ECCV 2026 论文列表

Vision-Language Models (VLMs) frequently `hallucinate'_x0015_generate plausible yet factually incorrect statements_x0016_posing a criticalbarrier to their trustworthy deployment. In this work, we propose a newparadigm for diagnosing hallucinations, recasting them from static out-put errors into an auditable process-level anomaly. Our framework isgrounded in a normative principle of computational rationality, allowingus to model a VLM's generation as a dynamic cognitive trajectory. Wedesign a suite of information-theoretic probes that project this trajec-tory onto an interpretable, low-dimensional Cognitive State Space.Our central discovery is a governing principle we term the geometric-information duality: a cognitive trajectory's geometric abnormalitywithin this space is fundamentally equivalent to its high information-theoretic surprisal. Hallucination detection is thus elegantly re-framedas a geometric anomaly detection problem. Evaluated across diversesettings_x0016_from rigorous binary QA (POPE) and comprehensive reason-ing (MME) to unconstrained open-ended captioning (MS-COCO)_x0016_ourframework achieves state-of-the-art performance. Crucially, it operateswith high e_x001E_ciency under weak supervision and remains highly robusteven when calibration data is heavily contaminated. This approach en-ables a causal attribution of failures, mapping observable errors to dis-tinct pathological states: perceptual instability (measured by Percep-tual Entropy, HEvi ), logical-causal failure (measured by InferentialCon_x001D_ict, SConf ), and decisional ambiguity (measured by Decision En-tropy, HAns ). Ultimately, this opens a path toward building AI systemswhose reasoning is transparent, auditable, and diagnosable by design.

While many vision language models (VLMs) are developed to passively answer well-defined, straightforward questions with highly specified targets, as in most benchmarks, they often struggle in practice with complex open-ended tasks, which usually require multiple rounds of active exploration and reasoning in the visual space. Such visual thinking paths not only provide step-by-step exploration and verification as an AI detective but also produce better interpretations of the final answers. However, these paths are challenging to evaluate due to the large exploration space of intermediate steps. To bridge the gap, we develop an evaluation suite, “Visual Reasoning with multi-step EXploration (V-REX)”, which is composed of a benchmark of challenging visual reasoning tasks requiring native multi-step exploration and an evaluation protocol. V-REX casts the multi-step exploratory reasoning into a Chain-of-Questions (CoQ) and disentangles VLMs’ capability to (1) Planning: determining the exploratory direction to gather information at each step for solving the task; and (2) Following: answering curated CoQ sequentially to collect information for deriving the final answer. By curating finite options of questions and answers per step, V-REX achieves a reliable quantitative and finegrained analysis of the intermediate steps. By assessing SOTA proprietary and open-sourced VLMs, we reveal consistent scaling trends, significant differences between planning and following abilities, and substantial room for improvement in multi-step exploratory reasoning.

Real-world image super-resolution (Real-SR) has been rev-olutionized by leveraging the powerful generative priors from DiffusionModels (DMs) and Flow Matching (FM). However, existing one-stepmethods typically replace Gaussian noise with degraded low-resolution(LR) latents at initialization, introducing a substantial distribution shiftthat further leads to trajectory deviation and prior collapse under ex-treme acceleration. To overcome these limitations, we propose Allo{SR}2 ,a novel FM-based framework that rectifies one-step SR flows via allomor-phic generative flows to maintain high-fidelity generative realism. Specifi-cally, we utilize SNR-Guided Trajectory Initialization to identify a statis-tically aligned intermediate state along the pre-trained path to integrateLR representations into the generative flow. To ensure a stable, low-curvature path for one-step inference, we propose Flow-Anchored Trajec-tory Consistency (FATC), which explicitly regularizes the velocity fieldof the underlying probability flow. Furthermore, we develop AllomorphicTrajectory Matching (ATM), a self-adversarial distillation strategy thatjointly models the SR flow and the generative flow within a unified ve-locity field, enabling one-step Real-SR while preserving the generativeprior. Extensive experiments on both synthetic and real-world bench-marks demonstrate that Allo{SR}2 achieves state-of-the-art performancein one-step Real-SR, offering a superior balance between fidelity and re-alism while maintaining extreme efficiency.

