ECCV 2026 Papers

Despite recent advancements in 3D generation, models oftenstruggle to produce assets with high visual fidelity. To bridge this gap,we propose OREO, an alignment framework that enhances the realismof 3D generators by leveraging rich 2D diffusion priors. Instead of relyingon static datasets, OREO establishes a dynamic optimization loop thatproduces on-the-fly pseudo ground truths. At its core, OREO combinestwo components: (i) Reinforced Editing, which uses an image editor torefine rendered views of the 3D output, improving the visual fidelity ofrendered views while preserving geometric structure; and (ii) ContrastiveDistillation, which treats the edited views as positive and the originalrenderings as negative anchors, distilling the fidelity gap into the 3Dgenerator via a latent contrastive objective. Experiments demonstratethat OREO effectively improves upon pre-trained baselines, producing3D assets with enhanced visual realism. Our project page is at https://theericma.github.io/oreo/.

Encoding videos into discrete tokens can enable concise andunified multi-modal LLMs. Previous discrete video VAEs suffer from un-stable training, long training time, and degraded reconstruction quality.We revisit the relationship between continuous and discrete VAEs andfind that bridging discrete and continuous representations improves dis-crete token learning. Based on this insight, we propose a unified progres-sive training framework that (i) jointly optimizes continuous and discretereconstructions within a single network, and (ii) progressively derives afamily of VAEs at different compression ratios, leading to faster conver-gence and better final performance. Furthermore, leveraging this unifiedtraining, a single VAE can achieve competitive performance for bothcontinuous and discrete representations. Meanwhile, we propose two ar-chitectural improvements to further boost the performance of discreteVAEs. First, inspired by the use of enlarging continuous latent dimensionto boost reconstruction, we propose multi-token quantization, improvingPSNR by nearly 1 dB at the same token compression ratio. Second, weintroduce first-frame enhancement, which uses a lower-compression firstframe as an anchor in causal VAEs, alleviates the limited-context is-sue of the first frame and significantly improves reconstruction underhigh-compression settings (e.g., 4 × 16 × 16). The code is available athttps://github.com/HVision-NKU/OneVAE.

Understanding how neural networks rely on visual cues offers a human-interpretable view of their internal decision processes. The cueconflict benchmark has been influential in probing shape-texture preference and in motivating the insight that stronger, human-like shape bias is often associated with improved in-domain performance. However, we find that the current stylization-based instantiation can yield unstable and ambiguous bias estimates. Specifically, stylization may not reliably instantiate perceptually valid and separable cues nor control their relative informativeness, ratio-based bias can obscure absolute cue sensitivity, and restricting evaluation to preselected classes can distort model predictions by ignoring the full decision space. Together, these factors can confound preference with cue validity, cue balance, and recognizability artifacts. We introduce REFINED-BIAS, an integrated dataset and evaluation framework for reliable and interpretable shape–texture bias diagnosis. REFINED-BIAS constructs balanced, humanand modelrecognizable cue pairs using explicit definitions of shape and texture, and measures cuespecific sensitivity over the full label space via a ranking-based metric, enabling fairer cross-model comparisons. Across diverse training regimes and architectures, REFINED-BIAS enables fairer cross-model comparison, more faithful diagnosis of shape and texture biases, and clearer empirical conclusions, resolving inconsistencies that prior cue-conflict evaluations could not reliably disambiguate. Our code is available at REFINED-BIAS.

