ECCV 2026 论文列表

Reinforcement learning (RL) has become a standard tech-nique for post-training diffusion-based image synthesis models, as itenables learning from reward signals to explicitly improve desirable as-pects such as image quality and prompt alignment. In this paper, wepropose an online RL variant that reduces the variance in the modelupdates by sampling paired trajectories and pulling the flow velocityin the direction of the more favorable image. Unlike existing methodsthat treat each sampling step as a separate policy action, we considerthe entire sampling process as a single action. We experiment with bothhigh-quality vision language models and off-the-shelf quality metrics forrewards, and evaluate the outputs using a broad set of metrics. Ourmethod converges faster and yields higher output quality and promptalignment than previous approaches.

A recent cutting-edge topic in multimodal modeling is tounify visual comprehension and generation within a single model. How-ever, the two tasks demand mismatched decoding regimes and visualrepresentations, making it non-trivial to jointly optimize within a sharedfeature space. In this work, we present Cheers, a unified multimodalmodel that decouples patch-level details from semantic representations,thereby stabilizing semantics for multimodal understanding and improv-ing fidelity for image generation via gated detail residuals. Cheers in-cludes three key components: (i) a unified vision tokenizer that encodesand compresses image latent states into semantic tokens for efficient LLMconditioning, (ii) an LLM-based Transformer that unifies autoregressivedecoding for text generation and diffusion decoding for image generation,and (iii) a cascaded flow matching head that decodes visual semanticsfirst and then injects semantically gated detail residuals from the vi-sion tokenizer to refine high-frequency content. Experiments on popularbenchmarks demonstrate that Cheers matches or surpasses advancedUMMs in both visual understanding and generation. Notably, Cheersoutperforms the Tar-1.5B on the popular benchmarks GenEval and MM-Bench, while requiring only 20% of the training cost, indicating effectiveand efficient (i.e., 4× token compression) unified multimodal modeling.Code and checkpoints are available at github.com/AI9Stars/CHEERS.

This paper introduces a scalable paradigm for supervisedstyle transfer by inverting the problem: instead of learning to stylizedirectly, we learn to destylize, reducing stylistic elements from artisticimages to recover their natural counterparts and thereby producing au-thentic, pixel-aligned training pairs at scale. To realize this paradigm, wepropose DeStylePipe, a progressive, multi-stage destylization frameworkthat begins with global general destylization, advances to category-wiseinstruction adaptation, and ultimately deploys specialized model adap-tation for complex styles that prompt engineering alone cannot han-dle. Tightly integrated into this pipeline, DestyleCoT-Filter employsChain-of-Thought reasoning to assess content preservation and style re-moval at each stage, routing challenging samples forward while discard-ing persistently low-quality pairs. Built on this framework, we constructDeStyle-350K, a large-scale dataset aligning diverse artistic styles withtheir underlying content. We further introduce BCS-Bench, a benchmarkfeaturing balanced content generality and style diversity for systematicevaluation. Extensive experiments demonstrate that models trained onDeStyle-350K achieve superior stylization quality, validating destyliza-tion as a reliable and scalable supervision paradigm for style transfer. Ourproject page: https://wangyephd.github.io/projects/DeStyle/index.html

We present a novel paradigm for ultra-low-bitrate imagecompression (ULB-IC) that exploits the “temporal” evolution in gener-ative image compression. Specifically, we define an explicit intermediatestate during decoding: a compact anchor frame, which preserves the scenegeometry and semantic layout while discarding high-frequency details.We then reinterpret generative decoding as a virtual temporal transitionfrom this anchor to the final reconstructed image. To model this progres-sion, we leverage a pretrained video diffusion model (VDM) as a temporalprior: the anchor frame serves as the initial frame and the original imageas the target frame, transforming the decoding process into a next-frameprediction task. In contrast to image diffusion-based ULB-IC models, ourdecoding proceeds from a visible, semantically faithful anchor, which im-proves both fidelity and realism for perceptual image compression. Ex-tensive experiments demonstrate that our method achieves superior rate-distortion performance. On the CLIC2020 test set, our method achievesover 50% bitrate savings across LPIPS, DISTS, FID, and KID com-pared to DiffC, while also delivering a significant decoding speedup of upto ×5. Code will be released at https://github.com/UnoC-727/NeFIC.