This paper introduces Actor as Its Own Critic, a unifiedreinforcement learning framework, Cycle Group Relative Policy Opti-mization (CycleGRPO), that jointly optimizes region understanding andlocalization for Multimodal Large Language Models (MLLMs). Unlikeexisting separate pipelines, we leverage the inherent duality between thetwo tasks to construct a self-evaluating reinforcement learning paradigm:“region → text → region”. Specifically, a single MLLM first acts as theactor to generate region captions, then immediately transitions to a criticto ground its generated text back in the spatial domain. Therefore, Cy-cleGRPO requires only region inputs, e.g., masks or bounding boxes,entirely bypassing the need for textual ground truths. A quality-awaretoken-level cycle-consistency reward is employed to assess the seman-tic discriminability of text captions via their physical localization ac-curacy. Empirically, built upon SAMTok, our CycleGRPO frameworksuccessfully bootstraps both capabilities simultaneously. Without anytask-specific fine-tuning, the framework yields consistent performancegains across a wide range of benchmarks, including region captioning,region VQA, grounded dialogue, and referring segmentation. Overall,CycleGRPO offers a straightforward and scalable way to advance pixel-level capabilities in MLLMs. Code and models are released at https://github.com/devinxzhang/CycleGRPO.

Model editing promises a fast, targeted way to correct postdeployment mistakes in medical vision–language models (VLMs) without costly retraining. However, existing multimodal model editing benchmarks focus on general-purpose tasks and do not reflect realistic clinical domain requirements and variability. To address this, we introduce M3Bench, a clinically grounded benchmark for multimodal model editing that evaluates whether an edit remains reliable, precise and generalizable under the challenges of image and text variation, modality and protocol shifts, clinical knowledge composition, and temporal progression. M3Bench contains 16,276 questions spanning diverse anatomy, modalities, and specialties, and supports both single and sequential edits: Evaluating 4 representative editors across 6 medical and general VLMs, we indicate that no method excels across all criteria. Gradient-based editors achieve strong transfer but suffer from catastrophic locality violations, whereas memory-based methods preserve locality but lack compositional generality and exhibit high backbone-dependent hyperparameter sensitivity. We further attribute these failures to the latent space geometry of VLMs and how different editing methods shift its landscape. Overall, M3Bench establishes a rigorous clinical stress test for multimodal model editing and offers actionable guidance for safer post-deployment adaptation. The benchmark is publicly available at https://github.com/BioMed-AI-Lab-U-Michgan/M3Bench.

Generative world models hold immense promise as scalablesimulators for autonomous systems, particularly for synthesizing rare butsafety-critical multi-agent interactions, such as vehicle collisions. How-ever, current evaluation paradigms index heavily on visual fidelity and se-mantic alignment, leaving a critical blind spot: they cannot reliably quan-tify whether generated dynamics actually obey the fundamental physicallaws required for reliable simulation. Assessing this physical plausibilityis inherently difficult due to a lack of physical metrics and the challengeof extracting metric-scale kinematics from uncalibrated video rollouts.To bridge this gap, we introduce CrashTwin, a physics-grounded eval-uation framework designed to stress-test the physical trustworthiness ofworld models. CrashTwin couples a diverse dataset of multi-agent colli-sion scenarios, comprising 25K controllable synthetic and 12K in-the-wildreal-world collision sequences with a novel calibration-free reconstruc-tion pipeline, enabling the recovery of 3D physical attributes directlyfrom world model rollouts. We propose a diagnostic suite that systemat-ically evaluates three dimensions: spatio-temporal consistency, momen-tum and kinetic energy conservation, and world-dynamics integrity. Ex-tensive benchmarking of state-of-the-art models reveals a crucial insight:high perceptual quality frequently masks severe physical violations dur-ing complex interactions. By quantitatively exposing these failure modes,CrashTwin provides a vital diagnostic tool for developing physicallygrounded world models capable of reliable real-world simulation. Codeand dataset are available at: https://github.com/phai-lab/CrashTwin.

Humans can navigate an unfamiliar city and gradually forma coherent spatial mental map spanning tens of square kilometers. CanAI build spatial representations at a comparable scale? Although recentfoundation models have advanced scene reconstruction and embodied in-telligence, scaling to entire cities remains an open challenge, primarilydue to the lack of city-scale data. To bridge the gap, we introduce Wild-City, a real-world multimodal dataset collected by autonomous fleetstraversing complex urban environments. Our dataset includes 18 tra-jectories, each averaging 83.7 kilometers in length, and preserves thecore challenges of in-the-wild perception, e.g., dynamic objects, lightingvariations, and imperfect camera poses. We further establish an urban-tailored reconstruction baseline and convert the reconstructed environ-ments into a closed-loop simulator. Beyond the dataset and baseline, wesystematically analyze the key challenges on the path to simulation-readyurban digital twins: scalability, extrapolation, and uncertainty. Ul-timately, WildCity aims to catalyze progress not only in city-scale render-ing, but more broadly in the pursuit of AI that can perceive, remember,and reason across space at a scale comparable to human cognition.

