ECCV 2026 Papers

Automatic extraction of road networks from aerial imageryis a fundamental task, yet prevailing methods rely on discrete poly-lines that inherently lack the high-order geometric continuity required bydownstream applications. While parametric curve representations, suchas our Bézier Graph, offer analytic smoothness and topological compact-ness, their extraction is bottlenecked by the absence of curve-level vectorground-truth (GT) and the fragility of heuristic vectorization pipelines.To overcome these challenges, we introduce DOGE, a novel differen-tiable rendering-driven framework for robust, curve GT-free road net-work extraction. DOGE elegantly decouples the extraction process viaa discrete-continuous coupling mechanism: a DiffAlign module contin-uously optimizes the graph’s geometric alignment against target masksusing differentiable rendering, while a TopoAdapt module applies dis-crete operators to evolve the network’s topology. This coupling preventstopological bottlenecks from paralyzing the gradient-based geometric op-timization. Extensive experiments demonstrate that our method sets anew state-of-the-art on the large-scale SpaceNet and City-Scale bench-marks. Driven purely by mask guidance, DOGE successfully reconstructshighly accurate, geometrically smooth, and topologically compact vectormaps, advancing a new paradigm for structured road network reconstruc-tion.

Diffusion models have achieved impressive results in gener-ative tasks such as text-to-image synthesis, yet they often struggle tofully align outputs with nuanced user intent and maintain consistentaesthetic quality. Existing preference-based training methods such asDiffusion Direct Preference Optimization help address these issues, butobtain their supervision targets from the forward process q(xt−1 | xt , x0 )derived from terminal samples, which is not directly aligned with themodel’s actual backward denoising transitions at each step. In this work,we introduce Direct Diffusion Score Preference Optimization (DDSPO),which defines stepwise preference supervision directly over backward de-noising transitions through a contrastive policy pair, rather than relyingon forward-process approximations from terminal samples. We proposetwo practical instantiations of the contrastive policy pair: training sepa-rate winning and losing models on preference data, and inducing a con-trastive policy pair without additional training by using a pretrainedreference model conditioned on an original prompt and a semanticallydegraded variant, requiring neither reward modeling nor manual annota-tions. Empirical results show that contrastive-policy-pair supervision ismore effective than forward-process-based supervision across text–imagealignment and aesthetic-quality tasks. Our implementation is availableat: https://dohyun-as.github.io/DDSPO

Autoregressive diffusion models have enabled high-qualityvideo generation, yet their sequential nature inherently suffers from erroraccumulation. In long-horizon video synthesis, minor prediction devia-tions compound over time, inevitably leading to unconstrained genera-tive drift, structural collapse, and severe visual degradation. To addressthis, we propose Cycle-World, a novel framework designed for stable andtemporally consistent long-video generation. Our approach tackles errordrift by enforcing strict temporal reversibility across both the trainingand inference phases. Theoretically, we demonstrate that forward gener-ative drift can be strictly bottlenecked by a cycle-consistency objective.During training, we integrate an efficient reverse-prediction model to im-plicitly embed causal constraints into the forward generator, compellingit to produce reversible sequences that tightly adhere to the natural videomanifold. At inference time, we repurpose this frozen reverse model as aruntime corrector. Through gradient-based cycle guidance, it iterativelyrefines the generated latent representations, actively suppressing accu-mulated errors before they are committed to the historical context. Ex-tensive experiments on the VBench benchmark demonstrate that Cycle-World’s dual-phase synergy significantly mitigates error drift, achievingstate-of-the-art overall generation quality and long-horizon temporal con-sistency in 60-second synthesis.

Accurate estimation of large displacement optical flow re-mains a critical challenge. Existing methods typically rely on iterativelocal search or/and domain-specific fine-tuning, which severely limits theirperformance in large displacement and zero-shot generalization scenarios.To overcome this, we introduce MegaFlow, a simple yet powerful model forzero-shot large displacement optical flow. Rather than relying on highlycomplex, task-specific architectural designs, MegaFlow adapts powerfulpre-trained vision priors to produce temporally consistent motion fields.In particular, we formulate flow estimation as a global matching problemby leveraging pre-trained global Vision Transformer features, which natu-rally captures large displacements. This is followed by a few lightweightiterative refinement to further improve the sub-pixel accuracy. Extensiveexperiments demonstrate that MegaFlow achieves state-of-the-art zero-shot performance across multiple optical flow benchmarks. Moreover, ourmodel also delivers highly competitive zero-shot performance on long-range point tracking benchmarks, demonstrating its robust transferabilityand suggesting a unified paradigm for generalizable motion estimation.Project Page: https://kristen-z.github.io/projects/megaflow/.