Multimodal CAD generation faces a fundamental scalability challenge. Design-history methods are confined to small annotated datasets, while BRep topology is discrete and non-differentiable. Meanwhile, millions of unannotated 3D meshes remain untapped, since existing CAD methods cannot leverage them without explicit CAD annotations. We propose DreamCAD, a multimodal generative framework that bridges this gap by representing shapes as C0-continuous Bézier patches with differentiable tessellation, enabling direct point-level supervision on large-scale 3D meshes without CAD-specific annotations. We further introduce CADCap-1M, the largest CAD captioning dataset with 1M+ GPT-5-generated descriptions to advance text-to-CAD research. DreamCAD achieves state-of-the-art performance on ABC and Objaverse across text, image, and point modalities, surpassing 75% user preference. Finally, we show that DreamCAD’s accurate, compact geometry enables topology recovery into production-ready CAD models, exportable as editable STEP files. Project page is available at https://s adilkhan.github.io/dreamcad2026/.

The development of unified multimodal large language mod-els (MLLMs) is fundamentally challenged by the granularity gap be-tween visual understanding and generation: understanding requires high-level semantic abstractions, while image generation demands fine-grainedpixel-level representations. Existing approaches usually enforce the twosupervision on the same set of representation or decouple these two su-pervision on separate feature spaces, leading to interference and incon-sistency, respectively. In this work, we propose EvoTok, a unified imagetokenizer that reconciles these requirements through a residual evolutionprocess within a shared latent space. Instead of maintaining separatetoken spaces for pixels and semantics, EvoTok encodes an image intoa cascaded sequence of residual tokens via residual vector quantization.This residual sequence forms an evolution trajectory where earlier stagescapture low-level details and deeper stages progressively transition to-ward high-level semantic representations. Despite being trained on a rel-atively modest dataset of 13M images, far smaller than the billion-scaledatasets used by many previous unified tokenizers, EvoTok achieves astrong reconstruction quality of 0.43 / 0.25 rFID at 256×256 / 384×384resolutions on ImageNet-1K, respectively. When integrated with a largelanguage model, EvoTok shows promising performance across 8 out of 9visual understanding benchmarks, and remarkable results on image gen-eration benchmarks such as GenEval, GenAI-Bench, and DPG. Theseresults demonstrate that modeling visual representations as an evolvingtrajectory provides an effective and principled solution for unifying visualunderstanding and generation.

Realistic handwritten text generation plays an important rolein numerous applications, such as font design, biometric authentication,and robotic calligraphy. Existing methods are typically divided into twoindependent paradigms: online approaches that estimate handwritingtrajectories and offline approaches that synthesize realistic handwritingimages. While online models capture structural and temporal dynamics,they often lack fine-grained textures, whereas offline models reproducerealistic appearance but discard stroke order. However, unifying onlineand offline models remains challenging due to (1) the lack of an explicitphysical model linking stroke kinematics to pixel-level appearance and(2) the absence of paired trajectory–image datasets. Moreover, enablingend-to-end learning requires a differentiable rendering process acrossmotion and appearance domains. To address these challenges, we proposea compact physical brush model that bridges stroke dynamics and visualappearance, together with a differentiable rendering module that convertsstroke trajectories into stylized images. By integrating these components,we propose a unified online–offline handwriting generation frameworkvia differentiable brush rendering. The proposed framework consistsof four core modules: 1) a text-to-stroke generator that predicts thetarget stroke conditioned on the given text and style image, 2) a brushparameter observer that extracts brush model parameters from stylereferences, 3) a differentiable brush renderer that maps a stroke sequenceand physical brush parameters into a handwritten image, and 4) a zero-shot image refiner that refines rendered images via diffusion models.Extensive experiments and real-world robotic calligraphy demonstrationsvalidate our approach, achieving both structural and visual fidelity.

Recovering an editable design file from a raster image is a common and costly bottleneck in modern design workflows, yet remains challenging since editability depends on recovering multi-modal attributes, such as typography, vector geometry, colors, grouping, and layer ordering. We present ReDesign, an agentic framework that grows an editable layer hierarchy by selecting and composing specialized tools across modalities. To keep this long decision process reliable despite imperfect tool outputs, we introduce graceful verification at each expansion, which provides local accept, prune, or retry feedback that prevents error accumulation and avoids large scale reruns. To evaluate editability at scale, we introduce the Figma Edit Replay Benchmark, consisting of 909 raw Figma files and 14,796 controlled edit instructions that replay edits on reconstructed outputs. Across this benchmark and standard reconstruction metrics, ReDesign achieves strong visual fidelity while delivering the highest editability across layout, color, and text edits, outperforming layered decomposition baselines and serial tool use pipelines.