Video Face Swapping (VFS) requires seamlessly injecting asource identity into a target video while meticulously preserving the orig-inal pose, expression, lighting, background, and dynamic information.Existing methods struggle to maintain identity similarity and attributepreservation while preserving temporal consistency. To address the chal-lenge, we propose a comprehensive framework to seamlessly transfer thesuperiority of Image Face Swapping (IFS) to the video domain. Wefirst introduce a novel data pipeline SyncID-Pipe that pre-trains anIdentity-Anchored Video Synthesizer and combines it with IFS modelsto construct bidirectional ID quadruplets for explicit supervision. Build-ing upon paired data, we propose a powerful Diffusion Transformer-basedframework OmniFace, employing a core Modality-Aware Conditioningmodule to discriminatively inject multi-model conditions. Meanwhile,we propose a Synthetic-to-Real Training mechanism and an Identity-Coherence Reward Weighting strategy to enhance visual realism andidentity consistency under challenging scenarios. To address the issue oflimited benchmarks, we introduce IDBench-V, a comprehensive bench-mark encompassing diverse scenes. Extensive experiments demonstrateOmniFace outperforms state-of-the-art methods and further exhibits ex-ceptional versatility, which can be seamlessly adapted to various swap-related tasks.

Multimodal large language models (MLLMs) struggle withalignment due to the limitations of existing reward models (RMs), whichare predominantly vision-centric, dependent on costly human labels, andprovide opaque scalar scores that fail to capture nuanced reasoning,leading to brittle alignment. We present Omni-RRM, an Omni-modalRubric-grounded Reward Model that generates multi-dimensional re-ward signals across text, image, video, and audio. To overcome the highcost and inherent inconsistency of human-centric evaluation in multi-dimensional reasoning, we introduce Omni-Preference, a high-qualitydataset constructed via automatic rubric-grounded preference synthe-sis. In this pipeline, teacher models reconcile raw preferences into ex-plicit justifications, ensuring that the synthesized supervision is bothhigh-fidelity and interpretable. Omni-RRM is trained using a progres-sive SFT + GRPO regimen, specifically optimized to sharpen rewarddiscrimination on low-margin, hard preference pairs. It achieves state-of-the-art accuracy on video (80.2% on ShareGPT-Video) and audiobenchmarks (66.8% on Audio-HH-RLHF and 65.0% on TA2T), yield-ing a five-benchmark Overall accuracy of 70.4% and a +17.0% relativegain over its backbone. Furthermore, Omni-RRM effectively guides Best-of-N selection and exhibits robust transfer to text-only alignment. Allresources, including the dataset, training and inference code, and modelcheckpoints are available at https://tmfk418.github.io/Omni-RRM.

In this paper, we propose RoomPlanner, the first fully au-tomatic 3D room generation framework for painlessly creating realisticindoor scenes with only short text as input. Without any manual layoutdesign or panoramic image guidance, our framework can generate explicitlayout criteria for rational spatial placement. We begin by introducing ahierarchical structure of language-driven agent planners that can auto-matically parse short and ambiguous prompts into detailed scene descrip-tions. These descriptions include raw spatial and semantic attributes foreach object and the background, which are then used to initialize 3Dpoint clouds. To position objects within bounded environments, we im-plement two arrangement constraints that iteratively optimize spatialarrangements, ensuring a collision-free and accessible layout solution.In the final rendering stage, we propose a novel ReachView Samplingstrategy for camera trajectory, along with the Interval Timestep FlowSampling (ITFS) strategy, to efficiently optimize the coarse 3D Gaussianscene representation. These approaches help reduce the total generationtime to under 30 minutes. Extensive experiments demonstrate that ourmethod can produce geometrically rational 3D indoor scenes, surpass-ing prior approaches in both rendering speed and visual quality whilepreserving editability. The code will be available at https://kaitlina-s.github.io/RoomPlanner/.

Neighboring-slice self-supervised denoising is attractive forvolumetric medical imaging, yet inter-slice misalignment breaks anatom-ical correspondence and often yields ghosting and blurred margins whenadjacent slices are used naïvely as targets. We propose Neighbor-GuidedPatch Sampling (NGPS), a lightweight framework that constructs neigh-boring supervision under local inter-slice misalignment. To avoid learn-ing from misleading targets, prior methods commonly mask discrepantregions, but this stabilizes training at the cost of leaving a non-trivialportion of neighboring evidence unexploited, particularly around high-frequency anatomical boundaries. NGPS addresses this by decouplingstructure matching from signal retrieval: for each masked location, itsearches a local neighborhood for structurally similar candidate patchesusing a simple guide image (e.g., fast bilateral filtering), while retriev-ing the supervision signal directly from the raw noisy neighbor at thematched coordinates. By matching on a noise-attenuated guide whileretrieving raw values from neighboring slices, NGPS constructs localpseudo targets without dense deformation-field estimation or spatial re-sampling. Across the evaluated CT and synthetic-Rician MRI settings,NGPS improves fidelity and structure-sensitive metrics. Code is availableat https://github.com/cv-cho/NGPS.

