ECCV 2026 论文列表

Ternary transformers offer extreme memory and compute ef-ficiency, but existing low-bit LoRA-based methods cannot directly fine-tune ternary weights. Current approaches either require dequantization,restoring low-bit base weights to higher precision to merge with adap-tation weight, or update only quantization parameters, preventing amerged model that remains ternary. We propose ternary multiplica-tive adaptation, which represents discrete updates of ternary weightssuch as sign flips or zeroing through a low-rank Kronecker factoriza-tion into two small ternary matrices applied element-wise to ternaryweights. This design is parameter-efficient and expressive, preserves theternary domain, and supports direct merging without dequantization.Experiments on six models across language and vision, including ternar-ized LLaMA-3 1B and 3B and a ternary ViT-B/16, demonstrate thatour method recovers much of the performance lost to quantization andoutperforms strong low-bit and ternary baselines. Code is available athttps://github.com/alexmanoo/ternary_adaptation.

Procedural material creation underpins applications in digitalcontent creation, visual effects, and 3D asset design. Achieving high-quality results requires more than reproducing node graphs—it demandsunderstanding the process by which experts construct materials. We for-mulate procedural material generation as retrieval-time process reasoningover expert demonstrations, elevating process to a first-class represen-tation beyond graph-only synthesis. Concretely, we represent expertworkflows as process traces: textual records of construction steps, param-eters, and design intent. To instantiate this idea, we use a pretrainedLLM-based ProcessSynthesizer to synthesize a process trace alignedwith a user’s intent and a pretrained LLM-based Compiler to groundthe process trace into an executable Blender material graph. Becauseprocedural expertise is most naturally conveyed through demonstrations,we leverage tutorial videos as a source of process knowledge and extracttextual, LLM-compatible traces using automated video analysis tools.In an expert study with five Blender artists (avg. 7.5 years of expe-rience), materials generated by reflecting expert demonstrations werefound to produce workflows requiring fewer edits, and more closely matchprofessional design strategies than methods operating solely on staticartifacts. A user study with 150 participants further shows that ourapproach achieves superior generation and editing performance comparedto prior procedural systems. All code, models, and data will be availableat https://materialapprentice.github.io.

Single-image reflection removal (SIRR) seeks to recover thetransmission layer from a mixture corrupted by reflections—a severelyill-posed problem. Existing methods operate in pixel space, where thenonlinear sRGB formation model entangles the two layers and limits gen-eralization. We observe that pretrained VAE latent spaces exhibit sub-stantially lower coherence between image layers compared to pixel space,providing a more favorable working space for decomposition. Buildingon this finding, we propose PRISM (Pretrained-latent Reflection ImageSeparation Model), which reinterprets SIRR as a latent linear separa-tion problem. Under an approximate additive formulation in latent space,PRISM learns a flow matching velocity field on a pretrained FLUX back-bone that recovers both transmission and reflection in a single forwardpass. To enforce robust disentanglement, we introduce a Latent Com-position Consistency (LCC) strategy that constructs synthetic mixturesby swapping reflection latents across samples and enforces consistent de-composition via a cycle loss. We further propose a Layer ContrastiveSeparation (LCS) loss that promotes semantic separation between layersthrough patch-level contrastive learning, without requiring explicit reflec-tion targets. Experiments on six benchmarks demonstrate that PRISMconsistently outperforms state-of-the-art methods by significant margins,with strong generalization to in-the-wild images.

Unified fashion generation integrates tasks like virtual try-onand garment reconstruction into a single model to reduce task-specificadaptation costs. However, naive parameter sharing across semanticallydistinct tasks induces negative transfer through severe inter-task gradientconflict. We propose OrthoTryOn, a unified framework mitigating thisinterference within a shared Low-Rank Adaptation (LoRA) module. ItsOrthogonal Subspace Projection (OSP) applies task-specific orthogonalrotations to bottleneck features, mapping them into decorrelated coordi-nate frames. To address residual semantic coupling at inference time, wefurther propose Fisher-guided Negative Guidance (FNG), a parameter-free strategy that utilizes diagonal Fisher information to quantify inter-task sensitivity overlap and explicitly repels generation trajectories fromthe most confusable task via Classifier-Free Guidance. Extensive exper-iments demonstrate that OrthoTryOn avoids the severe performancedegradation typical of naive unified training and even surpasses inde-pendently trained task-specific models, achieving state-of-the-art resultsacross multiple benchmarks while generalizing robustly across diversediffusion backbones. Code is available at https://github.com/NJU-PCALab/OrthoTryOn.

