ECCV 2026 论文列表

Visual signals play a crucial role in policy learning by en-abling models to capture object motion and interaction dynamics. Just ashumans reason about actions using both past experience and anticipatedoutcomes, effective policies should integrate past interactions with futurepredictions. However, existing visuomotor policies typically model eitherhistorical context or future dynamics in isolation, lacking a unified tem-poral representation of interaction dynamics. In this work, we introduceChronoFlow, a temporally unified representation that captures past,current, and future interaction dynamics through sparse 3D keypointsof both objects and the gripper. Based on this representation, we proposeChronoFlow-Policy, a diffusion-based visuomotor policy that jointlylearns ChronoFlow and action sequences through a co-training objective.Experiments on 14 simulated tasks and 5 real-world manipulation tasksdemonstrate that ChronoFlow-Policy consistently outperforms strongdiffusion-policy baselines and improves robustness in long-horizon andnon-Markovian manipulation scenarios. Code and models will be releasedat https://github.com/The-kamisato-Sii/ChronoFlow-Policy.

Vision–Language Models (VLMs) excel at reasoning in lin-guistic space but struggle with perceptual understanding that requiresdense visual perception, e.g., spatial reasoning and geometric aware-ness. This limitation stems from the fact that current VLMs have lim-ited mechanisms to capture dense visual information across spatial di-mensions. We introduce Chain-of-Visual-Thought (CoVT), a frameworkthat enables VLMs to reason not only with discrete text tokens but alsothrough continuous visual tokens—compact latent representations thatencode rich perceptual cues. With a small budget of roughly 20 tokens,CoVT distills knowledge from lightweight vision experts that capturecomplementary properties such as 2D appearance, 3D geometry, spatiallayout, and edge structure. During training, the VLM with CoVT au-toregressively predicts these visual tokens to reconstruct dense supervi-sion signals (e.g., depth, segmentation, edges, and DINO features). At in-ference, the model reasons directly in the continuous visual latent space,preserving efficiency while optionally decoding dense predictions for in-terpretability. Evaluated across more than ten diverse benchmarks, in-cluding CV-Bench, MME-RealWorld, MMVP, RealWorldQA, MMStar,WorldMedQA, and HRBench, integrating CoVT into strong VLMs suchas Qwen2.5-VL and LLaVA consistently improves performance by 3% to16% and demonstrates that compact continuous visual thinking enablesmore precise, grounded, and interpretable multimodal intelligence.

High-quality 2D assets have become increasingly abundantand easy to edit, providing a rich foundation for creative content acrossart, design, and virtual environments. However, while diffusion-based 3Dgeneration has achieved remarkable progress in single-object synthesis,leveraging such 2D assets for controllable 3D composition remains a chal-lenging problem. To address this, we introduce CAST3D, a training-freeframework that enables Customized Composition in 3D: transformingarbitrary 2D assets into a coherent 3D object or scene under textualguidance. CAST3D consists of two stages: 3D Layout Hinting and Com-positional Generation. To maintain structural consistency and eliminateartifacts, we further design stochastic trajectory manipulation (STM)for structure-preserving modification and a connectivity-based pruningstrategy for clean geometry integration. Extensive experiments demon-strate that CAST3D produces semantically consistent and visually faith-ful 3D compositions, bridging 2D asset creation and 3D world synthesis.

Incremental 3D object detection requires a detector to learnnovel object classes while remembering previously learned ones oversequentially arriving data. Previous methods, primarily based on pseudo-labeling, perform reasonably in short-incremental stages but still suf-fer from severe model forgetting when dealing with long-incrementalsequences. We investigate this failure and reveal a detrimental self-reinforcing cycle: data distribution shift of novel classes causes modelforgetting on old classes, which further produces accumulated error inpseudo-labeling that exacerbates model degradation. To address thisissue, we draw inspiration from the human learning process and proposethe Learning-Dynamics-driven Memory and Review (LDMR) framework.LDMR monitors per-class detection quality at periodic training check-points and uses these learning-dynamics signals to drive two innova-tive mechanisms, namely (i) human-like intra-stage review that divideseach incremental stage into multiple sub-stages’ training and concen-trates on remembering the most-forgotten objects, and (ii) scene-awarecross-stage memory evolution that evolves a memory bank to transferknowledge between two consecutive stages by jointly considering scenelearnability and diversity. Extensive experiments across multiple long-incremental protocols on indoor benchmarks SUN RGB-D and Scan-NetV2 show that LDMR substantially mitigates the model forgettingand outperforms all baselines by a clear margin. Code is available athttps://github.com/qianpeisheng/LDMR.