We introduce ArcDeck, a multi-agent framework that for-mulates paper-to-slide generation as a structured narrative reconstruc-tion task. Unlike existing methods that directly summarize raw text intoslides, ArcDeck explicitly models the source paper’s logical flow. It firstparses the input to construct a discourse tree and establish a global com-mitment document, ensuring the high-level intent is preserved. Thesestructural priors then guide an iterative multi-agent refinement process,where specialized agents iteratively critique and revise the presentationoutline before rendering the final visual layouts and designs. To evaluateour approach, we also introduce ArcBench, a newly curated benchmarkof academic paper-slide pairs. Experimental results demonstrate that ex-plicit discourse modeling, combined with role-specific agent coordination,significantly improves the narrative flow and logical coherence of the gen-erated presentations. The ArcDeck codebase and ArcBench benchmarkare available on our project webpage.

While Change Captioning systems have garnered substantialattention to respond to our evolving world, their true performance on di-verse real-world change contexts remains largely unexplored due to thelack of comprehensive evaluation frameworks. To fill this gap, we proposeC3-Bench, a comprehensive benchmark for evaluating Context-awareChange Captioning. C3-Bench features: (1) 4,996 human-labeled imagepairs of 51 real-world change contexts across four domains (e.g., naturalscenes, remote sensing imagery, image editing, and anomalies), each withdiverse, carefully curated scenarios derived from multiple change-centriccommunities; and (2) the first LLM-as-Judge evaluation framework in thechange captioning task that measure fine-grained dimensions (e.g., cor-rectness, specificity, fluency, and relevance), along with a novel reversibil-ity metric exploring whether models understand changes with symmetricconsistency. Based on C3-Bench, we benchmark 32 models—including* Corresponding authorconventional change captioning models, proprietary Large MultimodalModels (LMMs), and 2B-90B open-source LMMs. We reveal a fundamen-tal blind spot in the prevailing change captioning paradigm: Once thechange context departs from training-style regimes, conventional mod-els collapse, and even state-of-the-art LMMs such as GPT-5.2 exhibitsystematic domain- and position-dependent errors that distort reliablechange understanding. By making these hidden failure modes explicitand measurable, we delineate the next frontier for building generalizableand trustworthy change captioning systems. All codes and datasets arepublicly available on the project page.

3D Gaussian Splat (3DGS) enables high-fidelity, real-timenovel view synthesis by representing scenes with large sets of anisotropicprimitives, but often requires millions of Splats, incurring significant stor-age and transmission costs. Most existing compression methods rely onGPU-intensive post-training optimization with calibrated images, limit-ing practical deployment. We introduce NanoGS, a training-free andlightweight framework for Gaussian Splat simplification. Instead of re-lying on image-based rendering supervision, NanoGS formulates sim-plification as local pairwise merging over a sparse spatial graph. Themethod approximates a pair of Gaussians with a single primitive usingmass preserved moment matching and evaluates merge quality througha principled merge cost between the original mixture and its approx-imation. By restricting merge candidates to local neighborhoods andselecting compatible pairs efficiently, NanoGS produces compact Gaus-sian representations while preserving scene structure and appearance.NanoGS operates directly on existing Gaussian Splat models, runs ef-ficiently on CPU, and preserves the standard 3DGS parameterization,enabling seamless integration with existing rendering pipelines. Experi-ments demonstrate that NanoGS substantially reduces primitive countwhile maintaining high rendering fidelity, providing an efficient and prac-tical solution for Gaussian Splat simplification. Our project website isavailable at https://saliteta.github.io/NanoGS/ .