Previous feed-forward 4D reconstruction methods either pre-dict per-frame static point clouds, ignoring foreground motion, or esti-mate point cloud trajectories while being limited to small camera mo-tions. This limits their ability to aggregate observations over time andreconstruct complete dynamic scenes under large viewpoint changes. Toaddress this, we propose OmniX, a feed-forward 4D reconstruction frame-work that predicts dense 3D point trajectories for every pixel fromvideos with large camera motion. 1) OmniX separates dynamic mo-tion modeling from static geometry prediction and represents motionwith a small set of dynamic tokens. Leveraging the sparse and low-rankstructure of 3D motion, these tokens generate trajectory fields for allpixels in all images while efficiently preserving global interactions. 2) Tofacilitate training, we build an automatic UE5-based 4D data engine andintroduce a dataset of 80k scenes and 1.28M multi-view videos with fullgeometric annotations. OmniX achieves state-of-the-art results on dense3D point trajectory prediction and 3D point tracking, with competitiveperformance on video depth estimation and camera pose estimation.

Recent advancements in LVLMs necessitate robust bench-marks for complex, visually grounded reasoning. A critical limitation isidentified in many document understanding benchmarks: visual contentis often reducible to text, enabling high performance without genuine vi-sual grounding. To address this limitation, OmniMapBench is introducedto foster visual-centric reasoning for map documents. The benchmarkcomprises 2,096 manually annotated question-answer pairs across 1,603map documents from nine categories. It is designed to probe a hierar-chy of skills, ranging from perception to multi-step visual reasoning. Toquantify benchmark properties, a simple yet effective benchmark-levelmetric is proposed: the Visual Dependency Index (VDI), defined as theaccuracy drop when images are replaced with question-agnostic descrip-tions. OmniMapBench exhibits higher VDI than established benchmarks,which quantitatively validates its focus on irreducible visual reasoning.Comprehensive evaluations of 25 leading LVLMs are conducted on Om-niMapBench. A significant performance gap is observed, with the top-performing model achieving only 75.03% accuracy. This result under-scores the challenges posed by OmniMapBench to current LVLMs. Thiswork aims to catalyze progress in visual-centric reasoning for documentunderstanding of LVLMs. The dataset and code are publicly available athttps://github.com/SIGMME/OmniMapBench.

Data augmentation is fundamental to training modern deepvision and multimodal models. While individual methods, such as Ran-dAug, CutMix, Mixup, RandErase, and DropPath, offer strong regular-ization effects, their combined use has saturated in performance due tooverlapping functionalities, and aggressive pixel-level manipulations maydisrupt delicate cross-modal alignment. This saturation motivates thesearch for a new augmentation axis within the embedding space ratherthan the input space. We introduce Aether, a simple plug-in method thatapplies diffusion-style random perturbations in the embedding space viacontrolled alpha-mixing, specifically designed to provide isotropic regu-larization that remains semantically consistent. Inspired by feature-spaceperturbations in language models and image degradation in generativepretraining, Aether induces mild yet effective perturbations that smooththe representations without compromising the fine-grained structuralinformation required for strong vision-language encoders. Across diversearchitectures and across multiple recognition tasks, Aether delivers consis-tent gains over the advanced recipe combining CutMix, Mixup, DropPath,and RandAug—a level of improvement rarely observed with modern aug-mentation alternatives. Notably, Aether demonstrates superior effective-ness in multi-modal alignment, succeeding where traditional pixel-spaceaugmentations fail by providing a stable, isotropic regularization signalthat respects the integrity of the high-dimensional feature space. Code isavailable at https://github.com/naver-ai/aether.