We propose continuous adversarial flow models, a type ofcontinuous-time flow model trained with an adversarial objective. Unlikeflow matching, which uses a fixed mean-squared-error criterion, our ap-proach introduces a learned discriminator to guide training. This changein objective induces a different generalized distribution, which empiri-cally produces samples that are better aligned with the target data dis-tribution. Our method is primarily proposed for post-training existingflow-matching models, although it can also train models from scratch.On the ImageNet 256px generation task, our post-training substantiallyimproves the guidance-free FID of latent-space SiT from 8.26 to 3.63 andof pixel-space JiT from 7.17 to 3.57. It also improves guided generation,reducing FID from 2.06 to 1.53 for SiT and from 1.86 to 1.80 for JiT.We further evaluate our approach on text-to-image generation, where itachieves improved results on both the GenEval and DPG benchmarks.

Designing efficient vision backbones is increasingly challeng-ing. An intriguing avenue is mergeable re-parameterization, which canreshape training dynamics by introducing auxiliary parameterizationswithout inference overhead. In particular, RepVGG-style structures can beviewed as adding a fixed-support correction on top of a dense base kernelafter merging. In this paper, we propose Critical Parameter Rectification(CaRe), which generalizes fixed corrections to learned selective rectifi-cation. CaRe factorizes each layer as W = Wbase + Waux ⊙ ϕ(Wgate ),where a static parameter-space gate induces sparse activation in thecorrection term, concentrating updates on a small subset of critical pa-rameters during training. Since the gate depends only on parameters,CaRe is losslessly mergeable into a standard weight tensor, incurring zeroinference overhead. Built on CaRe, the CaReNet family establishes astrong accuracy-throughput Pareto frontier on ImageNet-1K, with consis-tent gains across model scales and strong transfer to downstream denseprediction tasks. Code is available at https://github.com/lime-j/care.

Multiple Instance Learning (MIL) is widely used for weaklysupervised learning, particularly in digital pathology, where fine-grainedannotations are costly. Most MIL methods aggregate instance featuresvia attention mechanisms. However, attention weights do not alwaysfaithfully reflect instance importance and may focus on spuriously cor-related regions. In this work, we propose CAR-MIL, a framework thatexplicitly guides attention learning through a counterfactual attentionregularization objective inspired by counterfactual explanations. Built ona standard attention-based MIL architecture, our approach introduces alightweight counterfactual attention branch trained to produce an alter-native prediction while remaining close to the factual attention distribu-tion. This encourages prediction changes to arise from minimal, struc-tured redistributions of attention, leading to more informative evidenceallocation. The resulting factual and counterfactual attention maps cap-ture complementary evidence: the former highlights regions supportingthe prediction, while the latter reveals regions whose reweighting wouldchallenge it. We evaluate our method on synthetic MIL benchmarks withinstance-level ground truth enabling controlled analysis of attention be-havior and on five digital pathology datasets across four tasks. CAR-MILmaintains competitive classification performance, with the largest gainsobserved on more challenging tasks, while improving attention reliability,demonstrating the benefits of integrating counterfactual explainabilityreasoning into attention learning.

Recent advances in Reinforcement Learning with VerifiableRewards (RLVR) for multimodal large language models (MLLMs) havemainly focused on improving final answer correctness and strengthen-ing visual grounding. However, a critical bottleneck remains: althoughmodels can attend to relevant visual regions, they often fail to effectivelyincorporate visual evidence into subsequent reasoning, leading to reason-ing chains that are weakly grounded in visual facts. To address this issue,we propose Trajectory-Guided Reinforcement Learning (TGRL), whichguides the policy model to integrate visual evidence into fine-grained rea-soning processes using expert reasoning trajectories from stronger mod-els. We further introduce token-level reweighting and trajectory filteringto ensure stable and effective policy optimization. Extensive experimentson multiple multimodal reasoning benchmarks demonstrate that TGRLconsistently improves reasoning performance and effectively bridges thegap between visual perception and logical reasoning.

