ECCV 2026 Papers

Direct Preference Optimization (DPO) has proven effectivefor aligning diffusion models with human preferences. However, by re-lying solely on endpoint supervision, existing DPO fine-tuning practicesignore the curved trajectory of the pretrained generation process. Thisforces a linear shortcut in the latent space, disregarding the intrinsicgeometry and leading to off-trajectory and off-manifold generation. Toaddress this, we propose Trajectory-Consistent Preference Optimization(TCPO), a fine-tuning paradigm respecting the model’s intrinsic genera-tive geometry. TCPO first performs Curvature-Adaptive Sampling (CAS)to dynamically select a curvature-aware intermediate latent as a geo-metric anchor. Then, Trajectory-Rectified Fusion (TRF) combines thisgeometric anchor with the conventional endpoint target which acts as asemantic anchor to form a trajectory-consistent supervision signal. Ex-tensive experiments show that TCPO achieves superior visual fidelityand text-to-image alignment on standard benchmarks, with improvedtraining efficiency. 4

Method illustrations (MIs) play a crucial role in conveyingthe core ideas of scientific papers, yet their generation remains a labor-intensive process. Here, we take inspiration from human authors’ drawingpractices and correspondingly propose FigAgent, a novel multi-agentframework for high-quality automatic MI generation. Our FigAgent dis-tills drawing experiences from similar components across MIs and encap-sulates them into reusable drawing middlewares that can be orchestratedfor MI generation, while evolving these middlewares to adapt to dynami-cally evolving drawing requirements. Besides, a novel Explore-and-Selectdrawing strategy is introduced to mimic the human-like trial-and-errormanner for gradually constructing MIs with complex structures. Exten-sive experiments show the efficacy of our method. Project page is here.

Multimodal Large Language Models (MLLMs) are critically hampered by hallucination—generating content inconsistent with the provided image. In this paper, we identify an internal signature of hallucination: progressive degradation of text-to-image cross-attention during generation, leading to specific failure patterns like unfocused or biased attention. Existing mitigation strategies are largely outcome-driven and do not explicitly target this failure mode. To address this problem, we propose ADAPT (Attention Dynamics Alignment with Preference Tuning), an attention-based framework that intervenes directly on text-toimage cross-attention dynamics. We propose ADAPT with three key contributions: a cross-attention visual anchor refined from early decoding to provide stable spatial grounding, an attention-supervised inference mechanism that detects and corrects attention drift online, and a Visual Attention Guidance DPO that aligns preferences toward visually grounded responses. Experiments show that each component of ADAPT contributes to hallucination reduction, and the full framework achieves new best results across multiple hallucination benchmarks, reducing hallucination rates by 40–60% across mainstream backbones while preserving general multimodal capabilities. Our work provides an attentionbased perspective on mitigating hallucinations by exploring the model’s internal text-to-image cross-attention behaviors. Code is available at: https://github.com/yao-ustc/ADAPT.

Synthetic data generation has emerged as a powerful tool for improving data scalability in computer vision. Recent diffusion-based pipelines have demonstrated strong photorealism. However, enforcing precise 3D structure and pose consistency in generated images remains challenging. Existing methods leverage visual prompts such as edge maps to guide diffusion models, but often suffer from over-conditioning artifacts that degrade image realism and limit dataset quality. In this paper, we present a diffusion-based image generation framework that enforces 3D structural alignment while preserving photorealism through adaptive conditioning. Our framework, Adaptive Conditioning for 3D-Aware Synthetic Data Generation (AC3S), introduces a self-supervised visual prompt modulator that dynamically adjusts the strength of ControlNet conditioning, preventing over-conditioning and enabling the diffusion model to retain its generative expressiveness. To further enhance diversity and semantic consistency, we develop a multi-agent vision language model framework that composes detailed and 3D-aware prompts aligned with the underlying geometric structure. Together, these components enable the scalable generation of high-quality synthetic datasets with accurate 2D and 3D annotations. Extensive experiments demonstrate that our method significantly improves image quality and downstream utility.1

Large Language Models (LLMs) and Vision Language Mod-els (VLMs) have shown impressive reasoning abilities, yet they strugglewith spatial understanding and layout consistency when performing fine-grained visual editing. We introduce a Structured Reasoning frameworkthat performs text-conditioned spatial layout editing via scene-graph rea-soning. Given an input scene graph and a natural-language instruction,the model reasons over the graph to generate an updated scene graphthat satisfies the text condition while maintaining spatial coherence. Byexplicitly guiding the reasoning process through structured relationalrepresentations, our approach improves both interpretability and controlover spatial relationships. We evaluate our method on a new text-guidedlayout editing benchmark encompassing sorting, spatial alignment, androom-editing tasks. Our training paradigm yields an average 15% im-provement in IoU and 25% reduction in center-distance error comparedto Chain of thought Fine-tuning (CoT-SFT) and vanilla GRPO base-lines. Compared to SOTA zero-shot LLMs, our best models achieve upto 20% higher mIoU, demonstrating markedly improved spatial precision.