WiFi Channel State Information (CSI) enablesprivacy-preserving human pose sensing in camera-denied environments,but existing WiFi-based pose estimators often fail under environmentshifts and rely on costly camera-based annotation pipelines that limitscale. We propose WiFi-JEPA, a self-supervised framework that learnsCSI-native representations by predicting masked latent embeddings in-stead of reconstructing raw CSI signals that may contain hardware-specific artifacts. WiFi-JEPA makes three contributions: (i) CSI-specifictokenization and link masking tailored to the CSI tensor over channel,time, and link (C, T, L); masking entire Tx–Rx antenna links forces themodel to predict one spatial link view from others, capturing cross-linkcorrelations informative of 3D spatial structure. (ii) A ray-tracing CSIsimulation pipeline that generates diverse unlabeled CSI from random-ized geometric primitives, providing scalable pre-training data withoutpose annotations. (iii) State-of-the-art results on Person-in-WiFi-3D:WiFi-JEPA outperforms prior WiFi-CSI baselines on both single- andmulti-person 3D pose estimation under the same evaluation protocol.We also show that simulated CSI provides complementary pre-trainingsignal to real CSI, and that four vision-native SSL objectives degrade per-formance below training from scratch, whereas WiFi-JEPA consistentlyimproves downstream pose estimation.

We present an RGB-only method for 6D object pose esti-mation that leverages monocular depth cues from a single image. Al-though monocular depth estimation has advanced substantially, its pre-dictions remain scale-ambiguous and locally unreliable, limiting theiruse in metric pose refinement. We address this gap by closing the loopbetween depth correction and pose refinement: monocular depth is notonly a regularizer but is iteratively calibrated and filtered using pose-induced geometric consistency, enabling stable metric pose updates fromRGB alone. We propose a dynamic depth outlier removal module basedon metric consistency and infer object pose from dense 2D correspon-dences. Both components are embedded into a recurrent optimizationloop, enabling iterative depth correction and pose refinement. Experi-ments on seven BOP datasets demonstrate state-of-the-art performanceamong RGB-only methods, without requiring real depth input.

Blind Face Restoration (BFR) encounters inherent challengesin exploring its large solution space, leading to missing details and identityambiguity in deterministic outputs. While diffusion models offer powerfulgenerative priors, their potential is often constrained by deterministic sam-pling processes that lack exploration in the vast solution space. To tacklethis, we propose Likelihood-Regularized Policy Optimization (LRPO),the first online reinforcement learning (RL) framework that uniquelyleverages the inherent stochasticity of diffusion models for the BFR task.Instead of relying on a single deterministic mapping, LRPO systematicallyexplores multiple diverse restoration candidates per input. By evaluatingthese trajectories, our framework steers the denoising policy towardsoptimal solutions, significantly increasing the likelihood of high-qualityoutputs. To effectively facilitate this exploration mechanism, LRPO incor-porates three key innovations: 1) a composite reward function balancinghuman preference and fidelity, 2) ground-truth-guided likelihood regular-ization, and 3) a noise-level advantage assignment. Extensive experimentsdemonstrate that LRPO significantly improves face restoration qualityover baselines and achieves state-of-the-art performance. Source code isavailable at the GitHub repository BFR_RL.

