ECCV 2026 Papers

Estimating the 3D trajectory of every pixel from a monocularvideo is crucial and promising for a comprehensive understanding of the3D dynamics of videos. Recent monocular 3D tracking works demon-strate impressive performance, but are limited to either tracking sparsepoints on the first frame or a slow optimization-based framework fordense tracking. In this paper, we propose a feedforward model, calledTrack4World, enabling an efficient holistic 3D tracking of every pixel inthe world-centric coordinate system. Built on the global 3D scene repre-sentation encoded by a VGGT-style ViT, Track4World applies a novel3D correlation scheme to simultaneously estimate the pixel-wise 2D and3D dense flow between arbitrary frame pairs. The estimated scene flow,along with the reconstructed 3D geometry, enables subsequent efficient3D tracking of every pixel of this video. Extensive experiments on multiplebenchmarks demonstrate that our approach consistently outperforms ex-isting methods in 2D/3D flow estimation and 3D tracking, highlighting itsrobustness and scalability for real-world 4D reconstruction tasks. Projectpage: https://jiah-cloud.github.io/Track4World.github.io/.

Tone style transfer for photo retouching aims to adapt thestylistic tone of the reference image to a given content image. However,the lack of high-quality large-scale triplet datasets with stylized groundtruth forces existing methods to rely on self-supervised or proxy objec-tives, which limits model capability. To mitigate this gap, we design adata construction pipeline to build TST100K, a large-scale dataset of100,000 content-reference-stylized triplets. At the core of this pipeline,we train a tone style scorer to ensure strict stylistic consistency for eachtriplet. In addition, existing methods typically extract content and ref-erence features independently and then fuse them in a decoder, whichmay cause semantic loss and lead to inappropriate color transfer and de-graded visual aesthetics. Instead, we propose ICTone, a diffusion-basedframework that performs tone transfer in an in-context manner by jointlyconditioning on both images, leveraging the semantic priors of generativemodels for semantic-aware transfer. Reward feedback learning using thetone style scorer is further incorporated to improve stylistic fidelity andvisual quality. Experiments demonstrate the effectiveness of TST100K,and ICTone achieves state-of-the-art performance on both quantitativemetrics and human evaluations. The project page is available online.

Effective Human-Robot Collaboration (HRC) requires robotsto accurately infer human intents under partial observations. However,existing benchmarks mostly rely on idealized views, overlooking the spa-tiotemporal incompleteness such as low-angle truncation and dynamicocclusions inherent in real-world quadruped robot perception. Further-more, current evaluations lack diagnostic transparency, failing to distin-guish whether Vision-Language Models (VLMs) perform reliable com-pensatory reasoning or merely hallucinate based on statistical biaseswhen visual features are degraded. To bridge these dual gaps, we in-troduce CoT-HRC, a large-scale benchmark built in Habitat 3.0 thatcouples simulated robot-centric views with explicit hierarchical Chain-of-Thought (CoT) annotations. We propose a diagnostic protocol featur-ing the Step-wise Consistency Assessment (SCAR ) to penalize spuriousaccuracy by enforcing intermediate logical consistency, and the Condi-tional Reasoning Accuracy (CRA) to explicitly decouple models’ un-derlying reasoning capabilities from visual perception failures. Exten-sive experiments across state-of-the-art VLMs reveal that while currentmodels inherently possess strong physical commonsense, their reason-ing chains are severely disrupted by the dynamic geometric constraintsof egocentric views. Our fine-grained ablations further highlight criti-cal bottlenecks in temporal aggregation and the compensatory role ofsemantic priors, establishing CoT-HRC as a vital stepping stone for ro-bust embodied intent inference.The full benchmark is publicly availableat https://github.com/Thus-cx/CoT-HRC.