Accurately recovering the camera’s calibration and metricposes for any unconstrained video would unlock large-scale training datafor navigation and scene understanding. The dominant approaches tothis problem are severely limited: Simultaneous Localization and Map-ping (SLAM) is sensitive to initialization and transient failures due to itscausal, incremental nature; it is often over-optimized for real-time oper-ation and generally requires known camera calibration; while Structure-from-Motion (SfM) typically forgoes any image ordering, enabling opti-mal initialization and global optimization, but lacks robustness to visualsymmetries and extreme motions. To bridge this gap, we introduce asystem that combines the strong sequential constraints of SLAM withthe flexibility and global optimization of offline SfM, enabling the met-ric reconstruction of arbitrary, long, uncalibrated videos. This systemleverages recent advances in wide-baseline dense image matching, treatstemporal ordering as a first-class citizen for reliable loop closure, andaugments global optimization with metric monocular depth priors. Asa result, thorough evaluations on diverse, challenging datasets that ex-hibit extreme motion and visual symmetries reveal that our approach issignificantly more robust and accurate than both state-of-the-art SLAMand SfM, classical or learned, with given or unknown camera calibration.

Unified multimodal models (UMMs) have shown great promise in integrating understanding and generation across diverse modalities. However, existing research rarely extends this paradigm to the tactile domain, where both object-level semantics and sensor-level configurations jointly determine the meaning of touch. To address this gap, we propose UniTac, the first UMM designed for tactile understanding and generation. UniTac models the tactile process as a transition from noncontact to contact, capturing the physical interaction between sensors and objects through a dual-level representation that encodes both sensor and object attributes. For tactile understanding, UniTac introduces two tasks, object property description and sensor identification, to enhance reasoning over physical and cross-sensor information. For tactile generation, we design a two-stage training paradigm consisting of reconstruction and alignment, together with a sensor-prior-based sampling strategy that simulates realistic tactile contact. Trained on large-scale multi-sensor datasets, UniTac achieves state-of-the-art performance in tactile understanding and generates realistic tactile signals across sensors.

Camera-controllable view synthesis aims to synthesize novelviews of a given scene under varying camera poses while strictly preservingcross-view geometric consistency. However, existing methods typicallyrely on fragmented geometric guidance, such as only injecting point cloudsat the representation level despite models containing multiple levels, andare mainly based on image diffusion models that operate on discreteview mappings. These two limitations jointly lead to geometric drift andstructural degradation under continuous camera motion. We observethat while leveraging video models provides continuous viewpoint priorsfor camera-controllable view synthesis, they still struggle to form stablegeometric understanding if geometric guidance remains fragmented. Tosystematically address this, we inject unified geometric guidance across thethree levels that jointly determine the generative output: representation,architecture, and loss function. To this end, we propose UniGeo, a novelcamera-controllable editing framework. Specifically, at the representationlevel, UniGeo incorporates a frame-decoupled geometric reference injectionmechanism to provide robust cross-view geometry context. Furthermore,at the architecture level, it introduces a geometric anchor attention toalign multi-view features, and at the loss function level, it proposes atrajectory-endpoint geometric supervision strategy to explicitly reinforcethe structural fidelity of target views. Experiments across multiple publicbenchmarks, encompassing both extensive and limited camera motionsettings, demonstrate that UniGeo significantly outperforms existingmethods in visual quality and geometric consistency.

Automatically generating pedestrian pathways from aerialimages requires producing a connected network suitable for routing, notjust detecting where sidewalks appear. Sidewalks and crossings, in con-trast to roads, may be partially occluded, implicitly defined, and exhibitcomplex connectivity patterns. Existing segmentation-based approachesfocus on labeling pixels to infer segments, but often produce disconnectedor fragmentary graphs that are unreliable for navigation. We introduceTraversRL, a vision-conditioned model that iteratively grows a path-way network from an aerial image, simulating a traveler navigating thebuilt environment. TraversRL uses an action space of short and long di-rection–distance segments designed to adapt to complex patterns andspan occlusions, and uses a combination of graph-level and step-wise re-wards to balance overall connectivity with precise edge placement. Acrossthree visual backbones and three intersection datasets, TraversRL sub-stantially improves buffered IoU with the ground-truth graph relative toa state-of-the-art segmentation baseline, and more than doubles metricsof connectivity. Moreover, combining global and local rewards producescleaner graphs with fewer spurious branches while further improvingoverall performance. These results demonstrate that modeling pathwayextraction as a sequential decision process from the perspective of a trav-eler, while optimizing for final graph quality with reinforcement learning,produces significantly more reliable pedestrian networks.