Estimating 3D geometry in underwater environments presentsunique challenges due to light attenuation, scattering, and the absenceof large-scale, high-quality 3D annotations. Pioneering methods rely onmassive dense annotations that are impractical in underwater settings.In this paper, we propose Wat3R, a cross-domain semi-supervised learn-ing framework designed to adapt feed-forward 3D reconstruction modelsfrom air to underwater scenes. Uniquely, our method eliminates the needfor any annotated underwater data following a teacher-student architec-ture, that learns robust geometry representations merely on abundantunlabeled real underwater video footage. We also design a cross-viewconsistency loss that leverages geometric cues from other views to com-pensate for the information degradation in the current view caused bywater attenuation and scattering. Furthermore, considering the lack ofcomprehensive evaluation benchmarks, we construct Water3D, a di-verse dataset covering various water bodies and underwater scenarios, de-signed for geometric task evaluation. Experimental results demonstratethat Wat3R outperforms current state-of-the-art methods in underwatermulti-view depth estimation and point cloud reconstruction. The datasetand code are available at https://github.com/LSXI7/Wat3R .

In-Image Machine Translation (IIMT) aims to translate scenetext in an image and render the translated text back into the original re-gions while preserving the overall visual appearance. Recent unified mul-timodal models provide a promising solution by combining visual-textunderstanding and image generation within a single framework. However,directly adapting such models to IIMT remains challenging. In particu-lar, they often suffer from understanding-generation conflicts, where thetranslation inferred during understanding is inconsistent with the textsupervision used in generation, and spatial position misalignment, wherethe rendered text does not accurately match the target text regions.To address these issues, we present UniTranslator, a unified multimodalframework for IIMT that tightly couples translation understanding andtext editing. Specifically, we introduce an Understand-Generation Align-ment Module (UGAM) to bridge the representation gap between un-derstanding and generation, encouraging semantic consistency betweentranslated content prediction and text rendering. We further propose aSpatial Mask Decoder (SMD) with pixel-level supervision over text re-gions to improve spatial grounding, geometric alignment, and layout con-trollability during generation. Extensive experiments on multiple bench-marks demonstrate that UniTranslator achieves state-of-the-art perfor-mance across diverse language directions and complex real-world layouts.Moreover, our results reveal a strong mutual reinforcement effect betweentranslation understanding and image generation, highlighting the advan-tage of unified translation multimodal learning. Code is available here.

In this paper, we address the challenging problem of 4D re-construction from sparse-view videos. This setup usually relies on monoc-ular depth estimation to provide priors for the reconstruction model. Akey challenge arises from limited cross-view overlap and temporal vari-ation, making monocular depth predictions inconsistent across viewsand time. Existing methods align spatial and temporal dimensions inseparate stages, requiring foreground segmentation masks while failingto leverage temporal cues for cross-view alignment. Contrary to thesemethods, we propose a unified spatial-temporal depth alignment frame-work that jointly resolves cross-view and cross-time inconsistencies with-out distinguishing foreground/background. Our method represents depthmaps across views and time as a set of spatio-temporal neural fields.This representation not only yields fast convergence, but also capturesspatio-temporal correlation among depth maps implicitly, without de-pendence on external segmentation/tracking models. We also propose amulti-view depth-order loss while leveraging the classic scale-and-shift-invariant loss to further improve the final depth quality. The aligneddepths initialize and supervise Gaussian splatting models for 4D recon-struction. Experiments on Ego-Exo4D and EgoHuman demonstrate thatour improved depth alignment substantially benefits dynamic Gaussian-splatting-based reconstruction methods for novel-time/view synthesis andgeometry accuracy/consistency.

