ECCV 2026 论文列表

Although recent multimodal large language models (MLLMs)have advanced rapidly in vision–language reasoning, their capability bound-aries, particularly in cross-modal integration and reasoning, remain un-derexplored. Existing benchmarks primarily focus on evaluating uni-modal or loosely coupled multimodal abilities, leaving a gap in assess-ing complex cross-modal reasoning. To address this gap, we introduceVisWordBench, a benchmark comprising 2,625 English and 2,000 Chi-nese visual word puzzles with detailed human annotations. The bench-mark is designed to evaluate not only fundamental perceptual under-standing and world knowledge but also deep cross-modal reasoning skillsthat require consistently integrating visual and linguistic information.Through evaluations on VisWordBench, we find that current MLLMs ex-hibit under-diversified hypothesis search during reasoning. We thereforepropose a data construction pipeline and a training-free inference-timesteering strategy that promotes more diverse hypothesis exploration dur-ing CoT reasoning. Experimental results show consistent improvementsin reasoning quality and hypothesis diversity, supporting the utility ofVisWordBench for studying multimodal reasoning. The benchmark isavailable at: https://zenodo.org/records/20957955.

Generating high-fidelity remote sensing images is severelychallenged by their specific overhead perspectives, complex structures,and rigorous physical attributes. Although recent controllable RS gen-eration methods have made progress, they are still limited by two keybottlenecks. The first is semantic conflict, where dominant global conceptsoften overshadow fine-grained details. The second is physical inconsis-tency, which easily leads to structural collapse and radiometric distortionduring the generation process. To address these issues, we propose SGP2(Synergizing Geometric and Physical Priors), a novel coarse-to-fine frame-work for controllable multimodal RS image generation. Specifically, toresolve semantic conflicts, we introduce the Grassmann Miner, whichconstructs dynamic geodesic trajectories on the Grassmann Manifold,enabling the model to gradually shift its focus from global backgrounds tofine-grained local details during the denoising process. To ensure physicalconsistency, we propose Physical Prior Expert, a module that dynamicallyaligns multi-level features from Visual Foundation Models and employsa frequency-aware Spectral Loss to enforce both amplitude and phaseconsistency. Furthermore, to bridge RS data gaps, we present MMEarth-1.5M, a 1.5 million text-image dataset featuring strictly co-registeredoptical and SAR pairs for robust cross-modal simulation. Our approach,without bells and whistles, achieves favorable performance in controllablemulti-modal remote sensing image generation task. The dataset will beavailable in https://github.com/cpy0029/MMEarth-1.5M.

Current Large Vision Language Models (LVLMs) excel atmany zero-shot tasks like image captioning, visual question answeringand OCR. However, these same models suffer from poor performance atimage classification tasks, underperforming against CLIP-based meth-ods. Notably, this gap is surprising because many LVLMs use CLIP-pretrained vision encoders. Yet LVLMs are not inherently limited byCLIP’s architecture with independent vision and text encoders. In CLIP,this separation biases classification toward class-name matching ratherthan joint visual–text reasoning. In this paper we show that, despite theirpoor raw performance, LVLMs can improve visual feature class separa-bility at inference using prompt conditioning, and LVLMs’ internal rep-resentations, especially attention heads, can outperform the model itselfat zero-shot and few-shot classification. We introduce Head EnsembleClassifiers (HEC) to bridge the performance gap between CLIP-basedand LVLM-based classification methods. Inspired by Gaussian Discrimi-nant Analysis, HEC ranks the most discriminative vision and text headsand combines them into a training-free classifier. We show that HECachieves state-of-the-art performance in few-shot and zero-shot classifi-cation across 12 datasets. Code: github.com/AdhemarDeSenneville/HEC