Event cameras capture per-pixel brightness changes with mi-crosecond resolution, offering continuous motion information lost be-tween RGB frames. However, existing event-based motion estimatorsdepend on large-scale synthetic data that often suffers from a signifi-cant sim-to-real gap. We propose TETO (Tracking Events with TeacherObservation), a teacher-student framework that learns event motionestimation from only ∼25 minutes of unannotated real-world record-ings through knowledge distillation from a pretrained RGB tracker. Ourmotion-aware data curation and query sampling strategy maximizes learn-ing from limited data by disentangling object motion from dominantego-motion. The resulting estimator jointly predicts point trajectoriesand dense optical flow, which we leverage as explicit motion priors tocondition a pretrained video diffusion transformer for frame interpola-tion. We achieve state-of-the-art point tracking on EVIMO2 and opticalflow on DSEC using orders of magnitude less training data, and demon-strate that accurate motion estimation translates directly to superiorframe interpolation quality on BS-ERGB and HQ-EVFI.

Deepfake detection faces challenges in generalization, prompt-ing recent studies to fine-tune pretrained Vision Transformers for im-proved performance. Existing fine-tuning-based detection approaches of-ten assign the same rank to the tunable residual components of alllayers. However, the shallow layers exhibit a distinctly low-rank struc-ture, whereas the middle and deep layers present more complex struc-tures. Therefore, we argue that only slight fine-tuning is required forthe shallow layers, which are able to effectively capture low-level forgerycues, while the middle and deep layers require higher residual ranks tomodel long-range dependencies and extract high-level forgery artifacts.To achieve these goals adaptively, we propose the Stable-Rank-BasedResidual Adaptation (SRRA) strategy. It utilizes singular value decom-position to decompose the pretrained weight matrices and adaptively al-locates frozen principal components and learnable residual componentsbased on the Stable Rank. In addition, we design the Residual En-ergy Constraint (REC) and Residual Subspace Orthogonalization (RSO),which suppress excessive perturbations and promote subspace decou-pling, avoiding overfitting and improving the generalization performance.Extensive experiments demonstrate that SRRA achieves remarkable im-provements in cross-domain generalization across multiple deepfake de-tection benchmarks. Our code is available at https://github.com/LHK-CodeLab/SRRA.

This paper tackles the challenge of simultaneously manipu-lating multiple objects in image editing—a scenario that extends beyondthe capabilities of existing methods focused on single-object manipula-tion. Our approach SIMON efficiently handles complex rearrangementssuch as object shuffling, achieving higher visual quality at both image andregion levels, while reducing inference time. To this end, we propose threekey modules: a content-aware attention mechanism for region-adaptivefocus, a multi-object energy guidance strategy for subtask-specific con-sistency, and a latent initialization technique for artifact suppression.Together, these components enhance visual coherence across inpainting,object relocation, and background preservation subtasks. To evaluate ourframework, we curate 1,000 images from the 3D-FUTURE dataset, eachcontaining 2 to 5 target objects. For every object, we annotate a newtarget location, ensuring that the resulting layouts are physically plausi-ble and semantically meaningful. We conduct comparative experimentsagainst three state-of-the-art methods, using both image-level, region-level, and instance-level metrics. Both quantitative results and humananalysis confirm that our framework delivers significant improvementsin both visual quality and computational efficiency. We further validateour design through extensive ablation studies with diverse architecturesand applications. The code and annotated dataset are publicly avail-able at https://github.com/yijichar/SIMON-SImultaneous-Multi-Object-Navigation.

Image-conditioned video diffusion models achieve impressivevisual realism but often suffer from weakened motion fidelity, e.g., re-duced motion dynamics or degraded long-term temporal coherence, es-pecially after fine-tuning. We study motion alignment in video diffusionmodels post-training. To address this, we introduce pixel-motion rewardsbased on pixel flux dynamics, capturing both instantaneous and long-term motion consistency. We further propose Smooth Hybrid Fine-tuning(SHIFT), a scalable reward-driven framework that unifies supervisedfine-tuning and advantage-weighted fine-tuning. Benefiting from noveladversarial advantages, SHIFT improves convergence speed and miti-gates reward hacking. Experiments show that our approach efficientlyresolves dynamic-degree collapse in modern video diffusion models su-pervised fine-tuning. Project page: https://xiye20.github.io/projects/SHIFT/.