We introduce a novel upper bound on the local Lipschitz con-stant of the dot-product self-attention block showing its dependence onthe attention map distributions. The proposed bound is not only tighterthan the prior art, but for the first time, reveals how the distributionof attention probabilities shapes the local Lipschitz constant of the self-attention block. The theoretical basis of the proposed upper bound lies inthe refined closed-form upper bounds on singular values of the Jacobianof softmax function. Leveraging these theoretical insights, we introduceJaSMin (Jacobian Softmax norm Minimization), a lightweight regular-izer that directly controls the local Lipschitz constant of each block and,consequently, the entire model. Additionally, we discuss how the natureof the attention map distribution contributes to the gradient dynamicsand, consequently, transformer training stability.

We present AHOY, a method for reconstructing complete,animatable 3D Gaussian avatars from in-the-wild monocular video de-spite heavy occlusion. Existing methods assume unoccluded input: afully visible subject, often in a canonical pose. This excludes the vastmajority of real-world footage where people are routinely occluded by fur-niture, objects, or other people. Reconstructing from such footage posesfundamental challenges: large body regions may never be observed, andmulti-view supervision per pose is unavailable. We address these challengeswith four contributions: (i) a hallucination-as-supervision pipeline thatuses identity-finetuned diffusion models to generate dense supervisionfor previously unobserved body regions; (ii) a two-stage canonical-to-pose-dependent architecture that bootstraps from sparse observations tofull pose-dependent Gaussian maps; (iii) a map-pose/LBS-pose decou-pling that absorbs multi-view inconsistencies from the generated data;(iv) a head/body split supervision strategy that preserves facial iden-tity. We evaluate on YouTube videos and on multi-view capture datawith significant occlusion and demonstrate state-of-the-art reconstruc-tion quality. We also demonstrate that the resulting avatars are robustenough to be animated with novel poses and composited into 3DGSscenes captured using cell-phone video. Our project page is available athttps://miraymen.github.io/ahoy/.

for previously unobserved body regions; (ii) a two-stage canonical-to-pose-dependent architecture that bootstraps from sparse observations tofull pose-dependent Gaussian maps; (iii) a map-pose/LBS-pose decou-pling that absorbs multi-view inconsistencies from the generated data;(iv) a head/body split supervision strategy that preserves facial iden-tity. We evaluate on YouTube videos and on multi-view capture datawith significant occlusion and demonstrate state-of-the-art reconstruc-tion quality. We also demonstrate that the resulting avatars are robustenough to be animated with novel poses and composited into 3DGSscenes captured using cell-phone video. Our project page is available athttps://miraymen.github.io/ahoy/.

Recent 3D vision-language models (3D VLMs) construct ge-ometry aware tokens by projecting 2D visual features into world coordi-nates, enabling spatial reasoning for tasks such as 3D question answering.However, this design generates thousands of tokens per scene, resultingin substantial computational and memory overhead. While token com-pression has been extensively studied in 2D VLMs, existing approachesrely on semantic relevance or attention-based selection that overlook thestructured spatial nature of 3D tokens. Moreover, redundancy in 3D rep-resentations cannot be resolved by spatial proximity alone, as object-leveltoken imbalance persists even after spatial aggregation. To address this,we propose 3DZip, a three-stage token compression framework that firstapplies coarse voxelization to remove point-level redundancy, then selectsanchor tokens based on feature-space diversity via a Determinantal PointProcess, and finally merges remaining tokens under spatial constraintsto preserve geometric coherence. Experiments on three 3D question an-swering benchmarks demonstrate that 3DZip consistently outperformsexisting compression methods, retaining 94.7% of the original perfor-mance with only 128 tokens, achieving a 1.92× faster inference speed.

