ECCV 2026 论文列表

Classic 3D scene graph generation approaches fail to workin real-time due to the heavy computational cost of environment map-ping and the need to generate intermediate point-cloud representations.To alleviate this issue, a recent work eschews point clouds in favor ofa lightweight Gaussian distribution for each object. This approximationdrastically speeds up inference and enables real-time 3D scene graph gen-eration. However, the representation has two key weaknesses. 1) Each ob-ject is approximated by a single 3D Gaussian, which causes a severe lossof 3D geometric detail. 2) The discrepancy between this approximationand the true object geometry exacerbates the inaccurate merging of ob-ject candidates during online inference. To address these issues, we pro-pose NoPA, which represents each object as a separate non-parametricdistribution. This formulation retains 3D geometric information whilepreserving real-time inference of the parametric Gaussian formulation. Tobuild upon our novel object representation, we propose a tailored merg-ing strategy to recover coherent object instances. Specifically, we leveragemaximum mean discrepancy on kernel density estimates to enable robustmerging of object candidates during online exploration while minimizingadded computational complexity. The key is to maintain a fixed particleset per object. Furthermore, to rectify the relation loss caused by misclas-sified objects, NoPA propagates relationships between objects with highaffinity. Experiments show that NoPA substantially outperforms currentmethods without sacrificing real-time inference speed.

We present Point2Pose, a model-free method for causal 6Dpose tracking of multiple rigid objects from monocular RGB-D video.Initialized only from sparse image points on the objects, our approachtracks multiple unseen objects without requiring object CAD modelsor category priors. Point2Pose leverages a 2D point tracker to obtainlong-range correspondences, enabling instant recovery after complete oc-clusion. Simultaneously, the system incrementally reconstructs an on-line Truncated Signed Distance Function (TSDF) representation of thetracked targets. Alongside the method, we introduce a new multi-objecttracking dataset comprising both simulation and real-world sequences,with motion-capture ground truth for evaluation. Experiments show thatPoint2Pose trades some single-object pose accuracy for broader model-free tracking capabilities, including multi-object tracking and recoveryfrom complete occlusion. Project page: https://point2pose.github.io/.

End-to-end pixel-space diffusion models bypass the lossy com-pression of Latent Diffusion Models (LDMs) but struggle to jointly modellow-frequency semantics and high-frequency signals in high-dimensionalspace. Existing works heavily rely on complex pixel decoders to alleviatethis issue. In this paper, we challenge this trend by revealing that thesedecoders primarily compensate for the optimization difficulties inherentto velocity prediction (v-prediction). Under the clean data paradigm (x-prediction), they are redundant. Motivated by this insight, we advocatefor simplicity over complexity and introduce PixelU, a minimalist, single-stage U-shaped Diffusion Transformer tailored for pixel space. PixelUabandons auxiliary decoders in favor of zero-cost skip connections, whichprovide an “information highway” that directly routes uncorrupted high-frequency spatial details from shallow to deep layers. To further enablethe backbone to focus exclusively on modeling low-frequency semantics,we introduce a constant-channel spatial down-sampling mechanism asa natural low-pass filter, which compresses deep features into a com-pact, low-frequency semantic manifold. Extensive experiments demon-strate that this decoupling of frequencies could outperform the strongbaseline (JiT-G) with only about 1/3 of its computation cost. On Im-ageNet 256 × 256 and 512 × 512, PixelU achieves FID of 1.63 and 1.92respectively, surpassing recent pixel-space methods and establishing asimple yet powerful new paradigm for end-to-end diffusion models. Codewill be available at https://github.com/gzp6688/PixelU.