The visual aesthetics of photographs are deeply influencedby lens characteristics such as aperture shape, optical vignetting and op-tical diffraction, which together define a camera’s unique optical style.Existing lens effect rendering methods primarily focus on accurately sim-ulating the blur transition from small to large apertures but overlook thestylistic aspects of lens effects. As a result, they fail to produce diversebokeh effects under large apertures or capture distinctive photographicphenomena such as starbursts that emerge under small apertures. In thiswork, we introduce LensStyle, a unified framework for controllable styl-ized lens effect rendering that explicitly models lens aesthetics throughjoint continuous–discrete control. Our model incorporates a Dual-PathController that disentangles continuous optical parameter modulation(e.g., focus distance and blur strength) from discrete lens-style condi-tioning (e.g., circular, polygonal, donut, cat-eye, and starburst effects),enabling fine-grained, interpretable, and physically grounded lens manip-ulation within a single unified framework. To support model training, wecurate a comprehensive MultiLens dataset containing multi-lens imagepairs synthesized under real optical constraints. Extensive experimentsdemonstrate that LensStyle achieves superior realism, controllability, andaesthetic quality compared with existing lens effect rendering approachesand diffusion-based image editing models, advancing computational pho-tography toward multiple-lens-style simulation.

Existing video temporal grounding methods typically oper-ate under a closed-world assumption, relying on the implicit premise thatevery natural language query must have a corresponding event some-where in the video. This always-answer paradigm severely compressesthe decision boundary during training, as models are seldom exposed tothe alternative hypothesis that the queried event may be absent. Con-sequently, when users ask about specific events that cannot be matchedin the video within an open-world environment, these models exhibitsystematic temporal hallucinations. This may stem from incorrect at-tribution of textual, visual, or multimodal information, leading them toconfidently generate plausible time frames for events that do not actuallyexist. To address this limitation, we propose HRVTG, a test-time adapta-tion framework that dynamically calibrates the decision boundary of themodel during inference. Built upon a frozen grounding backbone, the pro-posed framework constructs three types of counterfactual hallucinationprobes in a self-supervised manner, targeting the aforementioned halluci-nations. We dynamically reshape the decision boundary through absten-tion signals from counterfactual probes and consistency rewards fromgenuine queries. This is executed by introducing the GRPO algorithm,effectively reformulating the task into an online policy optimization prob-lem in reinforcement learning. Furthermore, we introduce decoupled met-rics to independently evaluate hallucination resistance and real-eventgrounding. Experiments indicate that our method suppresses halluci-nated responses to fabricated queries. It achieves accurate selective pre-diction while maintaining strong grounding performance on real events.The code will be released at https://github.com/CVL-hub/HRVTG.

VAR has gained widespread popularity due to its next-scaleprediction paradigm. However, it faces substantial performance bottle-necks when handling complex scenes with multiple objects and attributes.Existing diffusion-based enhancement methods fail to adequately addressthe unique challenge of cross-scale error propagation and accumulationin VAR. To this end, we propose SynVAR, the first training-free enhance-ment framework specifically tailored for the VAR paradigm, which intro-duces a spatial-semantic collaborative control strategy to effectively sup-press propagation error and improve generation quality. SynVAR com-prises three key components: (1) Global guidance to ensure reasonablespatial structure in the early stages, (2) Receptive field constraints tomitigate early-stage semantic confusion, (3) High-frequency compensa-tion to recover fine-grained details. Extensive quantitative and qualita-tive experiments demonstrate the significant improvements in the abilityof SynVAR to enhance the VAR’s capability for complex scene modeling.