Typical video object-centric learning (VOCL) approachesemploy slot-based frameworks that rely on reconstruction-driven en-coder–decoder architectures, where learning is mediated by two spatialmaps: attention maps from the encoder and object maps from the decoder.As these two distinct maps exhibit different properties, a recent densealignment strategy attempted to reconcile this discrepancy by enforcingagreement across all spatio-temporal patches via contrastive learning.However, this indiscriminate alignment inadvertently propagates the in-herent weaknesses of each module, such as noisy encoder predictionsand blurred decoder boundaries. Moreover, computing dense similaritiesacross all pairs incurs a computational cost quadratic in the total numberof spatio-temporal patches, severely limiting scalability. Motivated bythis, we propose Selective Synergistic Learning (SSync). Instead of ex-haustive patch-to-patch alignment, SSync prevents error propagation byselectively distilling only the most reliable cues: leveraging the encoderstrictly for boundary refinement and the decoder for interior denoising.This is realized via a pseudo-labeling with linear complexity, eliminatingthe need for quadratic spatial comparisons. Also, to prevent the rein-forcement of architectural biases like slot redundancy, we introduce atransitive pseudo-label merging that consolidates overlapping slots basedon spatio-temporal activation consistency. Extensive studies demonstratethat SSync improves decomposition quality and serves as a versatile,plug-and-play module while also exhibiting exceptional robustness to slotconfigurations. Code is available at github.com/wjun0830/SSync.

Recent advances in multimodal generative models have drivensubstantial improvements in image editing. However, current generativemodels still struggle with handling diverse and complex image editingtasks that require implicit reasoning, underscoring the need for a compre-hensive benchmark to systematically assess their performance across var-ious reasoning scenarios. Existing benchmarks primarily focus on single-object attribute transformation in realistic scenarios, which, while effec-tive, encounter two key challenges: (1) they largely overlook multi-objectinteractions as well as game-world scenarios that involve human-definedrules, which are common in real-life applications; (2) they only rely ontextual references to evaluate the generated images, potentially leadingto systematic misjudgments, especially in complex reasoning scenarios.To this end, this work proposes UniREditBench, a unified benchmarkfor reasoning-based image editing evaluation. It comprises 2,700 metic-ulously curated samples, covering both real- and game-world scenariosacross 8 primary dimensions and 18 sub-dimensions. To improve evalua-tion reliability, we introduce multimodal dual-reference evaluation, pro-viding both textual and ground-truth image references for each sampleassessment. Furthermore, we design an automated multi-scenario datasynthesis pipeline and construct UniREdit-Data-100K, a large-scalesynthetic dataset with high-quality chain-of-thought (CoT) reasoning an-notations. We fine-tune Bagel on this dataset and develop UniREdit-Bagel, demonstrating substantial improvements in both in-domain andout-of-distribution settings. Through thorough benchmarking of bothopen-source and closed-source image editing models, we reveal theirstrengths and weaknesses across various aspects.

Acquiring a complete set of magnetic resonance imaging (MRI)contrasts is time-intensive and uncomfortable for patients, despite thediagnostic value of multi-contrast imaging. This motivates synthesizingmissing contrasts from those already acquired, which is an inherently3D problem requiring anatomical coherence across axial, sagittal, andcoronal planes. However, fully 3D generative models are often impracti-cal under computational resources that scale cubically with volume size.We propose a unified Multi-Plane Autoregressive Diffusion (MPAD), alatent diffusion framework that achieves full-volume 3D synthesis usingefficient plane-wise 2D operations while preserving volumetric coherence.A 3D autoencoder first compresses MRI scans into an isotropic 3D la-tent representation. A 2D diffusion model is then trained to reconstructmasked latent slices of the target contrast, conditioned on both source-contrast slices and unmasked target-contrast slices. During inference,we introduce plane-wise autoregressive synthesis with inter-plane priors.Slices are generated autoregressively in random order within one planeorientation to maintain intra-plane continuity, then propagated as con-ditioning priors to orthogonal plane orientations to enforce inter-planeconsistency. Compared to 3D latent diffusion baselines, MPAD reducestraining and inference FLOPs by 7× and 3×, respectively, while alsolowering inference time and peak memory consumption. Experimentson multiple datasets demonstrate that MPAD achieves superior perfor-mance, generating high-fidelity 3D volumes and supporting one-to-manytranslation within a single unified model.