This paper presents SOMO, a scalable framework for crossembodiment grasp synthesis that transfers to novel robot hands using only their hand description (i.e., a Unified Robot Description Format (URDF) file), without requiring any hand–object interaction annotations. Unlike prior approaches that rely on hand-specific models or annotated grasp data for each embodiment, SOMO introduces a shared Morphology-Prior Diffusion model applicable across heterogeneous hands through three key designs. First, grasping is formulated as a 3D assembly problem by predicting per-link SE(3) poses, enabling geometry-driven generation independent of hand-specific kinematics, with feasibility enforced through a post joint optimization stage. Second, a classifier-free training strategy learns a morphology prior from physically valid hand configurations generated through forward-kinematic exploration without object–grasp annotations, enabling grasp synthesis for unseen robot hands given only their URDF files. Third, a 3D shape-aware VAE encodes link and object geometry into a shared embedding, enabling consistent reasoning about hand–object complementarity across embodiments. Experiments show that SOMO achieves state-of-the-art grasp synthesis across six robot hands and demonstrates strong annotation-free generalization to previously unseen hands. Code and models will be released soon.

Generating high-quality 3D characters from single images re-mains a significant challenge in digital content creation, particularly dueto complex body poses and self-occlusion. In this paper, we present RCM(Rotate your Character Model ), an advanced image-to-video diffusionframework tailored for high-quality novel view synthesis (NVS). Com-pared to existing diffusion-based approaches, RCM offers several keyadvantages: (1) transferring characters with any complex poses into acanonical pose, enabling consistent novel view synthesis across the entireviewing orbit, (2) high-resolution orbital video generation at 1024×1024resolution, (3) controllable observation positions given different initialcamera poses, and (4) multi-view conditioning supporting up to 4 inputimages, accommodating diverse user scenarios. Extensive experimentsdemonstrate that RCM outperforms state-of-the-art methods in novelview synthesis on character generations. The deliverables will be updatedhere.

Video understanding has largely relied on deep spatiotem-poral architectures, including 3D convolutional networks and opticalflow (OF) based models. While effective, these methods are often com-putationally expensive and depend on heuristic motion representationsthat are sensitive to illumination, scale, and structural changes. To ad-dress these limitations, we propose ReynoldsFlow, a physics-inspiredrepresentation grounded in the Reynolds transport theorem (RTT) andHelmholtz-Hodge decomposition (HHD). ReynoldsFlow decomposes mo-tion into curl-free (CF) and divergence-free (DF) components, providinga principled and interpretable characterization of scene dynamics. Bycoupling intensity information with decomposed motion cues, it producesdynamics-aware, texture-preserving features that boost downstream taskssuch as pose estimation, action recognition, and tiny object detection.Lightweight and modular, ReynoldsFlow can be readily integrated intoexisting architectures. Experiments across diverse benchmarks show thatReynoldsFlow consistently matches or surpasses existing approaches, of-fering improved generalizability and computational efficiency.

The rapid progress of visual generative models has madeAI-generated images increasingly difficult to distinguish from authen-tic ones, posing growing risks to social trust and information integrity.This motivates detectors that are not only accurate but also forensicallyexplainable. While recent multimodal approaches improve interpretabil-ity, many rely on post-hoc rationalizations or coarse visual cues, withoutconstructing verifiable chains of evidence, thus often leading to poor gen-eralization. We introduce REVEAL-Bench, a reasoning-enhanced multi-modal benchmark for AI-generated image forensics, structured aroundexplicit chains of forensic evidence derived from lightweight expert mod-els and consolidated into step-by-step chain-of-evidence traces. Basedon this benchmark, we propose REVEAL (Reasoning-enhanced Foren-sic Evidence Analysis), an explainable forensic framework trained withexpert-grounded reinforcement learning. Our reward design jointly pro-motes detection accuracy, evidence-grounded reasoning stability, and ex-planation faithfulness. Extensive experiments demonstrate significantlyimproved cross-domain generalization and more faithful explanations tobaseline detectors. All data and source codes are publicly available at:https://github.com/TrustMedia-zju/REVEAL.