While synthetic data has proven effective for improving sci-entific reasoning in the text domain, multimodal reasoning remains con-strained by the difficulty of synthesizing scientifically rigorous images.Existing Text-to-Image (T2I) models often produce outputs that arevisually plausible yet scientifically incorrect, resulting in a persistentvisual–logic divergence that limits their value for downstream reason-ing. Motivated by recent advances in next-generation T2I models, weconduct a systematic study of scientific image synthesis across genera-tion paradigms, evaluation, and downstream use. We examine both di-rect pixel-based generation and programmatic synthesis, and instantiateImgCoder as a structured implementation of the code-driven workflow,following an explicit “understand → plan → code” prompting strategyto encourage clearer structural specification. To rigorously assess scien-tific correctness, we introduce SciGenBench, which evaluates generatedimages based on information utility and logical validity. Our evaluationreveals systematic failure modes in pixel-based models and highlightsa fundamental expressiveness–precision trade-off. Finally, we show thatfine-tuning Large Multimodal Models (LMMs) on rigorously verified syn-thetic scientific images yields consistent reasoning gains, with potentialscaling trends analogous to the text domain, validating high-fidelity sci-entific synthesis as a viable path to unlocking massive multimodal rea-soning capabilities.

Visual generation is increasingly ubiquitous in diverse do-mains, from text-to-image/video synthesis to multimodal interactive cre-ation. Yet prevailing monolithic models remain fundamentally constrainedby their inability to learn cumulatively and evolve autonomously, whichis a limitation we term the “perpetual novice” problem. They lack mech-anisms for structuring experience into reusable knowledge and there-fore rely on brittle, “from-scratch” reasoning for each task, resultingin poor compositional generalization and inefficient knowledge reten-tion. Motivated by these limitations, we propose SymbOmni, an agenticomni-model designed for cumulative evolution through Symbolic Con-cept Learning. At its core is the Symbolic Concept Box, an optimizablememory module that abstracts low-level operations into reusable Sym-bolic Workflow Instructions. SymbOmni operates through an induction-transduction cycle: experiences are abstracted into symbolic concepts(induction), which are then adaptively composed to solve novel tasks(transduction). The training is done by verbalized backpropagation withlanguage-based feedback to enable continuous self-improvement withoutgradient-based model fine-tuning. Comprehensive experiments validatethat (I) SymbOmni significantly outperforms existing agent-based sys-tems for iterative creation and also surpasses closed-source models (e.g.,Nano Banana, GPT-Image-1) in both image quality and task successrates; (II) SymbOmni effectively reduces token consumption by over 40%while maintaining competitive generation quality; and (III) SymbOmnienables effective continual learning by achieving cumulative gains acrossmultiple online-learning benchmarks and setting a new state of the art.

We present SharpGS, a differentiable blur-driven density con-trol, which enhances the quality of 3D Gaussian Splatting (3DGS). Stan-dard 3DGS often struggles to capture intricate details, particularly intextured patterns and object boundaries, due to the limited sensitivityof its density control. Simple finer densification can result in excessiveprimitive counts with marginal quality gain. To address this, we intro-duce blur as an effective perceptual cue, leveraging CUDA-based differ-entiable blur. Our density control estimates per-primitive blur levels bycomparing reconstructions to blurred ground-truth images, and identifieshigh-frequency regions where additional primitives are required. Whilethis naturally produces more primitives, we counterbalance this by sup-pressing redundant primitives. Specifically, we penalize the blur levelsof inherently smooth regions such as sky, and regularize the opacitiesof potential split/clone candidates. We experimentally demonstrate thatSharpGS greatly improves the state-of-the-art 3DGS methods in termsof quality, while keeping learned primitives compact.

Establishing accurate point-to-point correspondences betweennon-rigid 3D shapes remains a critical challenge, particularly under non-isometric deformations and topological noise. Existing functional mappipelines suffer from ambiguities that geometric descriptors alone can-not resolve, and spatial inconsistencies inherent in the projection of trun-cated spectral bases to dense pointwise correspondences. In this paper,we introduce SGMatch, a learning-based framework that couples 3D-lifted semantic cues with trajectory-level feature transport regulariza-tion. Specifically, we design a Semantic-Guided Local Cross-Attentionmodule that integrates semantic features from vision foundation modelsinto geometric descriptors while preserving local structural continuity.Furthermore, we adapt conditional flow matching as a time-conditionedfeature transport regularizer that promotes spatially coherent point-wiserecovery. Experimental results on multiple benchmarks demonstrate thatSGMatch achieves competitive performance across near-isometric set-tings and consistent improvements under non-isometric deformations andtopological noise.