With the advancement of imaging technology, ultra-highdefinition images have become increasingly essential in modern visual applications. However, existing multi-focus image fusion remains largely confined to low-resolution images and faces three major barriers in UHD scenarios, namely data availability, model adaptability, and deployment feasibility, which severely hinder its practical application. To shatter these barriers, first, we propose the UHD-MFF dataset, the first largescale ultra-high-resolution multi-focus fusion dataset. Second, we propose a scale-specialized lookup-table framework tailored for ultra-highresolution images, termed as UMF-LUT. It consists of Coarse-Region Lookup Table (C-LUT) and Detail-Edge Lookup Table (D-LUT). Specifically, C-LUT performs joint queries of multiple gradient cues and semantic cues at low-resolution scales to enable region-level decision-making. Also, D-LUT operates at high-resolution scales, leveraging efficient Laplacian cues to provide complementary edge-level decision information. Such a design makes the model particularly well-suited for ultra-high-resolution multi-focus image fusion. Finally, it offers strong deployability with minimal computational overhead, enabling real-time 4K multi-focus fusion and showing promising potential for smartphone. Extensive experiments demonstrate that it outperforms SOTA methods in both visual fidelity and quantitative metrics. It effectively advances the development of multifocus image fusion toward ultra-high-resolution imaging scenarios. The code is available at https://github.com/zyb5/UHD-MFF.

Enabling embodied agents to imagine future states is essen-tial for robust and generalizable visual navigation. Yet, state-of-the-artsystems typically rely on modular designs that decouple navigation plan-ning from visual world modeling, which often induces state–action mis-alignment and weak adaptability in novel or dynamic scenarios. We pro-pose UniWM, a unified, memory-augmented world model that integratesegocentric visual foresight and planning within a single multimodal au-toregressive backbone. UniWM explicitly grounds action selection in vi-sually imagined outcomes, tightly aligning prediction with control. Mean-while, a hierarchical memory mechanism fuses short-term perceptual cueswith longer-term trajectory context, supporting stable and coherent rea-soning over extended horizons. Extensive experiments on four challeng-ing benchmarks (Go Stanford, ReCon, SCAND, HuRoN) and the 1XHumanoid Dataset show that UniWM improves navigation success ratesby up to 30%, substantially reduces trajectory errors against strong base-lines, generalizes zero-shot to the unseen TartanDrive dataset, and scalesnaturally to high-dimensional humanoid navigation. These results po-sition UniWM as a principled step toward unified, imagination-drivenembodied navigation. All materials are committed to be open-sourced.

In real scenarios, the videos can span several minutes or evenhours, yet existing research on spatio-temporal video grounding (STVG),given a textual query, mainly focuses on localizing the target from a videoof tens of seconds, typically less than one minute, hindering its applica-tions. In this paper, we explore Long-Form STVG (LF-STVG), whichaims to locate the target from long-term videos. In LF-STVG, long-termvideos encompass a much longer temporal span and more irrelevant infor-mation, making it challenging for current short-form STVG approachesthat process all the frames at once. Addressing these, we propose a novelAutoRegressive Transformer architecture for LF-STVG, dubbed ART-STVG. Unlike current STVG methods requiring seeing the entire videosequence to make a full prediction at once, our ART-STVG regards thevideo as a streaming input and processes its frames sequentially, makingit capable of easily handling the long videos. To capture spatio-temporalcontext in ART-STVG, spatial and temporal memory banks are devel-oped and applied to decoders of ART-STVG. Considering that memoriesat different moments are not always relevant for localizing the target incurrent frame, we introduce simple yet effective memory selective strate-gies that enable the more relevant information for decoders, greatly im-proving the performance. Moreover, rather than parallelizing spatial andtemporal localization as done in existing approaches, we introduce a novelcascaded spatio-temporal design that connects spatial decoder to tempo-ral decoder during grounding, which allows ART-STVG to leverage morefine-grained target information to assist with complicated temporal lo-calization in complex long videos, further boosting performance. On thenewly extended datasets for LF-STVG, ART-STVG largely outperformscurrent approaches, while showing competitive results on Short-FormSTVG. Our code is at: xxx.