Complex-valued Gaussian primitives have recently been ex-plored for representing holographic radiance _x001C_elds in 3D novel view syn-thesis. In this work, we extend this line of research to the hologramoptimization domain and propose a structured representation based oncomplex-valued 2D Gaussian primitives. Inspired by Gabor's theory,we show that our primitive attains the minimum space_x0015_frequency un-certainty and reduces the parameter search space by 5:1 compared toper-pixel parameterization. To enable end-to-end training, we develop adi_x001B_erentiable rasterizer for our representation, integrated with a GPU-optimized light propagation kernel in free space. Extensive experimentsshow that our method reduces VRAM usage by up to 30% and acceler-ates optimization by 50% over standard autodi_x001B_-based implementations,delivers up to 13 dB higher PSNR than prior Gaussian-based methods,and achieves up to 3200× faster rendering while maintaining reconstruc-tion quality on par with existing CGH approaches. For evaluation, weintroduce a conversion procedure that adapts our representation to prac-tical hologram formats, including smooth and random phase-only holo-grams. By reducing the hologram parameter search space, our representa-tion enables a more scalable hologram estimation in the next-generationcomputer-generated holography systems.

Precise 3D spatial orchestration in text-to-video generationremains a signi_x001C_cant challenge, particularly for multi-object scenes wheresemantic layout and temporal dynamics are often entangled. While exist-ing depth-conditioned models achieve good structural _x001C_delity, they neces-sitate dense, frame-accurate guidance that is labor-intensive to authorfor dynamic events involving deformable objects. We present LooseC-ontrolVideo (LCV), a framework that enables intuitive and expressivecontrol by using sparse, oriented 3D boxes as a _x0010_blocking_x0011_ proxy. Thisallows users to author high-level layout and trajectory while leveraginga video generative model to generate realistic occlusions, dynamics andinteractions. We achieve this by _x001C_ne-tuning a Wan 2.2 backbone on avideo dataset annotated with DNOCS, a novel encoding for 3D size, ori-entation and depth-ordered occlusions. Furthermore, our method allowsfor localized re_x001C_nement_x0016_such as adjusting a jump trajectory or addingan interaction_x0016_with minimal disruption to the global scene context. Ex-tensive evaluations on the nuScenes, HO-3D, and BEHAVE benchmarksdemonstrate that LCV signi_x001C_cantly outperforms existing 2D-box and_x001D_ow-based baselines. Our _x001C_ndings indicate a 1.2-3x improvement in Tra-jectory Error; 2x improvement in Rigid Motion Consistency; and a 1.5-2x increase in Occlusion Accuracy over current state-of-the-art layout-conditioned models, demonstrating that oriented 3D primitives providegood geometric prior for complex, multi-agent video authoring.

Exo-to-Ego video generation aims to synthesize a first-personvideo from a synchronized third-person view and corresponding cam-era poses. While paired supervision is available, synchronized exo-egodata inherently introduces substantial spatio-temporal and geometricdiscontinuities, violating the smooth-motion assumptions of standardvideo generation benchmarks. We identify this synchronization-inducedjump as the central challenge and propose Syn2Seq-Forcing, a sequen-tial formulation that interpolates between the source and target videosto form a single continuous signal. By reframing Exo2Ego as sequentialsignal modeling rather than a conventional condition–output task, ourapproach enables diffusion-based sequence models, e.g. Diffusion Forc-ing Transformers (DFoT), to capture coherent transitions across framesmore effectively. Empirically, we show that interpolating only the videos,without performing pose interpolation already produces significant im-provements, emphasizing that the dominant difficulty arises from spatio-temporal discontinuities. Beyond immediate performance gains, this for-mulation establishes a general and flexible framework capable of uni-fying both Exo2Ego and Ego2Exo generation within a single contin-uous sequence model, providing a principled foundation for future re-search in cross-view video synthesis. The code will be released at https://github.com/insait-institute/Syn2Seq.

Subject-driven image generation faces an “Identity-DiversityParadox”, where strong identity preservation often leads to rigid and low-diversity outputs. We propose a post-training framework called DivRLthat jointly optimizes identity consistency and structural diversity si-multaneously by leveraging disentangled visual features from a robustsimilarity model. Specifically, we introduce a Negative Self-SimilarityMeasure (nSSM) to quantify structural diversity, and Visual Seman-tic Matching (VSM) to evaluate identity consistency. We propose an“Explore-and-Suppress” strategy that treats VSM as a gated constraint:the model freely explores structurally diverse configurations, and onlysamples that violate the identity threshold are penalized via a quadratichinge loss. This converts identity preservation from a competing objectiveinto a feasibility constraint, allowing nSSM and VSM to improve jointly.Experiments demonstrate that our method effectively pushes the modelto generate both consistent and diverse images and improves structuraldiversity while maintaining comparable identity consistency through agated optimization formulation.