Visual grounding (VG) aims to precisely localize the objectin input images based on its natural language descriptions. Most recentlyproposed methods address this task using Transformer-based architec-tures that can inject textual information into visual features. However,according to the image tokenization procedure, a large number of vi-sual tokens will be located in text-irrelevant background areas. Thesetokens can introduce noise into the attention calculation, thus reduc-ing the significance of foreground object tokens and ultimately affect-ing the effectiveness of these methods. To this end, we propose a novelToken Blurring (ToB) module, which dynamically merges image tokensbased on the pair-wise visual similarity between them and their tex-tual relevance with input expressions. By reducing the number of text-irrelevant background tokens and preserving the density of text-referredones, ToB can improve both model effectiveness and efficiency in solvingVG tasks. Extensive experiments on RefCOCO, RefCOCO+, and Re-fCOCOg datasets show that Transformer-based models equipped withour ToB module yield better results while reducing computational over-head compared to various existing VG methods. Our code is available athttps://github.com/Mr-Bigworth/ToB.

High-fidelity 4D physical simulation requires accurate phys-ical parameter assignments that align with diverse material properties.Previous methods struggle with critical bottlenecks: they either rely oncomputationally intensive per-scene optimization or predict parametersfrom purely visual inputs. The latter suffers from inherent ambiguity,e.g., the inability to distinguish the stiffness of rubber from its ap-pearance alone. To bridge this gap, we propose AutoPhyX, a text-conditioned framework for predicting spatially-varying physical proper-ties. Since training robust feed-forward models requires fine-grained su-pervision, we first introduce a novel part-controllable data generationpipeline that decomposes complex 3D assets into semantically distinctcomponents and pairs them with physically plausible, diverse parame-ters and corresponding text descriptions. Leveraging this dataset, Auto-PhyX employs a cross-modal modulation mechanism where text featuresdynamically modulate visual features for precise physical grounding. Byformulating property prediction within a voxel field, AutoPhyX en-sures compatibility with diverse 3D representations, including meshes,point clouds, Gaussian Splatting, and NeRF. Experiments demonstratethat our method enables physically plausible, text-driven parameter as-signment in a single forward pass. It achieves high accuracy and robustgeneralization to in-the-wild objects, paving the way for future physics-intensive embodied and robotic applications.

Existing ViT-based weather forecasting models apply uni-form computation across all spatial tokens, even though nearby atmo-spheric grid points often contain similar values and large regions evolvesmoothly over time. This makes much of the intermediate per-token com-putation redundant. Standard token-efficiency methods, such as pruningor merging, reduce cost by removing or fusing tokens. However, weatherforecasting is a spatiotemporal dense prediction problem in which a his-tory of atmospheric states must be mapped to future values on the orig-inal latitude-longitude grid. Thus, every grid cell must retain a physi-cally meaningful representation, especially under autoregressive rollout.We introduce Sparse-Reslim, a parameter-free plug-in routing modulethat makes sparse token processing compatible with this fixed-grid re-quirement. Sparse-Reslim routes only 25% of spatial tokens through theexpensive middle transformer blocks and treats those blocks as residualupdates: it computes the change produced for the routed tokens and scat-ters only this delta back to the full sequence. Unselected tokens keep theirpre-routing representations exactly, so no grid cell is dropped or replacedby a mask token, and no fusion layer or additional parameters are intro-duced. Across ERA5 resolutions up to the operational 0.25° standard andtwo model families, a deterministic Transformer and a diffusion model,Sparse-Reslim improves forecast accuracy on every evaluated variablewhile substantially reducing cost: training is about 2.5× faster in themain settings and reaches 3.18× speedup at 0.25°, with over 2.2× lowerpeak memory. A controlled decomposition shows that the accuracy gaincomes primarily from sparse routing itself, while random token selectionprovides an additional regularization benefit without selector overhead.Code is available at https://github.com/janet-sw/Sparse-Reslim.

Detecting and localizing defects in 3D point clouds is chal-lenging because abnormal samples are scarce and diverse, while train-ing is often limited to normal data. We propose Anomaly Factory 3D(AF3AD), a modular framework that synthesizes diverse pseudo-anomaliesfrom normal point clouds to expand the training data for unsupervised3D anomaly detection methods that rely on pseudo-anomalies. AF3ADuses a center-conditioned parametric deformation model defined in localPCA frames, with kernel-controlled spatial falloff, anisotropy, directionalgating, and normal/tangential displacement fields, enabling a broad setof geometric defect presets. We demonstrate its ease-of-use and effec-tiveness by integrating AF3AD with an offset-prediction detector and areconstruction-based anomaly detection method, showing that AF3ADtransfers across detection paradigms. Experiments on AnomalyShapeNetand Real3D-AD show consistent improvements in object- and point-leveldetection and localization, supported by ablations on preset groups androbustness under noise. AF3AD is designed as a standalone synthesis toolto facilitate adoption across different 3D anomaly detection paradigms.Code is available at github.com/vpc-ccg/AF3AD.