Generative flow and diffusion models provide the continu-ous, multimodal action distributions needed for high-precision roboticpolicies. However, their reliance on iterative sampling introduces severeinference latency, degrading control frequency and harming performancein time-sensitive manipulation. To address this problem, we proposethe One-Step Flow Policy (OFP), a from-scratch self-distillation frame-work for high-fidelity, single-step action generation without a pre-trainedteacher. OFP unifies a self-consistency loss to enforce coherent transportacross time intervals, and a self-guided regularization to sharpen predic-tions toward high-density expert modes. In addition, a warm-start mech-anism leverages temporal action correlations to minimize the generativetransport distance. Evaluations across 56 diverse simulated manipulationtasks demonstrate that a one-step OFP achieves state-of-the-art results,outperforming 100-step diffusion and flow policies while accelerating ac-tion generation by over 100×. We further integrate OFP into the π0.5model on RoboTwin 2.0, where one-step OFP surpasses the original 10-step policy. These results establish OFP as a practical, scalable solutionfor highly accurate and low-latency robot control.

We present SynCity 3000, a framework for generating 3Dscenes that are globally coherent while enabling fine-grained layout con-trol. Building on the ability of current image-to-3D generators to pro-duce complex 3D assets from a single image, we extend this capabilityto the scale of entire scenes by adapting the generator to be applicableas a convolutional operator. We achieve this by fine-tuning the modelon scene-like data generated by a new synthetic data engine, which wepropose to address the scarcity of 3D scene data for training. The convo-lutional generator is then applied to a dimetric image of the entire scene,generated from the user prompt, resulting in 3D scenes of arbitrary sizeand complexity. Across diverse prompts and layouts, SynCity 3000 pro-duces large, coherent, and detailed scenes, addressing the shortcomingsof prior approaches to 3D scene generation.

Autoregressive diffusion models have shown remarkable prog-ress in long video generation by producing frames causally. To boundmemory and computational cost, these models rely on a fixed-size sliding-window KV cache. While attention sinks retain a few initial frames astemporal anchors, the vast majority of historical context is irrecover-ably discarded as the window advances, severing long-range dependen-cies essential for scene consistency and narrative coherence. Meanwhile,the fixed multi-step denoising schedule applies uniform effort to everychunk, ignoring the varying demands imposed by diverse scene dynamicsacross temporal segments. In this paper, we propose Surprise Forcing,a training-free framework that addresses both bottlenecks. We introducea Surprise-Gated Memory Bank that scores each evicted frame by† Corresponding authors.its aggregate divergence from the bank and nearest-neighbor novelty,admitting only the most informative frames via a budget-normalizedgate that adapts to video dynamics. A priority-based eviction policykeeps the bank compact yet informative, while dynamic routing retrievesthe most relevant historical frames at attention time. We further pro-pose Surprise-Aware Denoising, which uses intra-chunk cosine dis-tance among adjacent frames after a single denoising step as a self-referential difficulty predictor, enabling adaptive step reduction for easychunks while preserving full refinement for complex ones. Experimentson VBench show that Surprise Forcing maintains robust narrative coher-ence in long videos without compromising real-time responsiveness.