Transparency of neural networks’ internal reasoning is atthe heart of interpretability research, adding to trust, safety, and under-standing of these models. The field of mechanistic interpretability hasrecently focused on studying task-specific computational graphs, definedby connections (edges) between model components. Such edge-basedcircuits have been defined in the context of large language models, yetvision-based approaches so far only consider neuron-based circuits. Thesetell which information is encoded, but not how it is routed through thecomplex wiring of a neural network. In this work, we investigate whetheruseful mechanistic circuits can be identified through computational graphsin vision transformers. We propose an effective method for AutomaticVisual Circuit Discovery (Vi-CD) that recovers class-specific circuitsfor classification, identifies circuits underlying typographic attacks inCLIP, and discovers circuits that lend themselves for steering to correctharmful model behaviour. Overall, we find that insightful and actionableedge-based circuits can be recovered from vision transformers, addingtransparency to the internal computations of these models.

Large Multimodal Models (LMMs) have achieved remark-able success on images and short videos, yet scaling them to long videosremains challenging due to frame-centric tokenization and limited con-text windows. 3D geometry provides a natural compression mechanismfor visual streams: depth and camera pose enable observations from mul-tiple views and time steps to be fused into a persistent, world-aligned rep-resentation. While recent 3D LMMs leverage geometry-aware representa-tions to improve spatial reasoning, they continue to lag behind specialist3D perception systems on grounding and segmentation tasks. We arguethat a key limitation is geometry-aware decoding: existing methods com-municate 3D predictions through language tokens, proposal selection, orlightweight grounding queries, creating a bottleneck between languagereasoning and dense geometric prediction. Building on these insights, weintroduce Qwen-3D, a geometry-aware LMM that compresses visual in-formation within the Qwen backbone using multi-view geometric cues,enabling efficient long-horizon visual reasoning over static scenes. Qwen-3D augments visual tokens with 3D Rotary Positional Embeddings, al-lowing attention to operate directly in 3D scene space rather than acrossindependent image frames and thereby facilitating scalable cross-viewand temporal reasoning. To bridge language and geometry, Qwen-3D in-corporates a query-based segmentation decoder that grounds languagedirectly in the underlying 3D scene representation, unifying referentialgrounding, instance segmentation, and visual question answering acrossboth images and videos. Across a diverse set of benchmarks, Qwen-3Dsurpasses existing 3D LMMs and outperforms several large proprietary2D models. Notably, Qwen-3D achieves these improvements while main-taining strong performance on standard 2D vision–language benchmarksby jointly training on 2D and 3D data. Our code and checkpoints can befound at the project website https://qwen-3d.github.io/.

We introduce Point-supervised Multi-Object Tracking (PS-MOT) as a cost-effective alternative to traditional bounding box su-pervision, shifting the focus from spatial fitting to topological center-driven representation. However, PS-MOT faces challenges, e.g., spatialambiguity and identity drift due to the lack of explicit geometric struc-ture and scale constraints. To address these, we propose PS-Track, ahierarchical pipeline transitioning from points to instances across data,model, and loss levels. At the data level, we introduce Temporal-FeedbackPrompting (TFP) to evolve points into temporally consistent pseudo-labels using negative spatial cues and motion priors. At the model level,we design the Point-Excited Wavelet Attention (PEWA) module, whichleverages semantic correlations to activate high-frequency components,“hallucinating” object boundaries. At the loss level, Uncertainty-GuidedGaussian Learning (UGL) models pseudo-labels as probabilistic distri-butions, dynamically calibrating supervision intensity. Experiments onDanceTrack, EmboTrack, SportsMOT, and JRDB demonstrate that PS-Track provides a feasible and effective point-supervised alternative acrossdiverse tracking scenarios, establishing a new state-of-the-art for point-supervised tracking. The source code is available at https://github.com/xifen523/PS-MOT.