Monocular 3D scene reconstruction has recently seen signif-icant progress. Powered by the modern neural architectures and large-scale data, recent methods achieve high performance in depth estimationfrom a single image. Meanwhile, reconstructing and decomposing com-mon scenes into individual 3D objects remains a hard challenge due tothe large variety of objects, frequent occlusions and complex object rela-tions. Notably, beyond shape and pose estimation of individual objects,applications in robotics and animation require physically-plausible scenereconstruction where objects obey physical principles of non-penetrationand realistic contacts. In this work we advance object-level scene re-construction along two directions. First, we introduce MessyKitchens,a new dataset with real-world scenes featuring cluttered environmentsand providing high-fidelity object-level ground truth in terms of 3D ob-ject shapes, poses and accurate object contacts. Second, we build onthe recent SAM 3D approach for single-object reconstruction and ex-tend it with Multi-Object Decoder (MOD) for joint object-level scenereconstruction. To validate our contributions, we demonstrate MessyK-itchens to significantly improve previous datasets in registration accuracyand inter-object penetration. We also compare our multi-object recon-struction approach on three datasets and demonstrate consistent andsignificant improvements of MOD over the state of the art. Our newbenchmark, code and pre-trained models are publicly available on ourproject website: https://messykitchens.github.io/.

Existing 3D relighting methods operate through either ex-plicit material decomposition, diffusion-based view-space generation, ora combination of both, requiring full recomputation for each new light-ing condition. We observe that recent latent scene representations, whichencode multi-view images into a set of compact tokens with no fixed phys-ical semantics, open up a novel design space for relighting. We presentLumiTokens, a framework that formulates 3D relighting as a directtransformation on latent scene tokens, without explicit 3D representa-tions, rendering equations, or physics-based decomposition. Our modelintroduces a Scene Token Editor that processes scene tokens jointly withlight-ray tokens through self-attention, producing updated tokens thatcan be decoded into multi-view-consistent relit images. To support di-verse lighting types through a unified interface, all lighting signals, in-cluding environment maps, point lights, and area lights, are parameter-ized as Plücker ray tokens, enabling native 3D user interaction with arepresentation that carries no explicit spatial structure. Crucially, thisdesign supports progressive relighting: because the editor’s output re-mains in the same latent space as its input, a user can incrementallybuild up illumination one light source at a time, with each edit compos-ing in token space. Experiments demonstrate that LumiTokens achievescomparable or superior relighting quality to other methods and supportsprogressive, composable lighting edits.

Evaluating the physical consistency of embodied world mod-els (EWMs) is a critical open challenge. While closed-loop evaluationvia simulator rollouts offers a more faithful assessment of physical plau-sibility than open-loop alternatives, existing frameworks almost exclu-sively rely on Inverse Dynamics Models (IDMs) for action extraction.Due to the intricate mapping from 2D pixel space to 3D kinematicspace, the learned IDMs can be brittle to data outside their trainingdistribution, resulting in unreliable action extraction from the gener-ated videos with novel objects and scenarios. This creates an unavoid-able attribution ambiguity between world model inaccuracies and ex-tractor errors. To reduce this ambiguity, we present KineBench, anIDM-free closed-loop benchmark for EWMs, built upon an explicit kine-matic grounding pipeline. Given a generated video, KineBench employscascaded visual foundation models to directly extract 6D end-effectorposes from individual frames, which are then executed in a physics sim-ulator for closed-loop validation. This explicit grounding directly teststhe physical feasibility rather than visual plausibility, while remainingsensitive to general physical hallucinations such as gripper vanishing orspatial inconsistency. Beyond execution-based task success, KineBenchincorporates two classical 3D kinematic metrics—Spectral Arc Length(SPARC) and the Maruyama Manipulability Index—to characterize tra-jectory smoothness and kinematic feasibility from a robot-centric per-spective. Across the evaluated models and tasks, these metrics exhibittask- and model-dependent associations with physical success rates, sug-gesting that they provide complementary diagnostic signals for assessingembodied generation quality. Built on 20 diverse manipulation tasks inManiSkill3, KineBench evaluates EWMs across four progressive suites:basic execution, task transfer, visual out-of-distribution generalization,and complexity-conditioned scaling. Evaluation across frontier models re-veals task-complexity-bounded nonlinear scaling in embodied video gen-eration, providing empirical guidance for future data-scaling strategies.The code and datasets are available on GitHub at https://github.com/minecraft-zzz/KineBench and on Hugging Face at https://huggingface.co/datasets/Zorkzak/KineBenchDatasets.