As embodied AI transitions to real-world deployment, thesuccess of the Vision-and-Language Navigation (VLN) task tends to evolvefrom mere reachability to social compliance. However, current agentssuffer from a “Goal-driven trap”, prioritizing physical geometry (“can Igo?”) over semantic rules (“may I go?”), frequently overlooking subtleregulatory constraints. To bridge this gap, we establish Rule-VLN, the firstlarge-scale urban benchmark for rule-compliant navigation. Spanning amassive 29k-node environment, it injects 177 diverse regulatory categoriesinto 8k constrained nodes across four curriculum levels, challenging agentswith fine-grained visual and behavioral constraints. We further proposethe Semantic Navigation Rectification Module (SNRM), a universal, zero-shot module designed to equip pre-trained agents with safety awareness.SNRM integrates a coarse-to-fine visual perception VLM framework withan epistemic mental map for dynamic detour planning. Experimentsdemonstrate that while Rule-VLN challenges state-of-the-art models,SNRM significantly restores navigation capabilities, reducing CVR by19.26% and boosting TC by 5.97%. The project page is available athttps://jeffry-wen.github.io/Rule-VLN/.

Three-dimensional ultrasound (US) is a safe, radiation-freecomplementary modality to CT and X-rays for longitudinal monitoring,yet its segmentation-derived partial point clouds are extremely artifact-laden. Consequently, it is challenging to recover a clean and completeanatomical structure from such US point clouds. In this paper, we presentUBone3D, a novel framework based on physics-rectified conditional flowmatching (CFM) that performs point cloud completion directly from par-tial US observations. UBone3D models deterministic physics artifacts(e.g., surface thickening, streaking, dropouts) via a simulated physicsproxy, and introduces test-time physics rectification to steer the shapecompletion. At inference, the completion is jointly steered by two decou-pled forces: (1) anatomical plausibility enforced by a CT-trained gen-erative shape prior, BoneFM, and (2) physics consistency enforced byUSimNet in the ultrasound formation space. Extensive experimentson simulated and in-vivo data demonstrate significant improvements inreconstruction accuracy and anatomical fidelity over existing baselines.Project page: https://answerrtx.github.io/UBone3D-Proj/.

Reinforcement learning (RL) has become indispensable forpushing Vision-Language-Action Models (VLA) beyond static imitationlearning. However, existing RL methods typically necessitate externalenvironmental feedback, relying on predefined success signals to guidepolicy updates. In this work, we demonstrate that VLA models possessstrong internal evaluative capabilities: in discrete-action VLAs, trajecto-ries with higher generation confidence are significantly more likely to suc-ceed. Based on the observation, we introduce T2 VLA (Test-time VLA),an architecture-agnostic test-time RL framework that enables VLA mod-els to achieve self-bootstrapping policy improvement. Instead of relyingon external rewards, T2 VLA leverages the trajectory-level similarityto high-confidence expert demonstrations as an intrinsic reward signal.In addition, we propose a Confidence-Driven Dual Expert Bootstrappingmechanism. By dynamically balancing a Local Pseudo-Expert for aggres-sive exploration and a Global Expert Pool for training stability, T2 VLAprevents policy collapse while discovering further breakthroughs. Exten-sive experiments on the LIBERO and RoboTwin benchmarks show thatT2 VLA consistently outperforms supervised baselines and approachesoracle RL performance with ground-truth rewards, achieving effectiveimprovement without external reward feedback. Furthermore, T2 VLAcan adapt to distinct VLA paradigms, including both OpenVLA-OFTand the π series.

Traditional video highlight detection relies on a narrow, event-centric definition of saliency, which often fails to generalize to uncon-strained personal videos where highlights are heterogeneous and perspective-dependent. To address this, we introduce TRINITY, a multi-perspectivebenchmark that decomposes highlight saliency into three complemen-tary dimensions, Event, Emotion, and Nature, within a unified tempo-ral framework. Leveraging this multi-faceted view, we propose a shared-backbone multi-branch architecture designed for parallel multi-perspectiveprediction via view-specific experts. Comprehensive experiments demon-strate that our method significantly outperforms state-of-the-art base-lines, achieving gains of +7.15/ + 3.62 mAPρ=15%/50% on Mr. HiSumand +10.82 mAP on YouTube Highlights. These results validate thatmulti-perspective modeling provides a more robust and comprehensiveformulation of video saliency, especially for complex real-world scenar-ios. The benchmark and relevant codes will be released upon acceptance.The benchmark is available at https://huggingface.co/datasets/vanilladucky/TRINITY.