While instruction-guided image editing has seen significantstrides, most existing models fail to preserve physical laws and causal-ity. This limitation stems from the sparsity of physical supervision sig-nals, coupled with the static, non-causal architecture of prevailing frame-works, which together hinder a deep comprehension of the physical world.To bridge this gap, we present PhysEdit, a novel framework that en-forces physical consistency in image editing through causal generationand physics-aware reinforcement learning. Specifically, by harnessing theintrinsic physical priors of video generative models, we curate a high-quality dataset, dubbed PhysEdit-50K, wherein every sample encapsu-lates the complex physical causalities and dynamics involved in the tran-sition between input and edited images. Building upon this, a specializedCausal Image Editor is devised to factorize the image editing processalong the temporal evolution of visual transformations, thereby internal-izing the underlying physical laws and boosting causal coherence. Simul-taneously, an additional regularization term is incorporated to maintainvisual consistency in unedited regions beyond the simple flow matchingobjective, mitigating the impact of undesired viewpoint shifts or cam-era motions in the dataset synthesized by video models. Furthermore,we integrate physics-aware reinforcement learning with a tailored fine-grained reward to steer the editing process toward better adherence tophysical laws. Extensive experiments on PICABench and ImgEdit-Benchdemonstrate that our PhysEdit significantly outperforms state-of-the-artbaselines, yielding physically consistent and plausible edits. Code is pub-licly available at: https://github.com/HiDream-ai/PhysEdit/.

Vision-Language Models (VLMs) have revolutionized docu-ment parsing by enabling end-to-end mapping from images to structuredtext, imposing a significant latency bottleneck, particularly for token-dense documents. While Multi-Token Prediction (MTP) has emergedas a promising approach for accelerating inference, its potential is con-strained by optimization instability when scaling to deeper look-aheaddepth. In this paper, we propose P-MTP, a framework that leveragesProgressive Multi-Token Prediction with a lightweight MTP mod-ule to scale the look-ahead depth for high-throughput document parsing.Specifically, we introduce Progressive Curriculum Loss that adaptivelyre-weights different look-ahead depths using cumulative path reliabilityand retrospective target consistency. By effectively suppressing gradientnoise in long-range predictions, P-MTP facilitates an automated easy-to-hard optimization transition, enabling the model to master increasinglydistant look-ahead depths. Furthermore, we propose Confidence-GatedDynamic Drafting to maximize the effective look-ahead depth and ac-ceptance rate by adaptively calibrating speculative length during infer-ence, thereby minimizing computational waste and further pushing theboundaries of inference speedup. Experimental results across multiplebenchmarks and architectures demonstrate that P-MTP achieves up toa 5× speedup with negligible loss in accuracy, providing the first suc-cessful validation of extensive look-ahead MTP in the document parsingdomain.

Anatomical understanding, which is the ability to identify,localize, or segment anatomical structures, is critical in medical imageanalysis; however, its progress is constrained by the scarcity of expert-labeled data. A promising remedy is to leverage an annotated referenceimage to guide the interpretation of an unlabeled target. Although re-cent vision–language models (VLMs) exhibit non-trivial visual reasoning,their reference-based understanding and fine-grained localization remainlimited. We introduce RAU, a framework for reference-based anatomicalunderstanding with VLMs. We first show that a VLM learns to identifyanatomical regions through relative spatial reasoning between referenceand target images, trained on a moderately sized dataset. We validatethis capability through visual question answering (VQA) and boundingbox prediction. Next, we demonstrate that the VLM-derived spatial cuescan be seamlessly integrated with the fine-grained segmentation capabil-ity of SAM2, enabling localization and pixel-level segmentation of smallanatomical regions, such as vessel segments. Across two in-distributionand two out-of-distribution datasets, RAU consistently outperforms aSAM2 fine-tuning baseline using the same memory setup, yielding moreaccurate segmentations and more reliable localization. More importantly,its generalization ability to unseen modalities makes it scalable to unseendatasets, a property crucial for medical image applications. To the bestof our knowledge, RAU is the first to explore the capability of VLMs forreference-based identification, localization, and segmentation of anatom-ical structures in medical images. Its promising performance highlightsthe potential of VLM-driven approaches for anatomical understandingin automated clinical workflows.