Photorealistic and animatable hand avatars are essential forapplications such as AR/VR, gaming, and telepresence. Recent 3D Gaus-sian Splatting (3DGS) based avatar methods enable real-time renderingof articulated humans, but modeling hands remains challenging due totheir compact structure, frequent self-occlusions, and complex finger in-teractions. Existing approaches primarily rely on pose-driven transfor-mations while representing Gaussians in Euclidean space, lacking an ex-plicit structural association with the underlying skeleton, which makespreserving fine-grained hand structures under complex articulations diffi-cult. In this work, we introduce the Structural Coordinate Space (SCS),a skeleton-relative representation that assigns each Gaussian primitivean explicit structural coordinate with respect to the articulated handskeleton. SCS is constructed using a hybrid static–virtual bone basis to-gether with a distance–rotation structural descriptor that encodes thegeometric relationship between Gaussians and bones. Based on SCS, weenforce both intra-pose and cross-pose structural consistency by com-bining per-Gaussian residual embeddings for local appearance modelingwith structural correspondence across poses. Experiments demonstratethat our approach significantly improves structural consistency and pre-serves fine geometric details under challenging hand articulations com-pared with existing 3DGS-based avatar methods.

Sign language translation (SLT) converts continuous signvideos into spoken language text. Gloss-free approaches leverage pre-trained visual encoders and language models but rely on implicit cross-modal alignment from translation supervision alone. We present VTaMo,a framework that introduces explicit multi-granularity alignment at threelevels: (1) local alignment via entropy-regularized optimal transport witha learnable null token for fine-grained frame-to-token correspondences;(2) global alignment via a learnable orthogonal transformation that cali-brates embedding space geometry through Earth Mover’s Distance; and(3) position-aligned contrastive learning for discriminative token-levelrepresentations. Experiments on Phoenix-2014T, CSL-Daily, How2Sign,and OpenASL demonstrate consistent state-of-the-art performance, withablations confirming the complementary contributions of each compo-nent. Code is available at https://github.com/junyi2005/vtamo.

Current text-to-video models can make individual frames look convincing while still getting simple interactions wrong: objects move before contact, an intended action is skipped, a placed object keeps drifting, or a support relation breaks. Our starting point is that standard frame-first denoising updates every latent region at every step, even when the prompt implies that only a local interaction should be active. We introduce Event-Driven Video Generation (EVD), a small DiT-compatible intervention that gives the sampler an explicit event signal. A lightweight head predicts token-level event activity; training losses tie that activity to latent state change; and event-gated sampling, with hysteresis and an early-step schedule, applies the update field mainly where an interaction is forming. On EVD-Bench, EVD improves human preference and VBench dynamics for state persistence, spatial accuracy, support relations, and contact stability, while keeping appearance quality comparable to the base model. The results suggest that a modest amount of event structure can correct several interaction failures that otherwise remain hidden behind good frame-level appearance. Project webpage: https://evd-project-website.pages.dev

Controlled video generation has seen drastic improvementsin recent years. However, editing actions and dynamic events, or insert-ing contents that should affect the behaviors of other objects in real-world videos, remains a major challenge. Existing trained models strug-gle with complex edits, likely due to the difficulty of collecting relevanttraining data. Similarly, existing training-free methods are inherently re-stricted to structure- and motion-preserving edits and do not supportmodification of motion or interactions. Here, we introduce DynaEdit,a training-free editing method that unlocks versatile video editing ca-pabilities with pretrained text-to-video flow models. Our method relieson the recently introduced inversion-free approach, which does not in-tervene in the model internals, and is thus model-agnostic. We showthat naively attempting to adapt this approach to general unconstrainedediting results in severe low-frequency misalignment and high-frequencyjitter. We explain the sources of these phenomena and introduce novelmechanisms for overcoming them. Through extensive experiments, weshow that DynaEdit achieves state-of-the-art results on complex text-based video editing tasks, including modifying actions, inserting objectsthat interact with the scene, and introducing global effects (see website).