Analyzing whole-slide pathology images (WSIs) requires aniterative, evidence-driven reasoning process that parallels how pathol-ogists dynamically zoom, refocus, and self-correct while collecting ev-idence. However, existing computational pipelines often lack this rea-soning trajectory, resulting in opaque and unjustifiable predictions. Tobridge this gap, we present PathAgent, a training-free, large languagemodel (LLM)-based agent framework that emulates the reflective, step-wise analytical approach of human experts. PathAgent can autonomouslyexplore WSIs, iteratively and precisely locating significant micro-regionsusing the Navigator module, extracting morphological visual cues usingthe Perceptor, and integrating these findings into the continuously evolv-ing natural language trajectories in the Executor. The entire sequenceof observations and decisions forms a structured evidence trajectory,yielding traceable and evidence-grounded analyses. Evaluated across fivechallenging datasets spanning both whole-slide and patch-level settings,PathAgent exhibits strong zero-shot generalization and achieves perfor-mance comparable to human experts, surpassing task-specific baselinesacross diverse pathology tasks, including molecular subtype classifica-tion, histological grading and tumor type diagnosis. Human collaborativeevaluations demonstrate that PathAgent produces diagnostic decisionshighly consistent with expert-selected regions. These results demonstratethat PathAgent enables interpretable and clinically grounded WSI analy-sis. The code is available at https://github.com/G14nTDo4/PathAgent.

3D reconstruction from sparse views is a challenging task in3D computer vision. Recent studies on 3D Gaussian Splatting (3DGS)have achieved remarkable results with sparse views in novel view syn-thesis, yet reconstructing high-quality geometric surfaces from sparseviews remains a challenge, due to the limited geometry clues and thediscreteness of Gaussians. In this paper, we propose a novel 3DGS-basedmethod for high-fidelity surface reconstruction from sparse views. Ourkey insight is to introduce a normal-guided depth propagation approach,which can extend depth information from high-confidence regions to con-strain the depth in low-confidence areas. Additionally, we propose an ab-normal depth edge-aware regularization to address depth discontinuitiescaused by the discreteness of Gaussians. Extensive experiments on DTUand Tanks-and-Temples datasets demonstrate that our method outper-forms the state-of-the-art methods in sparse view surface reconstruction.Project page: https://hanl2010.github.io/DP-GS.

With the growing demand for realistic virtual humans, para-metric body models have become a cornerstone of modern medicine, sportsor entertainment applications. However, most of these models are inherentlylimited: they only capture the 3D surface of the skin, offering no insightinto the complex bio-mechanical structures that generate motion. As moreapplications expand towards biomechanics, the need for virtual humanmodels that go beyond the skin has become increasingly evident. Traditionalsoft-tissue simulations, such as FEM, are accurate but non-scalable and toocomputationally expensive for most common applications. Alternatively,existing biomechanical tools can simulate muscular forces and activations,but do not model changes in external shape, restricting how activationscorrelate with actual observable anatomy. This motivates a novel inverseresearch problem: recovering muscle deformations directly from visiblesurface observations - i.e., from the skin, and thus the pose. In this work,we present SOMA (from Surface Observations to Muscle Anatomy), aperson-specific model that infers spatio-temporal muscle behavior fromsurface signals obtained using RGB cameras, and SKIM, a subject-specificsoft-tissue deformation dataset. To the best of our knowledge, this is thefirst method that attempts to recover muscle deformations from multi-viewRGB data. We show how our method provides anatomically groundedanimations without the complexity of traditional simulations, leading to ascalable and cost-effective solution. Data and code are available.

Sign language generation (SLG), also known as text-to-signgeneration, aims to bridge the communication gap between signers andnon-signers. Unlike many other generative tasks, SLG must satisfy twofundamental linguistic constraints. First, sign language expresses mean-ing through a sequence of gestures aligned with word-like units calledglosses, and therefore requires correct lexical ordering to preserve in-tended meaning. Second, each gesture should faithfully reflect the in-tended gloss (semantic accuracy). Despite recent progress, existing SLGmethods frequently produce signs with incorrect lexical order and lowsemantic accuracy. A common limitation of prior approaches stems fromglobally fused conditioning strategies, which weaken temporal ground-ing, the temporal correspondence between glosses and their realized signsegments. This often leads to incorrect lexical order and semanticallyambiguous signs. To address this limitation, we propose SIGNER, aSIGN language generation framework with timE-Resolved conditioningto ensure temporal grounding, leveraging a temporal-gloss condition andlocal temporal fusion (LTF). SIGNER constructs a temporal-gloss con-dition by estimating a gloss sequence and its durations from input text,and assigning gloss semantics across the temporal dimension. We thenintroduce LTF, a temporally grounded fusion module that integratesthe temporal-gloss condition within a constrained temporal window dur-ing denoising. By enforcing temporal locality in condition fusion, LTFpreserves temporal grounding, leading to correct lexical ordering andclearer per-gloss semantics. Experiments on Phoenix-2014T and CSL-Daily demonstrate state-of-the-art performance, further supported bymotion-smoothness analysis. The project page is available here.