Normalizing Flows (NFs) is a principle generative framework,which learns to generate images by establishing a bijective mapping tra-jectory between noise and data with likelihood maximization. In thispaper, we leverage the self-play paradigm, a general trajectory fine-tuningmethod to improve generative capability of NFs without auxiliary fine-tuning data. The general self-play mechanism formulates competitionbetween model and its historical counterpart in favoring real data trajec-tory (i.e., “wining trajectory”) while forgetting trajectory of self-generatedsynthesized data (i.e., “losing trajectory”), improving model capabilitythrough preference alignment. However, NF-based self-play faces twochallenges. (1) Enforcing NF to forget data synthesized with historicalmodel contradicts with NF’s original training goal, leading to potentialtraining collapse. (2) Self-play is naturally plausible for generative modelwith pre-defined data trajectory (e.g., diffusion models), but faces com-patibility problem with NFs due to its implicit wining data trajectorydefinition. We thus propose jigsaw self-play optimization for NFs (JSON)to overcome these challenges. Our core contribution is the integration ofa jigsaw puzzle reassembly task to serve as an intrinsic spatial coherenceevaluator, identifying “winning” trajectories that exhibit superior spatiallogic for self-play fine-tuning. We further propose a bounded self-play lossand a nested-loop optimization strategy to mitigate training collapse andensure stable fine-tuning. By aligning the model with high-quality struc-tural anchors mined during self-play, JSON achieves promising results onImageNet-1K, providing a robust framework for high-fidelity NF-basedgeneration.

We introduce VSeek, an agentic framework that transformslong-video question answering (LVQA) from a passive, single-pass per-ception task into a multi-turn retrieval process. VSeek utilizes a naturallanguage-driven search to identify relevant context within long videosand is post-trained with reinforcement learning (RL) to jointly formulatetargeted search queries and reason over retrieved clips for LVQA. WhileRL post-training has revolutionized reasoning in symbolic domains suchas mathematics and code, its application to long-video understandingremains hindered by a lack of verified rewards. To ensure that the re-trieved context is relevant, we propose a novel neuro-symbolic approachthat bridges open-ended natural language with discrete visual verifica-tion. Specifically, complex user queries are compiled into formal tempo-ral logic specifications for systematically decomposing natural languagequestions into a definitive checklist of required atomic visual primitives,such as key objects and activities, along with their temporal ordering.These systematically derived grounding events provide the critical feed-back signal for RL post-training, enabling dense, verifiable rewards basedon the successful retrieval of these specific visual elements rather thanrelying entirely on outcome-only answer accuracy. By explicitly optimiz-ing for this verifiable evidence-seeking behavior, VSeek improves Pass@1scores by up to 8% and Pass@4 scores by 15% on long-video understand-ing benchmarks compared to base models. We open-source our code athttps://utaustin-swarmlab.github.io/VSeek.

The advancement of generative AI models capable of pro-ducing text and image marks a critical step forward in the realm of* Equal Contribution, B Corresponding Author, † Project Leadermultimodal intelligence, particularly for tasks involving the interleavingof both modalities. To advance this intelligence to the next stage, it iscrucial for models to autonomously generate free-form interleaved text-image sequences. In this paper, we introduce ILLUME-X, an advancedunified multimodal paradigm that enables high-quality, free-form inter-leaved text-image generation by improving multimodal data efficiencyand stabilizing the multimodal training process. ILLUME-X comprisesthree key components: (i) an expanded training data pipeline optimizedfor interleaved text-image generation, (ii) a progressive training strategywith self-adaptive objectives for free-length multimodal token sequences,and (iii) an objective and comprehensive evaluation method ILScore forinterleaved text-image sequences. Notably, our ILLUME-X outperformsprevious unified models across multiple interleaved text-image generationtasks like style transfer, image decomposition and storytelling.