Vision-language models (VLMs) pre-trained on large-scaledatasets demonstrate strong generalization in few-shot scenarios, oftensurpassing conventional visual models on natural samples. However, theyremain vulnerable to adversarial perturbations, and directly applyingadversarial training may disrupt the cross-modal alignment, degradinggeneralization to novel classes. To mitigate these issues, we proposean Inverse-Adversarial and Difficulty-Adaptive Robust Vision-LanguageModel (IADA-RVLM) for robust few-shot classification. Specifically, toconstrain the optimization trajectory, our method jointly leverages ad-versarial and inverse-adversarial samples, where inverse-adversarial sam-ples serve as semantic anchors to guide representations toward high-confidence regions. To improve robustness, we incorporate visual prompttuning and text adapters to inject task-specific knowledge into bothmodalities. To balance robustness and generalization under varying sce-narios, we quantify adversarial transfer difficulty via semantic distancesand dynamically adjust the fusion of general and specialized knowl-edge. Extensive experiments across multiple datasets and attack settingsdemonstrate that our method achieves a superior overall trade-off be-tween robustness and generalization.

Vision-Language Pretrained Models (VLPMs) offer a scal-able path to open-vocabulary chest radiology understanding, yet two as-pects remain underexplored: how structured clinical semantics extractedfrom medical reports can reduce in-batch noise during contrastive learn-ing, and how cross-modal fusion can be designed to produce more faithfulspatial grounding without added complexity. We introduce AlphaRad,addressing these opportunities through two contributions. First, we con-struct a large-scale structured medical concept space from medical re-ports parsed by a Large Language Model for training, thereby mit-igating in-batch learning noise and removing heuristic pair matchingin contrastive learning, and thus naturally positioning AlphaRad asa medical concept discriminator trained via α-Corrected Binary Cross-Entropy. Second, we propose FLaS (Factorized Latent Supervision), anextremely simple yet effective cross-modal feature fusion module thatfactorizes VLPM representations into independent subspaces, using ded-icated alignment supervision to enhance the expressiveness of spatialgrounding without introducing additional model parameters. Throughextensive empirical validation, AlphaRad shows strong zero-shot gen-eralization across diverse chest radiology tasks. Notably, it establishesstate-of-the-art average performance across 16 classification benchmarks,while achieving individual state-of-the-art results via distinct gains on 7grounding/phrase grounding and 3 segmentation datasets.

Despite the rapid progress in data-driven 3D vision, aerialgeometric 3D vision remains a formidable challenge due to the severescarcity of large-scale, high-fidelity training data. Existing benchmarks,predominantly biased toward ground-level or object-centric views, donot account for complex viewpoint transformations and diverse envi-ronmental conditions in UAV-based sensing. To bridge this critical gap,we propose AirZoo, a unified large-scale dataset and benchmark forgrounding aerial geometric 3D vision. AirZoo possesses three appeal-ing properties: 1) Scalable Generation Pipeline: Leveraging freelyavailable, world-scale photogrammetric 3D meshes, it renders vast out-door environments with customizable UAV flight trajectories and config-urable weather/illumination. 2) Comprehensive Scene Diversity: Itprovides extensive global coverage (22 countries, 95 base flight sequences,and 377 weather-conditioned trajectories), systematically encompassingboth highly structured urban landscapes and complex unstructured nat-ural environments. 3) Rich Geometric Annotations: Each frameprovides synchronized, pixel-level metric depth and precise 6-DoF geo-referenced poses, essential for geometry-aware learning. Through threerigorous evaluation tracks—aerial image retrieval, cross-view matching,and multi-view 3D reconstruction—we demonstrate that AirZoo servesas a powerful "pre-training engine." Extensive experiments on both pub-lic and newly collected real-world benchmarks reveal that fine-tuning onAirZoo yields substantial performance gains for state-of-the-art models(e.g., MegaLoc, RoMa, VGGT, and Depth Anything 3), establishing anew performance upper bound for aerial spatial intelligence.