While recent diffusion models have achieved remarkable suc-cess in visual content editing, extending them to stereo image editingremains highly challenging due to the lack of explicit stereo geomet-ric consistency alignment, which leads to cross-view inconsistencies andgeometric distortions. We propose StereoEdit, a training-free diffusionframework designed to achieve geometrically consistent and structurallycoherent stereo editing. StereoEdit introduces two complementary mod-ules that explicitly enforce geometric reliability during the denoising pro-cess. The Geometrically-Gated Attention Reference (GGAR) adaptivelyfilters cross-view attention based on disparity-aware confidence, ensur-ing stable correspondences across epipolar regions, while the Geometry-Aware Latent Alignment (GALA) performs confidence-weighted, bidi-rectional latent corrections to maintain parallax alignment throughoutthe diffusion process. Together, these modules enable robust and consis-tent stereo edits. Extensive experiments on diverse stereo benchmarksdemonstrate that StereoEdit significantly improves cross-view coherenceand visual realism, establishing a new foundation for controllable andimmersive stereoscopic content editing.

Recent advances in video diffusion models have greatly im-proved visual fidelity, yet their generated motions often violate physicalplausibility. We observe a common kinematic failure, “motion entangle-ment”, the unintended coupling of independent motion sources, such ascamera movement and object motion. We identify that this issue stemsfrom data bias and the reconstruction-based training design of diffusionmodels. Training on noisy videos that still retain coarse motion cues in-advertently encourages the model to replicate existing motion without anincentive to learn how to model kinematically-grounded motions. To ad-dress this, we propose a Self-Imagination Fine-Tuning (SIFT) paradigm,which enables the model to learn from its own generated videos ratherthan directly reconstructing real ones, breaking the reconstruction short-cut. We further employ motion-aware discriminative supervision and aprogressive hard-case replay strategy to stabilize and accelerate learning.By leveraging freely-generated text prompts, our method can denselycover a broad motion space, including rare or finely-disentangled scenar-ios that would be costly to collect as video data. Extensive experimentsdemonstrate that our approach substantially improves the physical real-ism, motion disentanglement, and controllability of generated videos.

Modern Video Object Segmentation (VOS) involves track-ing and segmenting user-specified targets. While recent approaches haveachieved remarkable performance in single-target scenarios, extendingthem to multi-target settings typically involves replicating the single-target processing for each individual object, resulting in reduced framerates (FPS) with unbounded latency as target density scales. Built uponSegment Anything 2 (SAM2), we propose SAM-MT, which addressesthis by transforming the model into an interactive framework for real-time Multi-Target video segmentation. SAM-MT uses explicit queriespropagated and updated across frames to represent different individualtargets, in parallel with a shared representation for global context. It em-ploys decoupled masked attention to keep individual identities distinctfrom cross-target interference, and sparse memory for stable temporalevolution, along with specialized strategies for occlusion handling andoverlap prevention. SAM-MT successfully decouples latency from thenumber of targets, achieving real-time speed on par with single-targetbaselines (>36 FPS for 10 targets) while maintaining SAM2’s robustvideo segmentation performance.

Text-to-image (T2I) diffusion models have achieved strik-ing progress but still struggle to synthesize rare concepts involving un-usual attribute-object pairings, often resulting in concept omission orsemantic drift where a dominant entity overwhelms the generation. Trac-ing these failures to a lack of compositional balance during the denois-ing trajectory, we propose RADIANCE, a training-free framework thattreats inference as a closed-loop feedback process. RADIANCE augmentspretrained backbones with three modular components: (1) a Composi-tional Similarity Monitor (CSM) that tracks the emergence of objectsand attributes in intermediate latents via CLIP-based feedback; (2) aBidirectional Scale Controller (BSC) that applies a reactive "restoringforce" using positive and negative IP-Adapter scales to rebalance biasedtrajectories; and (3) a Feedback Guidance Scheduler (FGS) that coordi-nates these updates across timesteps without additional training. We fur-ther extend the framework to multi-object prompts via Delayed AdapterActivation (DAA) and Layer-wise Alternating Guidance (LAG) to pre-vent premature concept fusion. By overlapping monitoring and denoisingthrough pipelined execution, RADIANCE maintains competitive latencywhile significantly enhancing the per-sample success rate and effectivethroughput. Experiments on RareBench and T2I-CompBench demon-strate that RADIANCE consistently enhances compositional alignmentand perceptual quality over state-of-the-art baselines.

