heightand attitude among the collaborators. In this paper, we propose a novelcollaborative 3D object detection framework called CoGoal3D, whichextracts and refines the 3D feature gradually in a two-stage pipeline. Inthe first stage, a multiscale 3D-aware global fusion module is designedto mitigate the 3D spatial misalignment. The resulting proposals arethen refined in the second stage with an auxiliary task of 3D point re-construction. An effective multi-agent collaborative data augmentationstrategy is further proposed to enrich the training data while minimizinginformation loss. Extensive experiments on public real-world datasetsdemonstrate that our CoGoal3D achieves new state-of-the-art perfor-mance, with 3D AP@0.7 improvements of 10.86%, 10.34%, and 10.18%on the DAIR-V2X, V2V4Real, and V2X-Real datasets, respectively.

Robust egocentric hand-gesture recognition remains chal-lenging under the rapidly changing illumination typical of intelligent re-ality environments. Existing first-person gesture datasets rarely quantifylighting, limiting controlled analysis of illumination robustness. Unlikeprior egocentric datasets that describe lighting qualitatively (e.g., “in-door”, “bright”), BiCE-HG is the first to provide calibrated lux measure-ments at 80 spatial points, enabling reproducible illumination robustnessbenchmarking. BiCE-HG (Bi-Conditional Egocentric Hand Gestures)dataset addresses this gap as the first egocentric gesture dataset withmeasured bi-conditional lighting and spatial illuminance documentation.BiCE-HG contains 19 gestures performed by 23 participants across sixconditions combining sitting, standing, and walking with full and lowlighting setups. Walking trials capture realistic motion through a 4 mlighting corridor with transitions between bright and dim zones. Spatialilluminance is recorded every 25 cm using a calibrated meter, enablingprecise modelling of lighting gradients and reproducible experimentalconditions. Baseline evaluation reaches 81.24% validation accuracy (F1= 0.80) on the static (single-frame) release with a no-attention ST-GCN,the dynamic (temporal) release lifts accuracy to 88.57% with the identicalarchitecture, isolating a 7.3-percentage-point gain from temporal infor-mation alone, while adding tri-attention contributes a further marginalgain to 89.24% (F1 = 0.88). Lighting-robustness analysis shows stablespatial measurements, with a coefficient of variation below 5% for mostconditions across a 4.95× illuminance range (29-523 lux). BiCE-HG pro-vides quantified environmental metadata and realistic mobility lightinginteractions, supporting the development of illumination-robust egocen-tric gesture-recognition systems for AR/VR applications.The dataset ispublicly available at https://doi.org/10.15131/shef.data.32374137

Recent Vision-Language-Action (VLA) models report im-pressive success rates on robotic benchmarks, but whether these scoresreflect genuine embodied reasoning remains questionable. To addressthis gap, we introduce BeTTER, a diagnostic Benchmark for TestingTrue Embodied Reasoning. By applying targeted causal interventions(e.g., spatial layout shifts and temporal extrapolation) under strict kine-matic isolation, BeTTER decouples high-level cognitive failures fromlow-level execution errors. Systematic evaluations reveal that state-of-the-art VLAs break down under these interventions, exhibiting lexical-kinematic shortcuts, semantic feature collapse, behavioral inertia, andcausal state-tracking failures. Our analysis traces these failures to twocoupled sources: real-time deployment constraints, such as capacity com-pression and myopic perception, which weaken semantic representations,and behavioral cloning, which can amplify predictive but non-causal cor-relations into shortcut policies under static training distributions. Wefurther show that highly static evaluations mask these defects by per-mitting overfitting to sensorimotor priors. Real-world robotic validationconfirms that these failures are not simulation artifacts, highlighting theneed for future VLA paradigms to preserve semantic and causal reason-ing while maintaining efficient continuous control. Code and benchmark:https://github.com/BeingBeyond/BeTTER.

Generative priors have advanced VSR, yet their multi-step inference remains costly. Recent trajectory-based one-step VSR methods accelerate generation from low-resolution (LR) inputs, but can struggle under complex real-world degradations. We analyze this limitation as trajectory drift, caused by initial state mismatch and a global timestep constraint. We propose ART-VSR, a state-time adaptation framework that estimates a continuous token-wise timestep map with an Adaptive Timestep Estimator (ATE) and uses it to guide a Latent Trajectory Rectifier (LTR) toward a compatible starting state. Conditioned on the rectified latent and spatial timestep map, ARTVSR achieves highly competitive one-step VSR performance, producing crisp, artifact-reduced details while preserving temporal consistency. Code: https://github.com/Roveer/ART_VSR.

While 3D Vision Foundation Models (3DVFMs) have demon-strated remarkable zero-shot capabilities in visual geometry estimation,their direct application to generalizable novel view synthesis (NVS) re-mains challenging. In this paper, we propose AirSplat, a novel train-ing framework that effectively adapts the robust geometric priors of3DVFMs into high-fidelity, pose-free NVS. Our approach introduces twokey technical contributions: (1) Self-Consistent Pose Alignment (SCPA),a training-time feedback loop that ensures pixel-aligned supervision toresolve pose-geometry discrepancy; and (2) Rating-based Opacity Match-ing (ROM), which leverages the local 3D geometry consistency knowl-edge from a sparse-view NVS teacher model to filter out degraded prim-itives. Experimental results on large-scale benchmarks demonstrate thatour method significantly outperforms state-of-the-art pose-free NVS ap-proaches in reconstruction quality. Our AirSplat highlights the potentialof adapting 3DVFMs to enable simultaneous visual geometry estimationand high-quality view synthesis.

Current feed-forward 3D reconstruction methods predict pixelaligned Gaussian primitives, resulting in highly redundant representa-tions. A natural solution is to prune the redundant Gaussians, but naivepruning introduces severe artifacts and often requires inference time fine-tuning, breaking the feed-forward paradigm. Based on previous works,high frequency regions require more Gaussian primitives, while low fre-quency regions can be represented with significantly fewer primitives.Motivated by this, we propose a novel approach to explicitly controlthe number of Gaussians by leveraging local texture information. Ourapproach achieves this through three key components: (1) texture es-timation to capture spatial variation in scene detail, (2) texture-awarepruning that removes redundant Gaussians from low frequency regions,and (3) an adaptive Gaussian head that predicts the modified attributesof the retained primitives without breaking the feed-forward paradigm.Experiments on RE10K, ACID, DL3DV, Tanks and Temples, and DTUdemonstrate the effectiveness of our approach, while ablation studiesvalidate the contributions of its key components.