While 10B-level industrial foundation models have pushedthe boundaries of image inpainting, their prohibitive computational costsseverely hinder practical deployment. Constructing a highly optimizedtask-specific specialist offers a promising solution; however, extreme struc-tural compression inevitably triggers a severe representation bottleneck.To conquer this, we propose Moebius, a highly efficient lightweight in-painting framework. We systematically reconstruct the diffusion back-bone by introducing the Local-λ Mix Interaction (LλM I) block. Com-prising Local-λ and Interactive-λ modules, it elegantly summarizes spa-tial contexts and global semantic priors into fixed-size linear matrices,preserving complex latent interactions while drastically shedding param-eters. Furthermore, to unlock the full representational capacity of thishighly compact architecture, we synergistically pair it with an adap-tive multi-granularity distillation strategy. Operating strictly within thelatent space to avoid expensive pixel-space decoding, this strategy dy-namically balances multiple gradient-based losses to achieve high-fidelityalignment. Extensive experiments across natural and portrait bench-marks demonstrate that this optimal synergy enables Moebius to rival oreven surpass the generation quality of the 10B-level industrial generalistFLUX.1-Fill-Dev. Remarkably, Moebius achieves this using less than 2%of the parameters (0.22B vs. 11.9B) while delivering a > 15× accelerationin total inference time, setting a new efficiency standard for high-fidelityinpainting. Project page at https://hustvl.github.io/Moebius.

Grounding natural-language instructions into executable con-tinuous control remains a core challenge for quadruped vision-language-action (VLA) systems due to the gap between high-level semantic reason-ing and low-level locomotion actuation. Existing approaches often relyon implicit reasoning or purely behavioral supervision, leading to unsta-ble long-horizon grounding and limited robustness in real-world deploy-ment. To address these issues, we present MobileVLA-R1, a unifiedquadruped VLA framework that explicitly aligns hierarchical reason-ing with continuous control. Our method introduces a multi-granularitychain-of-thought (CoT) supervision scheme for embodied trajectories, in-stantiated in MobileVLA-CoT, together with a two-stage training paradigmthat combines supervised CoT alignment and reinforcement learning forreasoning-to-control consistency. We conduct comprehensive evaluationson VLN and quadruped VLA benchmarks, including VLN-CE, QUARD,and real-world Unitree Go2 deployment. MobileVLA-R1 consistently out-performs strong baselines, achieving about 5% gains on key metrics whiledemonstrating deployment-time closed-loop execution under a fixed hy-brid inference setup.

Detecting covariate shift is critical for building reliable vi-sion systems. While most prior work focuses on improving robustness toshift, explicitly detecting covariate shift remains underexplored. Existingapproaches typically rely on fully supervised training, requiring labeledexamples from both original and shifted distributions, which is often im-practical. In this paper, we show that covariate shift detection can beeffectively addressed with weaker supervision using Positive–Unlabeled(PU) learning. However, under covariate shift, in-distribution and shifteddata overlap significantly, making classical PU methods unstable and sen-sitive to noise. To overcome this challenge, we introduce Spectral PUNeighborhood Annotation (S-PUNA), a geometry-aware frame-work that progressively discovers shifted data by leveraging the localmanifold structure of visual features. Extensive experiments show thatS-PUNA achieves state-of-the-art performance in PU settings and re-markably matches the performance of fully supervised methods. More-over, our approach transfers robustly across different types of shifts,demonstrating strong generalization capabilities. Code is available athttps://github.com/fira7s/S-PUNA.

Answering questions accurately and efficiently in embodiedscenarios presents significant challenges due to limited computationaland memory resources for Vision Language Model (VLM) inference. Ex-isting methods adopt visual search key frame retrieval method to se-lect critical question-related key frames for VLM input. However, visualsearch methods are inefficient because they require visual search amongthousands of video frames for each individual user query. In this work, wepropose a memory tree guided key frame selection paradigm for efficient3D question answering in embodied scenarios. Our method leveragesa compact and reusable 3D scene representation, termed MemTree3D,which supports real-time online construction leveraging camera 6-DoFposes. MemTree3D captures multi-level 3D scene information, enabling aLarge Language Model to efficiently query and retrieve question-relevantkey frames through our scoring-based frame selection without repro-cessing the entire video stream. On OpenEQA, our method improvesthe LLM-Match of GPT-4o by 17.4%, LLaVA-OneVision-7B by 5.8%,outperforms existing visual search methods. Our code is available athttps://github.com/hsiangwei0903/MemTree3D.