High-resolution 3D asset generation is vital in various 3D ap-plications. Existing state-of-the-art diffusion-based models remain con-strained by fixed resolutions, limiting their ability to produce details. Inthis paper, we tackle the challenge of generating more detailed, higher-resolution 3D objects by introducing a 3D super-resolution (SR) frame-work built on existing 3D generative foundation models. To this end,we design PLSR, a progressive and localized super-resolution solutionto achieve this goal effectively and memory efficiently. Technically, givena coarse geometry from a pretrained 3D generator, we decompose theglobal SR task into localized sub-tasks via an associative input de-composition scheme, adapt a flow-based 3D generator into a local-ized super-resolution model through low-cost finetuning, and unifythem in an iterative patch-wise denoising pipeline for seamless high-resolution output. Experiments on challenging objects show that our ap-proach is able to generate 3D details with new strong fine-detail fidelitywhile significantly reducing the computational cost, offering a new andpractical solution for high-resolution 3D asset generation.

World models enable agents to reason about future outcomesand learn policies from their knowledge of state transition, but existingapproaches primarily focus on reconstructing future observations or fea-tures, which introduces unnecessary complexity and limits their effec-tiveness for decision making. In this work, we propose a compatibilityprediction Latent World Model (LWM) for robot navigation that pre-dicts action-conditioned latent feature compatibility rather than recon-structing observations. Our key insight is that spatial proximity corre-lates with latent feature similarity, enabling action consequences to beevaluated directly in latent space. To support counterfactual training,our model leverages action sequences sampled across trajectories andlearns to predict which sequences lead closer to the goal. Furthermore,we demonstrate how the learned world model can supervise policy learn-ing from unlabeled video data and further improve policies through re-inforcement learning entirely within the world model. This imagination-driven framework eliminates the need for action annotations and addi-tional environment interaction. Extensive experiments on multiple real-world robot navigation datasets show that our approach significantlyoutperforms prior world model and imitation learning methods in pre-diction accuracy, policy learning, and real-world navigation performance.The code, pretrained models, and additional materials are available athttps://wzm206.github.io/latent-world-model-nav/.

Affordance prediction serves as a critical bridge between perception and action in embodied AI. However, existing research is confined to pinhole camera models, which suffer from narrow Fields of View (FoV) and fragmented observations, often missing critical holistic environmental context. In this paper, we present the first exploration into Panoramic Affordance Prediction, utilizing 360-degree imagery to capture global spatial relationships and holistic scene understanding. To facilitate this novel task, we first introduce PAP-12K, a largescale benchmark dataset containing over 1,000 ultra-high-resolution (12k, 11904×5952) panoramic images with over 12k carefully annotated QA pairs and affordance masks. Furthermore, we propose PAP, a trainingfree, coarse-to-fine pipeline inspired by the human foveal visual system to tackle the ultra-high resolution and severe distortion inherent in panoramic images. PAP employs recursive visual routing via grid prompting to progressively locate targets, applies an adaptive gaze mechanism to rectify local geometric distortions, and utilizes a cascaded grounding pipeline to extract precise instance-level masks. Experimental results on PAP-12K reveal that existing affordance prediction methods designed for standard perspective images suffer severe performance degradation and fail due to the unique challenges of panoramic vision. In contrast, PAP framework effectively overcomes these obstacles, significantly outperforming state-of-the-art baselines and highlighting the immense potential of panoramic perception for robust embodied intelligence.