Learning robust navigation policies directly from visual ob-servations remains a fundamental challenge in vision-based robotic nav-igation. In end-to-end imitation learning approaches, the visual encoderand action decoder are jointly optimized using a single action loss, whichprovides only an indirect supervisory signal to the encoder. This indi-rect supervision frequently results in the encoder learning ambiguous,action-agnostic representations. The problem is further complicated bysubstantial variations in scene structure and appearance across diverseenvironments, as well as the prevalence of visual distractors inherent toreal-world navigation settings. Such action-agnostic features cause thenavigation policy to produce inconsistent actions at ambiguous deci-sion points, leading to navigation failure. To overcome these limitations,we propose ORION (Ordinal Neural Collapse for Visual Navigation), amethod that explicitly organizes the encoder’s representation space ac-cording to the ordinal structure of navigation actions. In the context ofgoal-directed navigation, ego-centric control categories from Far Left toFar Right exhibit a natural ordinal relationship in which neighboringclasses share similar visual contexts, while semantically opposing classesdiffer substantially in appearance. We encourage class representations tobe arranged sequentially along a single discriminative axis, while sup-pressing off-axis variance within each class. The pretrained encoder isthen integrated into a diffusion-based navigation framework, and the fullpipeline is fine-tuned end-to-end. Extensive experiments in both simula-tion and real-world settings show that ORION consistently outperformsend-to-end and neural collapse baselines in navigation success rate andgoal progress, with notable gains in visually challenging scenarios suchas complex multi-way intersections.

When evaluating identity-focused tasks such as personalizedgeneration and image editing, existing vision encoders entangle objectidentity with background context, leading to unreliable representationsand metrics. We introduce a principled framework for matched-contextidentity confusion using Near-identity (NearID) distractors, where se-mantically similar but distinct instances are placed on the exact samebackground as a reference image, eliminating contextual shortcuts andisolating identity as the sole discriminative signal. Based on this prin-ciple, we present the NearID dataset (19K identities, 316K matched-context distractors) together with a strict margin-based evaluation pro-tocol. Under this setting, pre-trained encoders perform poorly, achievingSample Success Rates (SSR), a strict margin-based identity discrimina-tion metric, as low as 30.7% and often ranking distractors above truecross-view matches. We address this by learning identity-aware repre-sentations on a frozen backbone using a two-tier contrastive objectiveenforcing the hierarchy: same identity > NearID distractor > randomnegative. This improves SSR to 99.2%, enhances part-level discrimina-tion by 28.0%, and yields stronger alignment with human judgments onDreamBench++, a human-aligned benchmark for personalization.

We present ScaleMoGen, a scale-wise autoregressive frame-work for text-driven human motion generation. Unlike conventional au-toregressive approaches that rely on standard next-token prediction, Scale-MoGen frames motion generation as a coarse-to-fine process. We quantize3D motions into compositional discrete tokens across multiple skeletal-temporal scales of increasing granularity, learning to generate motionby autoregressively predicting next-scale token maps. To maintain struc-tural integrity, our motion tokenizers and quantizers are explicitly de-signed so that discrete tokens at every scale strictly preserve the skeletalhierarchy. Additionally, we employ bitwise quantization and prediction,which efficiently scale up the tokenizer vocabulary to preserve motiondetails and stabilize optimization. Extensive experiments demonstratethat ScaleMoGen achieves state-of-the-art performance, establishing anFID of 0.030 (vs. 0.045 for MoMask) on HumanML3D and a CLIP Scoreof 0.693 (vs. 0.685 for MoMask++) on the SnapMoGen dataset. Further-more, we demonstrate that our skeletal-temporal multi-scale representa-tion naturally facilitates training-free, text-guided motion editing.