Global geolocation, the task of predicting precise coordi-nates from street-view imagery, is inherently plagued by visual ambi-guity. Resolving such ambiguity necessitates a transition from static,one-shot predictions to an interactive geolocation paradigm drivenby multi-turn deductive reasoning. However, most of existing geoloca-tion models and general-purpose MLLMs fail to support this dynamicprocess, mainly due to the lack of interaction capabilities and the ge-ographic knowledge gap. Driven by the imperative to actualize thisinteractive paradigm, we introduce MG-Geo, the first large-scale mul-timodal geolocation dataset explicitly structured for spatial reasoning.Comprising 4.87M geo-tagged Meta entries, 70K image-grounded Cluesamples, and 73K multi-turn Dialog samples across 210 countries andterritories, MG-Geo separates large-scale geographic alignment fromreasoning-oriented supervision. Experiments demonstrate that GaGAachieves SOTA performance across several benchmarks. Notably, on theGWS15k dataset, it surpasses the strong Hybrid Model by 4.57% and2.92% at the country and city levels, respectively, while securing thehighest city-level accuracy (7.46%) on OSV-5M-test. More importantly,we formalize the “Similarity Trap”—a phenomenon where distributivevisual features mislead static models—and demonstrate that GaGA ef-fectively navigates this challenge through a Tiered Interaction Protocol.By dynamically integrating user-provided geographic anchors, GaGAachieves significant localization improvements. Our dataset is accessiblevia: https://huggingface.co/datasets/kendouvg/MG-Geo.

Open-vocabulary 3D scene graph methods typically operatein two stages: first reconstruct, then enrich with vision-language mod-els, leaving the graph unqueryable during exploration. We argue thatthis sequential coupling is unnecessary and propose an asynchronous ar-chitecture in which lightweight online mapping runs concurrently withheavyweight semantic refinement. A probabilistic voxel-based backbonemaintains stable object identities incrementally, while background VLMagents progressively enrich the graph. This framework resolves duplicateobject tracks through semantic loop closure, attaches fine-grained visualattributes and derives spatial relations between objects. A multi-targetframe scheduler amortizes VLM cost by selecting a small set of infor-mative frames that jointly cover multiple targets. The resulting scenegraph is queryable during exploration and grows in semantic richnessover time. Our method matches or outperforms existing open-vocabulary3D scene graph methods on semantic segmentation (ScanNet, Replica)and surpasses the prior state-of-the-art across three visual groundingbenchmarks (Sr3D+, Nr3D, ScanRefer) by 15.3 to 18.8 A@0.25. Projectpage: https://denizbickici.github.io/thinkgraphs/

We present Soft Tail-dropping Adaptive Tokenizer (STAT),a discrete tokenizer that learns adaptive visual representations. STATadjusts the number of tokens allocated to each image according to itsperceptual complexity. Specifically, it encodes an image into discrete to-kens together with token-wise keep probabilities indicating whether eachtoken is necessary for faithful reconstruction or can be safely dropped.Through this learned adaptivity, STAT achieves state-of-the-art recon-struction quality while using fewer tokens on average. When integratedwith vanilla causal autoregressive (AR) modeling, STAT enables a content-aware generative model with adaptive-length sampling. The model achievescompetitive or superior visual generation quality compared with othergenerative model families while exhibiting favorable scaling behavior thathas been elusive in prior vanilla AR visual generation attempts.

Synthesizing large-scale, explorable, and geometrically accurate 3D urban scenes is a challenging yet valuable task for immersive and embodied applications. The challenge lies in the lack of large-scale and high-quality real-world 3D scans for training generalizable generative models. In this paper, we take an alternative route to create largescale 3D scenes by leveraging readily available satellite imagery for realistic coarse geometry and open-domain diffusion models for high-quality close-up appearance synthesis. We propose Skyfall-GS, a novel hybrid framework that synthesizes immersive city-block scale 3D urban scenes by combining satellite reconstruction with diffusion refinement, eliminating the need for costly 3D annotations, and also featuring real-time, immersive 3D exploration. We tailor a curriculum-driven iterative refinement strategy to progressively enhance geometric completeness and photorealistic texture. Extensive experiments demonstrate that SkyfallGS provides improved cross-view consistent geometry and more realistic textures compared to state-of-the-art approaches. Project page: https://skyfall-gs.jayinnn.dev/