Generative video compositing, which involves inserting ex-ternal assets seamlessly into existing video sequences, is essential forcontent creation and visual effects. However, existing approaches suf-fer from a control-fidelity trade-off: they either hallucinate motion fromstatic images, failing to preserve the dynamics of pre-animated assets, orlack fine-grained spatial control for precise asset placement along user-defined trajectories. We propose FlexComposer, a unified framework thatstandardizes video compositing as a trajectory-guided conditional gen-eration task, enabling the seamless integration of both static images anddynamic footage. Our approach introduces three key designs: (1) a Uni-fied Canonical Foreground Representation that decouples an object’s in-trinsic motion from its global displacement, standardizing heterogeneousinputs into a stabilized, centered latent space; (2) a Spatial-Aware La-tent Injection strategy that exploits the translation equivariance of VAElatent spaces to transport canonical features onto target trajectories viaa parameter-free mechanism; and (3) a Hybrid Dataset and Synthetic-to-Real Curriculum that synergizes procedural simulation, real-world cine-matic footage, and generative data to implicitly learn physically plausibleillumination and shadow harmonization. This unified design handles di-verse inputs—from product photos to dynamic subjects—achieving high-fidelity motion control and environmental integration without the needfor explicit 3D reconstruction or auxiliary learnable adapters. Extensiveexperiments demonstrate that FlexComposer outperforms state-of-the-art methods in visual quality, temporal consistency, and trajectory ad-herence.

We present GCA (Gaussian Constitutive Alignment), a frame-work for learning implicit constitutive laws from monocular dynamicvideo of deformable objects represented by 3D Gaussians. Given a staticmulti-view scan for geometric initialization, our method learns intrin-sic physical dynamics solely from a single fixed-viewpoint video of themoving object. Existing implicit methods often suffer from local minimaunder noisy supervision and lack physical interpretability, while explicitapproaches rely on predefined constitutive equations, limiting generaliz-ability and becoming unstable in monocular settings. To address thesechallenges, our framework unifies LoRA-based adaptation with two keyalignment modules. First, we propose Rank-based Depth-Geometric An-chors (RDGA) to establish robust geometric constraints from monoc-ular dynamic observations via scale-invariant rank-based depth align-ment, reducing the reliance on unreliable pixel-level color supervision.Second, a Constitutive Prior Regularizer (CPR) integrates classical con-stitutive models as soft differentiable priors, regularizing the optimiza-tion while preserving the flexibility of implicit modeling—even when theactual material is absent from the hypotheses. Extensive experiments onsynthetic, real-to-sim, and real-world datasets demonstrate that GCAoutperforms existing methods, achieving 48% lower Chamfer Distancethan the strongest baseline on synthetic benchmarks while remainingrobust under monocular supervision.

Creating photorealistic, animatable 3D human avatars from monocular images still largely depends on Linear Blend Skinning (LBS) and parametric body models which constrain expressivity and often introduce artifacts due to imperfect fitting. We propose LUNA, an LBSfree universal neural animation model that directly maps multiple 2D controls like images, keypoints, sketch and unseen characters into 3D Gaussian deformations, bypassing explicit body fitting. At its core, a transformer-based motion regressor disentangles global rigid motion from fine-grained local dynamics to capture both coherent movement and subtle non-rigid effects. To resolve the inherent ambiguity of 2D-to-3D lifting while scaling beyond fitted datasets, we introduce hybrid supervision that distills soft structural priors from an LBS teacher and a loss that supports training on both limited fitted data and large in-the-wild unlabeled videos. Extensive experiments show LUNA achieves competitive visual fidelity compared to LBS-based approaches, while delivering realistic human motion and zero-shot cross-identity generalization across diverse driving modalities. To the best of our knowledge, LUNA is the first end-to-end 3D animatable model that supports implicit 2D driving.