Training deep neural networks (DNNs) on classi(cid:28)cation tasks can be performed with a number of di(cid:27)erent losses, but cross-entropy (CE) loss is the de-facto standard. Here, we propose an alternative loss, high error margin (HEM), which is a margin based loss modi(cid:28)ed to improve the training dynamics of neural networks. HEM loss is evaluated extensively using a wide range of DNN architectures and benchmark datasets with all experimental settings and training hyper-parameters taken from the literature, and hence, optimised for CE loss. HEM is found to be more e(cid:27)ective than CE loss across a range of image-based tasks: unknown class rejection, adversarial robustness, learning with imbalanced data, continual learning, and semantic segmentation (a pixel-wise classi- (cid:28)cation task). HEM is inferior to CE only in terms of clean and corrupt image classi(cid:28)cation with balanced training data, and this di(cid:27)erence is small. We also compare HEM to specialised losses that have previously been proposed to improve performance for speci(cid:28)c vision tasks. LogitNorm, a loss achieving state-of-the-art performance on unknown class rejection, produces similar performance to HEM for this task, but is much poorer for continual learning and semantic segmentation. Logit-adjusted loss, designed for imbalanced data, has superior results to HEM for that task, but performs worse on unknown class rejection and semantic segmentation. DICE, a popular loss for semantic segmentation, is inferior to HEM loss on all tasks, including semantic segmentation. Overall, HEM is competitive with the best alternative loss for all the tasks we have used and performs better than all other tested losses in terms of rejecting out-of-distribution examples, for continual learning, and by a substan- 1 tial margin for semantic segmentation . • • •

Object-centric models inspired by DETR have become thedominant paradigm for open-vocabulary video instance segmentation(OV-VIS). While recent efforts have reduced the computational cost ofpixel decoding, textual modality fusion, and object decoding to makethese architectures more suitable for mobile devices, real-time on-deviceinference at high frame rates remains an open challenge. In this paper,we introduce SegFS, a dual-stream fast-slow framework that significantlyimproves efficiency without sacrificing accuracy. On sparse keyframes, anopen-vocabulary object-based model predicts instance-level representa-tions. These representations are then projected back into the backbonefeature space to condition a lightweight fast network, which efficientlyrelocalizes and segments the instances in subsequent frames. By shiftinginstance propagation from object decoding to feature-space condition-ing, our approach decouples multimodal semantic understanding fromdense mask prediction and enables efficient temporal propagation. Theproposed fast branch achieves up to 14× lower latency than the mobile-oriented MOBIUS model, while maintaining competitive segmentationperformance on standard OV-VIS benchmarks.

Image forgery localization remains challenging due to di-verse manipulation techniques and distribution shifts. Existing recentforgery localization models achieve high accuracy on benchmarks butoften struggle with cross-domain generalization and robustness. In thispaper, we propose SARIF (Segment Anything for Robust Im-age Forensics), a framework that leverages Segment Anything Model(SAM), which has a promptable architecture and generalization abilityto overcome these limitations. SARIF introduces a feedback-guided maskdecoder and a dual-encoder design that extracts forgery-specific informa-tion to capture forensic traces while exploiting SAM’s architecture. Tolocalize manipulated regions, we design a block-wise prompting mecha-nism that derives forgery-specific cues from residual features between anadapted encoder and its frozen counterpart. These features are fused withthe previous mask prompt to drive a feedback-based mask refinementprocess, enabling automatic forgery segmentation without manual in-put. Extensive experiments on standard forgery-localization benchmarksshow that SARIF achieves strong average cross-dataset performance androbustness to common image corruptions. Our SARIF code is availablein GitHub Link.