Although autoregressive (AR) models have demonstrated re-markable success in image generation, extending these models to layout-conditioned generation remains challenging due to the sparse natureof layout conditions and the risk of feature entanglement. We presentStructured Masking for AR-based Layout-to-Image (SMARLI), a novelframework that effectively integrates spatial layout constraints into theAR generation process. To equip AR models with layout control, a struc-tured masking strategy is applied to the attention computation to governthe interaction among the global prompt, layout, and image tokens. Thisdesign prevents the misassociation of different regions with their corre-sponding descriptions while enabling the sufficient injection of layoutconstraints into the generation process. To alleviate the exposure bias ofAR models and further enhance generation quality and layout accuracy,we incorporate a Group Relative Policy Optimization (GRPO) post-training scheme. We adapt it to the next-set-based paradigm and intro-duce a specifically designed layout reward, which is coordinated with animage quality reward to guide policy optimization in a balanced manner.Experimental results demonstrate that SMARLI seamlessly integrateslayout tokens with text and image tokens without compromising genera-tion quality, and the proposed masking strategy and post-training schemecan also be transferred to standard next-token-based AR models. Theproposed framework achieves superior layout control while maintainingthe structural simplicity and generation efficiency of AR models.

Recent advancements in Unified Multimodal Models (UMMs)have significantly advanced text-to-image (T2I) generation, particularlythrough the integration of Chain-of-Thought (CoT) reasoning. How-ever, existing CoT-based T2I methods largely rely on abstract natural-language planning, which lacks the precision required for complex spa-tial layouts, structured visual elements, and dense textual content. In thiswork, we propose CoCo (Code-as-CoT), a code-driven reasoning frame-work that represents the reasoning process as executable code, enablingexplicit and verifiable intermediate planning for image generation. Givena text prompt, CoCo first generates executable code that specifies thestructural layout of the scene, which is then executed in a sandboxed en-vironment to render a deterministic draft image. The model subsequentlyrefines this draft through fine-grained image editing to produce the fi-nal high-fidelity result. To support this training paradigm, we constructCoCo-10K, a curated dataset containing structured draft–final imagepairs designed to teach both structured draft construction and correc-tive visual refinement. Empirical evaluations on StructT2IBench, OneIG-Bench, and LongText-Bench show that CoCo achieves improvements of+68.83%, +54.8%, and +41.23% over direct generation, while also out-performing other generation methods empowered by CoT. These resultsdemonstrate that executable code is an effective and reliable reasoningparadigm for precise, controllable, and structured text-to-image genera-tion. The code is available at: https://github.com/micky-li-hd/CoCo

Reconstructing 4D animals from monocular videos is chal-lenging due to large inter-species variation, complex articulations, andthe lack of reliable templates. Existing approaches typically rely on eitherstrict category-specific priors that restrict generalization, or unconstrainedgenerative models that sacrifice input fidelity. To bridge this gap, wepresent a progressive test-time optimization framework built on 3D Gaus-sian Splatting for high-fidelity 4D animal reconstruction from a singlevideo. Our key insight is that a coarse shape prior suffices when coupledwith a progressive strategy that disentangles articulated pose from non-rigid deformation. Specifically, we employ a symmetry-aware temporalencoding that exploits bilateral cues while absorbing camera estimationdrift and a part-conditioned deformation mechanism guided by learn-able part anchors and a learnable skinning field. Extensive experimentsdemonstrate that our approach generalizes robustly across diverse species,achieving superior geometric accuracy, temporal consistency, and visualfidelity compared to existing baselines, even under severe prior mismatch.Project page: https://syl-322.github.io/ReWild4D/