Decades of orbital missions have produced multi-modal re-mote sensing data for the Moon, spanning optical imagery, spectroscopy,thermal emission, radar, gravity, and elemental composition. Yet thesedatasets remain fragmented across archives, and no benchmark exists forevaluating machine learning on lunar data. We introduce Moonstone, thefirst multi-modal foundation model benchmark for lunar remote sensing.Our contributions are: (1) a 28-channel, 128 pixels-per-degree (∼237 m)global lunar pretraining dataset from seven instrument families acrossfive missions, (2) MG-MAE, a modality-grouped masked autoencoderwith per-group convolutional tokenizers, a shared Vision Transformerencoder, attention masking for missing modalities, coverage-adaptivemasking for heterogeneous spatial coverage, and spectral continuity reg-ularization for physically plausible reconstructions, and (3) a benchmarkof six downstream tasks covering classification, regression, and segmen-tation. MG-MAE pretrained features outperform scratch baselines onall tasks and surpass both ImageNet-pretrained and vanilla MAE base-lines by large margins. We release the pretraining dataset, code, and thebenchmark suite.3

Referring Video Object Segmentation (RefVOS) seeks to seg-ment target objects in videos guided by natural language descriptions,demanding both temporal reasoning and fine-grained visual comprehen-sion. Existing sampling strategies for LLM-based approaches typicallyrely on either handcrafted heuristics or external keyframe models. Theformer often overlooks essential temporal cues, while the latter increasessystem complexity. To address this, we propose a unified frameworkthat jointly optimizes Temporal Sentence Grounding (TSG) and Re-fVOS, naturally incorporating key moment grounding capability. Duringtraining, we introduce a novel TSG paradigm that employs a dedicated[FIND] token for key moment identification through temporal token sim-ilarity matching, thereby avoiding the need for external timestamp en-codings. For inference, we design a Moment-Centric Sampling (MCS)strategy that densely samples informative moments while sparsely sam-pling non-essential frames, preserving both motion details and globalcontext. To further enhance tracking stability, we develop BidirectionalAnchor-updated Propagation (BAP), which leverages the most relevantmoment as start point for high-quality mask initialization and dynam-ically updates at sampled points to mitigate accumulated errors. Codewill be released at https://github.com/Dmmm1997/MomentSeg.

Multi-View Multi-Object Tracking (MV-MOT) aims to localize and maintain consistent identities of objects observed by multiple sensors. This task is challenging, as viewpoint changes and occlusion disrupt identity consistency across views and time. Recent end-to-end approaches address this by jointly learning 2D Bird’s Eye View (BEV) representations and identity associations, achieving high tracking accuracy. However, these methods offer no principled uncertainty accounting and remain tightly coupled to their training configuration, limiting generalization across sensor layouts, modalities, or datasets without retraining. We propose ModTrack, a modular MV-MOT system that matches end-to-end performance while providing cross-modal, sensoragnostic generalization and traceable uncertainty. ModTrack confines learning methods to just the Detection and Feature Extraction stage of the MV-MOT pipeline, performing all fusion, association, and tracking with closed-form analytical methods. Our design reduces each sensor’s output to calibrated position-covariance pairs (z,R); cross-view clustering and precision-weighted fusion then yield unified estimates (zˆ,Rˆ) for identity assignment and temporal tracking. A feedback-coupled, identityinformed Gaussian Mixture Probability Hypothesis Density (GM-PHD) filter with HMM motion modes uses these fused estimates to maintain identities under missed detections and heavy occlusion. ModTrack achieves 95.5 IDF1 and 91.4 MOTA on WildTrack, surpassing all prior modular methods by over 21 points and rivaling the state-of-the-art end-to-end methods while providing deployment flexibility they cannot. Specifically, the same tracker core transfers unchanged to MultiviewX and RadarScenes, with only perception-module replacement required to extend to new domains and sensor modalities.