Vision-Language Models (VLMs) have advanced rapidly invisual and semantic understanding, yet they remain weak at dynamic 4Dspatial grounding in video. On spatio-temporal reasoning benchmarks,raw-video input often yields low and variable accuracy. 4D Scene Graphs(4DSGs) can supply missing structure, but two interface gaps remain:identity facets (appearance, geometry, motion) are often encoded throughlanguage-side descriptors rather than modality-separated inputs, andmost training targets static spatial relations instead of dynamic 4D at-tributes such as trajectories and velocity. We introduce Visual Spatio-Temporal Anchor (VISTA), a multimodal node encoding that representseach tracked object as a persistent, queryable entity, delegating appear-ance to anchor crops processed by the VLM’s vision encoder and geome-try and motion to explicitly structured tokens. We introduce Real-TimeOpen-World Scene Understanding (ROSE), a training-free monocularpipeline (single GPU) that runs at 10 Hz. Beyond inference-time prompt-ing, the 4DSGs produced by ROSE can serve as structured training-timecontext: fine-tuning a VLM on 4DSG-augmented data teaches it to usetrajectory and velocity tokens for dynamic reasoning, yielding a new stateof the art on DSR-Bench (67.5, +8.6 over the previous best), with gainsacross all 13 subtasks. ROSE achieves competitive accuracy across allthree benchmarks under real-time, single-GPU monocular constraints.

Human perception of surface slant from texture exhibits systematic, graded biases that emerge reliably in psychophysical experiments. Prior work showed that unsupervised CNNs reproduce several human-like biases,whilesupervisedCNNsdonot.DoVision-LanguageModels(VLMs) exhibit similar competences? Across multiple VLM families and model scales,zero-shotandin-contextpromptingbothproducedistinctivefailures: slantispredictedatonlyasmallsetofanchors(e.g.,0°,±25°,±45°)withlittle dependence on stimulus field of view, optical slant, or surface curvature. Supervised fine-tuning partially remediates the failure, but residual anchoring persists. While success in high-level vision-language benchmarks might notrequiresensitivitytolow-levelgeometriccues,weinterpretanchoringas a failure at the representation-to-output language interface: not necessarily anabsenceofgeometricencoding,butafailuretoexpressitinagradedform.

Recent advancements in 3D Gaussian Splatting (3DGS) have enabled language-guided scene understanding. However, existing Referring 3D Gaussian Splatting (R3DGS) methods are fundamentally restricted to single-target queries. To reflect the ambiguity of real-world instructions, we introduce the Generalized Referring 3D Gaussian Splatting Segmentation (GR3DGS) task, which requires dynamically segmenting an arbitrary number of targets (0, 1, or N). To facilitate comprehensive evaluation of this new task, we construct two new benchmarks: GR-LERF and GR-ScanNet. Crucially, existing R3DGS paradigms exhibit fundamental technical bottlenecks that severely limit their performance on the GR3DGS task: they lack intrinsic 3D point-level understanding by operating merely on 2D rendered pixels, and they incur prohibitive computational overhead by requiring per-scene optimization to embed heavy semantic features. To dismantle these bottlenecks, we propose ZeroSplat, a novel training-free and zero-feature framework. ZeroSplat lifts 2D Vision-Language Model (VLM) priors into 3D space through robust multi-view geometric constraints. This strategy enables intrinsic point-level understanding without incurring any additional feature storage. Extensive experiments demonstrate that ZeroSplat significantly outperforms state-of-the-art methods across generalized and single-target scenarios while maintaining exceptional efficiency. Project Page: https://inkmind-ai.github.io/ZeroSplat

Estimating accurate camera poses, 3D scene geometry, and object mo-tion from in-the-wild videos is a long-standing challenge for classical structurefrom motion pipelines due to the presence of dynamic objects. Recent learning-based methods attempt to overcome this challenge by training motion estimatorsto filter dynamic objects and focus on the static background. However, their per-formance is largely limited by the availability of large-scale motion segmentationdatasets, resulting in inaccurate segmentation and, therefore, inferior structural 3Dunderstanding. In this work, we introduce the Dynamic Prior (D!"#$%) to robustlyidentify dynamic objects without task-specific training, leveraging the powerfulreasoning capabilities of Vision-Language Models (VLMs) and the fine-grainedspatial segmentation capacity of SAM2. D!"#$% can be seamlessly integrated intostate-of-the-art pipelines for camera pose optimization, depth reconstruction, and4D trajectory estimation. Extensive experiments on both synthetic and real-worldvideos demonstrate that D!"#$% not only achieves state-of-the-art performanceon motion segmentation, but also significantly improves accuracy and robustnessfor structural 3D understanding. Code is available.