Reasoning-induced vision-language models (VLMs) advanceimage quality assessment (IQA) with textual reasoning, yet their scalarscores often lack sensitivity and collapse to a few values, so-called discretecollapse. We introduce ME-IQA, a plug-and-play, test-time memory-enhanced re-ranking framework. It (i) builds a memory bank and re-trieves semantically and perceptually aligned neighbors using reason-ing summaries, (ii) reframes the VLM as a probabilistic comparator toobtain pairwise preference probabilities and fuse this ordinal evidencewith the initial score under Thurstone’s Case V model, and (iii) per-forms gated reflection and consolidates memory to improve future deci-sions. This yields denser, distortion-sensitive predictions and mitigatesdiscrete collapse. Experiments across multiple IQA benchmarks showconsistent gains over strong reasoning-induced VLM baselines, existingnon-reasoning IQA methods, and test-time scaling alternatives.

High-quality world-scale 3D data is scarce, making completeand generalizable 3D world generation difficult. Existing approaches typi-cally rely on either generated images/videos or domain-specific 3D worlddatasets, resulting in worlds that are often incomplete or restricted tonarrow domains. In this paper, we present Map2World, a map-conditionedframework that converts semantic layouts into complete and generalizablelarge-scale 3D worlds. Given a segment map with per-region text prompts,Map2World uses the map as a controllable world-level interface and lever-ages TRELLIS, a powerful pretrained 3D asset generator, as a general3D prior. To scale asset-level priors to world-level synthesis, Map2Worldenables arbitrary spatial expansion via MultiDiffusion, controls globalscale through initial noise optimization, and enhances local geometry andappearance with a dedicated enhancer. Experiments demonstrate thatMap2World generates complete, layout-controllable, scale-consistent, anddetailed 3D worlds across diverse layouts and semantics.

Self-supervised RGB-T tracking seeks to learn robust cross-modal representations without dense manual annotations. A commonsupervision strategy relies on forward-backward tracking loops, but inpractice, it is highly vulnerable to error accumulation and modality driftcaused by the inconsistent reliability of visible and thermal cues. To ad-dress this issue, we propose LGCTrack, a self-supervised RGB-T trackerunderpinned by a local-to-global coordination strategy within a closed-loop verification paradigm. At the local level, a hierarchical coordina-tion module performs layer-wise, reliability-aware token interaction toenable adaptive cross-modal fusion. At the global level, a turning-phasecoordination module refines the target representation during the criticalforward-to-backward transition, anchoring target semantics and improv-ing backward tracking. By jointly coordinating cross-modal interactionat two granularities, LGCTrack intrinsically filters pseudo-label noise andmitigates drift. Experiments on standard RGB-T tracking benchmarksshow that LGCTrack consistently outperforms existing self-supervisedmethods and substantially narrows the gap to fully supervised trackers.Our code is available at https://github.com/zyyyya/LGCTrack.

Current multimodal latent reasoning often relies on externalsupervision (e.g., auxiliary images), ignoring intrinsic visual attentiondynamics. In this work, we identify a critical Perception Gap in distil-lation: student models frequently mimic a teacher’s textual output whileattending to fundamentally divergent visual regions, effectively relyingon language priors rather than grounded perception. To bridge this, wepropose LaViT, a framework that aligns latent visual thoughts ratherthan static embeddings. LaViT compels the student to autoregressivelyreconstruct the teacher’s visual semantics and attention trajectories priorto text generation, employing a curriculum sensory gating mechanism toprevent shortcut learning. Extensive experiments show that LaViT sig-nificantly enhances visual grounding, achieving up to +16.9% gains oncomplex reasoning tasks and enabling a compact 3B model to rival largermodels on selected evaluated benchmarks.

In this paper, we present JoVA, a streamlined frameworkthat unifies joint video-audio generation and editing. While existingmethods often rely on fragmented, task-specific architectures or complexfusion mechanisms, JoVA employs native joint representation learningfor direct video, audio, and text interaction in a dual-branch architec-ture. This design eliminates redundant alignment modules and effectivelyunifies diverse multimodal tasks within a single model. Furthermore, weutilize channel-wise conditioning for flexible image and video reference toavoid massive token expansion, alongside a mouth-area loss to enhancelip alignment. To fully empower and systematically evaluate this frame-work, we construct a comprehensive training corpus encompassing video-audio generation and editing datasets, and introduce unified benchmarkstailored for these multimodal tasks. Extensive experiments demonstratethat JoVA achieves state-of-the-art performance across benchmarks, es-tablishing it as an extensible framework for versatile content creation.Project page: https://visual-ai.github.io/jova