Prompt learning has become a dominant paradigm for adapt-ing vision-language models (VLMs) such as CLIP to downstream taskswithout modifying pretrained weights. While extending prompts to bothvision and text encoders across multiple transformer layers significantlyboosts performance, it dramatically increases the number of trainableparameters, with state-of-the-art methods requiring millions of parame-ters and abandoning the parameter efficiency that makes prompt tuningattractive. In this work, we propose MMLoP (Multi-Modal Low-RankPrompting), a framework that achieves deep multi-modal promptingwith only 11.5K trainable parameters, comparable to early text-onlymethods like CoOp. MMLoP parameterizes vision and text prompts ateach transformer layer through a low-rank factorization that constrainsprompts to a compact subspace, providing parameter efficiency whilemotivating the need for our complementary regularization components.To further close the accuracy gap with state-of-the-art methods, we intro-duce three complementary components: a self-regulating consistency lossthat anchors prompted representations to frozen zero-shot CLIP featuresat both the feature and logit levels, a uniform drift correction that re-moves the global embedding shift induced by prompt tuning to preserveclass-discriminative structure, and a shared up-projection that couplesvision and text prompts through a common low-rank factor to enforcecross-modal alignment. Extensive experiments across three benchmarksand 11 diverse datasets demonstrate that MMLoP achieves a highly favor-able accuracy-efficiency tradeoff, outperforming the majority of existingmethods including those with orders of magnitude more parameters, whileachieving a harmonic mean of 79.70% on base-to-novel generalization.Code is available at https://github.com/sajjad-ucsb/MMLoP.

While text-guided image editing has made remarkable progress,it remains limited in structural portrait retouching. Textual descriptionsstruggle to convey fine-grained changes to facial features and body pro-portions. To address this gap, we introduce Exemplar-Based PortraitPhoto Retouching, where the model is given an exemplar pair and taskedwith inferring and applying the same retouching operations to a newquery image. Existing exemplar-based editing methods primarily focuson tasks with pronounced visual transformations. In contrast, structuralportrait retouching involves extremely delicate and localized modifica-tions, making accurate extraction and transfer of these edits challeng-ing. To tackle this, we propose MirrorPPR, a novel framework specifi-cally designed to capture and transfer subtle structural retouching oper-ations. Our method uses a Retouching Operation Extractor to capturethe subtle differences from the exemplar pair. The extracted represen-tations are then injected into a pre-trained Diffusion Transformer (DiT)through a connector and Low-Rank Adaptation (LoRA) modules. Fur-thermore, constructing perfectly aligned cross-identity training pairs isseverely hindered by operation misalignment. To overcome this, we pro-pose an advanced data self-augmentation paradigm that ensures strictlyaligned retouching operations. To alleviate data scarcity and supportthis novel task, we introduce MirrorPPR47M, a large-scale dataset withover 47 million retouched pairs. By structuring the dataset into simu-lated and professional subsets, we enable progressive curriculum learningto smoothly optimize the network. Extensive experiments demonstratethat MirrorPPR significantly outperforms existing baselines in both re-touching quality and identity preservation. The project page is availableat https://sjtu-deng-lab.github.io/MirrorPPR.

Optical character recognition (OCR) has evolved from line-level transcription to structured document parsing, requiring models torecover long-form sequences containing layout, tables, and formulas. De-spite recent advances in vision-language models, most existing systemsrely on autoregressive decoding, which introduces sequential latency andamplifies error propagation in long documents. In this work, we revisitdocument OCR from an inverse rendering perspective, arguing that left-to-right causal generation is an artifact of serialization rather than anintrinsic property of the task. Motivated by this insight, we proposeMinerU-Diffusion, a unified diffusion-based framework that replaces au-toregressive sequential decoding with parallel diffusion denoising undervisual conditioning. MinerU-Diffusion employs a block-wise diffusion de-coder and an uncertainty-driven curriculum learning strategy to enablestable training and efficient long-sequence inference. Extensive experi-ments demonstrate that MinerU-Diffusion consistently improves robust-ness while achieving up to 3.2× faster decoding compared to autoregres-sive baselines. Evaluations on the proposed Semantic Shuffle benchmarkfurther confirm its reduced dependence on linguistic priors and strongervisual OCR capability.