Despite rapid progress in video diffusion models (VDMs),ensuring semantic adherence and physical commonsense remains a fun-damental challenge, as even state-of-the-art systems frequently violatereal-world physical laws, such as dynamics, thermodynamics, and op-tics. While Direct Preference Optimization (DPO) has become a popu-lar post-training strategy for aligning generative models, existing videoDPO methods to improve physical commonsense suffer from high compu-tational costs, poorly matched preference pairs, and ambiguous global su-pervision that fails to localize physical violations. We propose PhysPO,a physics-aware local preference optimization framework for physicallyconsistent video generation. We first introduce CounterPhyPipe, a physics-aware counterfactual data construction pipeline that forms video pref-erence pairs with consistent global semantics and local physical viola-tions, enabling meaningful comparisons for preference learning. Thenwe leverage high-frequency decomposition to introduce a physical mask,which localizes regions where physical state transitions may occur. Fi-nally, we develop PhysPO, a physics-aware DPO framework that con-centrates supervision on physically relevant regions while enforcing neu-trality on background regions. This mechanism reduces gradient noiseand mitigates shortcut optimization, encouraging the model to focus ongenuine physical discrepancies rather than superficial cues. Extensiveexperiments demonstrate that PhysPO significantly improves physicalcommonsense without compromising semantic adherence. Our projectpage is here3 .

Dense video captioning aims to generate temporally groundeddescriptions of video events, benefiting both event-level video understand-ing and generation. In this domain, autoregressive video large languagemodels have emerged as a prevalent paradigm due to their strong gen-erative and cross-modal modeling capacity. However, generating densecaptions under the token-by-token paradigm severely limits inference effi-ciency and hinders scalability as video length and event density increase.In this work, we propose a parallelized autoregressive framework thatnot only improves generation efficiency but also enhances temporallygrounded captioning performance. Our key insight is to exploit the weaklocal dependencies across temporally distinct events to restructure thecausal dependency graph, thereby enabling lossless parallel generation.Specifically, tokens with weak cross-event dependencies can be decoded inparallel, while tightly coupled tokens within each event retain sequentialdecoding to preserve local semantic coherence. To realize this insight, weintroduce two key components for lossless parallel decoding: (1) a latentglobal planning mechanism that automatically learns the event-level struc-ture and produces compact tokens encoding global inter-event causalitywhile adaptively aggregating event-level audio-visual semantics, guidingsubsequent dependency restructuring and parallel decoding; and (2) anevent-factorized parallel decoding mechanism that effectively balanceslocal focus with global inter-event awareness. Experiments on variousbenchmarks demonstrate the clear advantage of our approach in bothefficiency and performance in omni-modal event grounding and captioning.Project website: https://github.com/showlab/PadCaptioner.

3D Visual Grounding (3DVG) is a critical bridge from vision-language perception to robotics, requiring both language understandingand 3D scene reasoning. Traditional supervised models leverage explicit3D geometry but exhibit limited generalization, owing to the scarcity of3D vision-language datasets and the limited reasoning capabilities com-pared to modern vision-language models (VLMs). We propose a gener-alizable 3DVG framework, PanoGrounder, that couples multi-modalpanoramic representation with pretrained 2D VLMs for strong vision-language reasoning. Panoramic renderings, augmented with 3D semanticand geometric features, serve as an intermediate representation between2D and 3D, and offer two major benefits: (i) they can be directly fed toVLMs with minimal adaptation and (ii) they retain long-range object-to-object relations thanks to their 360-degree field of view. We devise athree-stage pipeline that places a compact set of panoramic viewpointsconsidering the scene layout and geometry, grounds a text query on eachpanoramic rendering with a VLM, and fuses per-view predictions into asingle 3D bounding box via lifting. Our approach achieves state-of-the-artresults on ScanRefer and Nr3D, and demonstrates strong generalizationto unseen 3D datasets and text rephrasings.