Multi-task learning (MTL) has been widely adopted for itsability to simultaneously learn multiple tasks. While existing gradientmanipulation methods often yield more balanced solutions than simplescalarization-based approaches, they typically incur a significant compu-tational overhead of O(K) in both time and memory, where K is thenumber of tasks. In this paper, we propose LDC-MTL, a simple andscalable loss discrepancy control approach for MTL, formulated froma bilevel optimization perspective. Our method incorporates two keycomponents: (i) a bilevel formulation for fine-grained loss discrepancycontrol, and (ii) a scalable first-order bilevel algorithm that requiresonly O(1) time and memory. Theoretically, we prove that LDC-MTLguarantees convergence not only to a stationary point of the bilevelproblem with loss discrepancy control but also to an ϵ-accurate Paretostationary point for all K loss functions under mild conditions. Extensiveexperiments on diverse multi-task datasets demonstrate the superior per-formance of LDC-MTL in both accuracy and efficiency. Code is availableat https://github.com/OptMN-Lab/LDC-MTL.

Robust object detection under adverse visual conditions re-mains a long-standing challenge for multi-modal perception systems. Ex-isting fusion-based methods typically require both RGB and infrared(IR) inputs, and treat them equally during both training and inference,which compromises their robustness when the RGB modality becomesunreliable or unavailable. In this case, we propose InfraNet, an IR-centric quality-aware framework that regulates RGB guidance duringtraining and supports flexible RGB–IR or IR-only deployment. InfraNetemploys an asymmetric architecture where the primary IR pathway ex-tracts multi-scale infrared features for predictions, while the auxiliaryRGB pathway provides reliability-controlled supervisory signals. Thecore of InfraNet is QualGate, a quality-aware fusion module that learnsa task-oriented control signal to suppress unreliable RGB guidance andcompensate IR features during cross-modal training. Built upon InfraNet,we design two architectural variants: a lightweight IR-only architectureInfraNet-IR and an RGB–IR architecture InfraNet-RGB-IR. Our methodis evaluated through extensive experiments on four benchmark datasets(LLVIP, FLIR-Aligned, M3 FD, and DroneVehicle), showing strong orcompetitive accuracy in challenging low-light and adverse weather con-ditions. Notably, InfraNet maintains high efficiency in IR-only inference,making it both accurate and computationally efficient.

Achieving human-like reasoning in Vision-Language Models(VLMs) remains a long-standing challenge. Recent approaches leverageChain-of-Thought (CoT) rationales generated by human annotators orproprietary models to improve reasoning, which is costly and difficultto scale. Self-training offers a promising alternative by using model’sown outputs as supervision. However, existing methods often suffer fromvisual hallucinations - where rationales describe non-existent visual con-tent, and language shortcuts - where predictions rely on textual priorsrather than true visual grounding, as rationales are typically filtered onlyby answer correctness without verifying visual perception. To addressthis limitation, we propose a perception-verified self-training frameworkthat enforces visually grounded reasoning. First, our method employs aCoT template (caption-reasoning-conclusion) that disentangles percep-tion from reasoning, enabling independent verification of visual under-standing. To compensate for the absence of ground-truth captions, wepropose PerceptEval, an unsupervised method that evaluates captionquality based on its alignment with visual and textual elements presentin the image. Using caption verification together with answer correct-ness, we partition the data into three subsets: easy (correct caption andconclusion), medium (correct caption but incorrect conclusion), and hard(incorrect caption). Building on this partitioning, we design a two-stagecurriculum learning strategy. In Stage 1, the model is trained on easyexamples and subsequently in Stage 2, medium samples are incorporatedthrough a caption-guided reasoning enhancement procedure thatregenerates reasoning conditioned on verified captions. Only regeneratedsamples with the correct conclusions are retained. This ensures trainingin both stages is done exclusively on perceptually grounded reasoning,reducing hallucinations and language shortcuts. Extensive experimentsacross diverse domains and models demonstrate improvements of up to16% over standard self-training baselines. This demonstrates that theproposed framework is a scalable and cost-effective solution to advanc-ing multimodal reasoning without manually annotated CoT rationales.