We propose LF-MultiDiffusion, a training-free panorama gen-eration method that extends MultiDiffusion to support linear projectionsbetween target and reference image spaces. Our key idea is to reformulatelatent aggregation as a regularized least-squares problem and solve itefficiently with a Krylov-based iterative solver inside the denoising loop.This formulation enables denser and more natural mappings than priortraining-free methods, yielding more stable generation with far fewerperspective views. As a result, LF-MultiDiffusion reduces the number ofimage generator evaluations during denoising and significantly improvesinference efficiency. Experiments show that LF-MultiDiffusion achievesbetter visual quality, text alignment, and panoramic consistency thanthe strongest training-free baseline, while providing a 15.36× speedup.Our project page is available at: https://ahykw.github.io/lfmd.

Diffusion models (DMs) have achieved remarkable successin image and video generation. However, there are still challenges in(1) physical alignment and (2) counterfactual generation. To diagnosethe root causes, we introduce the Causal Scene Graph (CSG) to modelDM’s generative process, and build the Physical Alignment Probe (PAP)dataset to quantify the failure modes. Our analysis yields three key in-sights: (1) DMs struggle with physical elements not explicitly determinedin the prompt; (2) the prompt embedding contains disentangled repre-sentations for texture and physics; (3) visual causal structure is dispro-portionately established during the initial, computationally constraineddenoising steps. Based on these findings, we introduce Lina (LearningINterventions Adaptively), a novel framework that learns to predictprompt-specific interventions. Lina employs (1) targeted guidance in theprompt and visual latent spaces, and (2) a reallocated, causality-awaredenoising schedule. Without modifying pre-trained DM weights, Linaenforces both physical alignment and counterfactual generation in imageand video DMs, achieving state-of-the-art performance on challengingcausal generation tasks and the Winoground dataset. Our project pageis at https://opencausalab.github.io/LINA

Robust human-body understanding benefits from fusing het-erogeneous sensors—WiFi CSI, depth, LiDAR, mmWave, and RGB—yeteach modality carries distinct noise characteristics and may be intermit-tently unavailable. Existing fusion methods often assume fixed sensorsets and can collapse onto dominant modalities, reducing robustness un-der missing-modality conditions. We propose LEO-Fuse, a task-general,modality-agnostic Local Expert Orchestration (LEO) framework for uni-versal multimodal human sensing. LEO-Fuse freezes unimodal encoders,aligns modality features in a shared token space, fuses available modal-ity tokens with a mask-aware transformer, and applies task-specific lo-cal routing for downstream tasks including 3D human pose estimation(HPE) and human activity recognition (HAR). For each target task, asingle trained model handles all 31 non-empty subsets of five modalitieswithout retraining. On MM-Fi, LEO-Fuse obtains 75.5 mm mean per-joint position error (MPJPE), averaged over all 31 subsets. For HAR,HPE-trained encoders yield 93.4% average top-1 accuracy, improving to95.7% with HAR-trained encoders. On XRF55, LEO-Fuse reaches up to95.5% HAR accuracy. These results demonstrate scalable and robust uni-versal multimodal human sensing under realistic missing-modality set-tings.

Recent 3D human motion generation models demonstrateremarkable reconstruction accuracy yet struggle to generalize beyondtraining distributions. This limitation arises partly from the use of pre-cise 3D supervision, which encourages models to fit fixed coordinate pat-terns instead of learning the essential 3D structure and motion–semanticcues required for robust generalization. To overcome this limitation, wepropose LaxMotion, a framework that synthesizes realistic 3D motionswithout direct 3D pose supervision. Instead of regressing toward exactcoordinates, LaxMotion learns 3D motion as a consistent explanationof global trajectories and monocular 2D kinematic cues. We introducea structured motion factorization together with a reformulated trainingparadigm under relaxed observability. This design is further supported byrelaxed regularization objectives that enforce view-consistent alignment,orientation coherence, and structural stability. Under this relaxed super-vision paradigm, LaxMotion generates diverse, temporally coherent, andsemantically aligned 3D motions, achieving performance comparable toor surpassing fully 3D-supervised methods. These results indicate thatshifting supervision from exact coordinate matching to structural consis-tency promotes stronger reasoning and improved generalization, offeringa scalable and data-efficient paradigm for 3D motion generation.