While the field of co-speech gesture generation has seen sig-nificant advances, producing holistic, semantically grounded gestures re-mains a challenge. Existing approaches rely on external semantic re-trieval methods, which limit their generalisation capability due to de-pendency on predefined linguistic rules. Flow-matching-based methodsproduce promising results; however, the network is optimised usingonly semantically congruent samples without exposure to negative ex-amples, leading to learning rhythmic gestures rather than sparse mo-tion, such as iconic and metaphoric gestures. Furthermore, by modellingbody parts in isolation, the majority of methods fail to maintain cross-modal consistency. We introduce a Contrastive Flow Matching-basedco-speech gesture generation model that uses mismatched audio–textconditions as negatives, training the velocity field to follow the cor-rect motion trajectory while repelling semantically incongruent trajec-tories. Our model ensures cross-modal coherence by embedding text,audio, and holistic motion into a composite latent space via cosine andcontrastive objectives. Extensive experiments and a user study demon-strate that our proposed approach outperforms state-of-the-art methodson two datasets, BEAT2 and SHOW. Our project page is available at:https://marcos452.github.io/HoliticSemGes/

The text encoder is a critical component of text-to-image andtext-to-video diffusion models, fundamentally determining the semanticfidelity of the generated content. However, its development has been hin-dered by two major challenges: the lack of an efficient evaluation frame-work that reliably predicts downstream generation performance, and thedifficulty of effectively adapting pretrained language models for visualsynthesis. To address these issues, we introduce GRAN-TED, a paradigmto Generate Robust, Aligned, and Nuanced Text Embeddings for Diffu-sion models. Our contribution is twofold. First, we propose TED-6K, anovel text-only benchmark that enables efficient and robust assessmentof an encoder’s representational quality without requiring costly end-to-end model training. We demonstrate that performance on TED-6K,standardized via a lightweight, unified adapter, strongly correlates withan encoder’s effectiveness in downstream generation tasks. Notably, un-der our experimental setup, compared with training a diffusion modelfrom scratch, evaluating with TED-6K is about 750× faster. Second,guided by this validated framework, we develop a superior text encoderusing a novel two-stage training paradigm. This process involves an ini-tial fine-tuning stage on a Multimodal Large Language Model for bettervisual representation, followed by a layer-wise weighting method to ex-tract more nuanced and potent text features. Our experiments show thatthe resulting GRAN-TED encoder not only achieves state-of-the-art per-formance on TED-6K but also leads to demonstrable performance gainsin text-to-image and text-to-video generation.

Unified multimodal models target joint understanding, rea-soning, and generation, but current image editing benchmarks are largelyconfined to natural images and shallow commonsense reasoning, offeringlimited assessment of this capability under structured, domain-specificconstraints. In this work, we introduce GRADE, the first benchmarkto assess discipline-informed knowledge and reasoning in image editing.GRADE comprises 520 carefully curated samples across 10 academicdomains, spanning from natural science to social science. To supportrigorous evaluation, we propose a multi-dimensional evaluation proto-col that jointly assesses Discipline Reasoning, Visual Consistency, andLogical Readability. Extensive experiments on 20 state-of-the-art open-source and closed-source models reveal substantial limitations in currentmodels under implicit, knowledge-intensive editing settings, leading tolarge performance gaps. Beyond quantitative scores, we conduct rigor-ous analyses and ablations to expose model shortcomings and identifythe constraints within disciplinary editing. Together, GRADE pinpointskey directions for the future development of unified multimodal models,advancing the research on discipline-informed image editing and reason-ing. Our benchmark and evaluation code are publicly released.