Reinforcement learning with verifiable rewards (RLVR) and Rubrics as Rewards (RaR) have driven strong gains in domains with clear correctness signals and even in subjective domains by synthesizing evaluation criteria from ideal reference answers. But many real-world tasks admit multiple valid outputs and lack the single ideal answer that rubric generation depends on. We identify this reference-free setting as a gap in current post-training methods and propose Implicit Error Counting (IEC) to fill it. Instead of checking what a response gets right against a rubric, IEC enumerates what it gets wrong, applying severity-weighted scores across task-relevant axes and converting them into calibrated per-aspect rewards. We show that naïve explicit enumeration is too noisy for stable optimization, and that two design choices: implicit score emission and group calibration are necessary to make error counting a reliable reward. As a case study, we validate IEC on virtual try-on (VTO), a domain that is simultaneously too constrained for holistic scoring and too permissive for rubric-based evaluation: subtle garment errors are unacceptable, yet many output variations are correct. We introduce Cascaded Error Counting (CEC) as an evaluation metric, which tracks human preferences well (60% top-1 vs. 30% others), and curate Mismatch-DressCode (MDressBench), a benchmark with maximal attribute mismatch to stress-test reward designs. On MDressBench, IEC outperforms RaR across all metrics (CEC: 5.31 vs. 5.60 on flat references; 5.20 vs. 5.53 on non-flat). On VITON-HD and DressCode, IEC matches or surpasses six baselines on 6 of 8 perceptual metrics. These results suggest that when ideal answers are unavailable, counting errors provide a stronger signal than constructing rubrics.

Video reasoning requires models to locate and track question-relevant evidence across frames. While reinforcement learning (RL) withverifiable rewards improves accuracy, it still struggles to achieve reliablespatio-temporal grounding during the reasoning process. Moreover, im-proving grounding typically relies on scaled training data or inference-timeperception tools, which increases annotation cost or computational cost.To address this challenge, we propose VisionCoach, an input-adaptiveRL framework that improves spatio-temporal grounding through visualprompting as training-time guidance. During RL training, visual promptsare selectively applied to challenging inputs to amplify question-relevantevidence and suppress distractors. The model then internalizes theseimprovements through self-distillation, enabling grounded reasoning di-rectly on raw videos without visual prompting at inference. VisionCoachconsists of two components: (1) Visual Prompt Selector, which predictsappropriate prompt types conditioned on the video and question, and(2) Spatio-Temporal Reasoner, optimized with RL under visual promptguidance and object-aware grounding rewards that enforce object identityconsistency and multi-region bounding-box IoU. Extensive experimentsdemonstrate that VisionCoach achieves state-of-the-art performanceacross diverse video reasoning, video understanding, and temporal ground-ing benchmarks (V-STAR, VideoMME, World-Sense, VideoMMMU, Per-ceptionTest, and Charades-STA), while maintaining a single efficientinference pathway without external tools. Our results highlight the effec-tiveness of training-time visual prompting as a lightweight mechanismfor improving grounded video reasoning.

Recent diffusion- and flow-based image editing methods en-able high-quality text-guided image editing. However, improving editstrength often degrades identity and background preservation, a phe-nomenon known as the edit—preservation trade-off. This trade-off iswidely observed and commonly regarded as an inherent limitation ofgenerative editing. In this work, we revisit this assumption and iden-tify trajectory instability as a previously overlooked factor behind theedit–preservation trade-off. Through empirical analysis across multipleediting methods and backbones, we find that trajectory instability isclosely associated with degraded semantic preservation while showinglittle association with edit alignment. We characterize instability by us-ing directional consistency and update variability of latent update ve-locities during integration, revealing that unstable integration dynamicscontribute substantially to the edit–preservation trade-off. Motivated bythis insight, we propose V-HOLD, a simple training-free method thatstabilizes editing trajectories by maintaining a target-oriented updatedirection across multiple integration steps instead of recomputing veloc-ities at every step. Experiments on PIE-Bench and the FlowEdit-Datashow that our method consistently improves semantic preservation whilemaintaining competitive edit alignment across multiple backbones. Inaddition, V-HOLD reduces the number of velocity-field evaluations dur-ing editing, improving computational efficiency. These results highlighttrajectory stability as a key factor for improving semantic preservationwithout sacrificing edit alignment.

While various works address reflective symmetry understand-ing in 3D data and images, pixel-level semantic left-right prediction ofin-the-wild images remains challenging, due to certain difficulties includ-ing the lack of 3D information, occlusion, object pose variation, partial-ity, etc. In this work, we propose an unsupervised learning framework totackle this challenge. Leveraging recent advances in vertex-wise semanticleft-right understanding of 3D data, our unsupervised learning methodjointly utilises 3D shape and image datasets to infer pixel-wise seman-tic left-right predictions in single-view images. In particular, we showthat a medium-scale 3D shape dataset comprising mainly of human- andquadruped animal-like shapes, combined with diverse in-the-wild imagedata, are sufficient to achieve high-quality semantic left-right predictionin images, even for entirely unseen 3D object categories, such as carsor trains. Overall, our approach achieves superior performance in densepixel-wise semantic left-right predictions on both rendered and in-the-wild image datasets when compared to existing state-of-the-art methods.