ECCV 2026 Papers

The ultimate goal of spatial intelligence is to enable embod-ied agents to actively interact with the physical world. However, exist-ing benchmarks adopt a “Spectator View”—relying on single images orpassive video—neglecting the necessity of resolving geometric ambigu-ity through active viewpoint transformation. To address this, we presentWalkerBench, a global-scale interactive benchmark built on real-worldstreet-view and map data spanning 161 cities, with two categories ofhierarchically structured spatial tasks. Our empirical evaluation revealsa critical “representation misalignment”: VLMs’ one-dimensional linearcontext structures are fundamentally incompatible with the topologicalnature of 3D environments, causing inevitable spatial forgetting duringnavigation. To overcome this, we propose Spatial-IDE, which breaks fromlinear dialogue history via two mechanisms: (1) State Externalization,transforming implicit observations into an Explicit Topological Memory(ETM); and (2) Cognitive Decoupling, disentangling visual understand-ing into Goal-Directed Perception and Spatial Reasoning, letting theVLM focus on pure high-level decision-making. Spatial-IDE achieves sub-stantial improvements on WalkerBench and successfully generalizes zero-shot to real-world urban navigation on the Unitree humanoid robot. Theproject is available on https://github.com/lalayang123456- ctrl/WalkerBench.

An embodied kitchen assistant must do more than recog-nize food in isolated frames. It must track ingredient states over timeand integrate visual observations with recipe and nutritional knowledgeto support constraint-aware decision-making. We formalize this capa-bility as Embodied Nutrition Management: perceiving nutrition-relevantevents, maintaining a persistent food state, and using it for knowledge-grounded planning. Existing benchmarks evaluate static food under-standing or embodied cooking actions, but do not measure whether anagent can continuously update and use nutrition-relevant states in dy-namic kitchens. To fill this gap, we introduce NutriBench-Kitchen,a benchmark containing 1,500 manually verified question–answer pairsfrom 160 cooking videos. It covers five task families: Ingredient Entry,Memory Management, Recipe Query, Long-Term Planning, and Short-Term Planning, spanning food-state construction, maintenance, knowl-edge retrieval, and decision-making across different planning horizons.Evaluations of proprietary and open-source large vision-language modelsreveal a substantial gap from human performance, particularly in quan-titative ingredient estimation, long-term state tracking, and reasoningunder interacting constraints. We further introduce Nutri-Vgent, a di-agnostic long-video agent with separate episodic, food-state, and recipememories. Its consistent improvements demonstrate the value of explicitstate representations and structured memory for nutrition management.Together, NutriBench-Kitchen and Nutri-Vgent provide a testbed forstudying persistent state tracking and knowledge-grounded reasoning indynamic kitchens. Code is available at https://github.com/V1ol1n/NutriBench-Kitchen.

The emergence of Multimodal Large Language Models hasenabled a new generative paradigm for multimodal emotion recognition(MER), where emotional interpretations are produced through multi-modal token-level reasoning. In view of multimodal inputs, existing stud-ies commonly observe that reallocating attention from generated tokensto modality tokens during inference consistently improves performance.However, in this work, we reveal a counterintuitive phenomenon: such astrategy only benefits two-modality settings, but degrades performancein three- or more-modality scenarios, where the opposite reallocation di-rection instead leads to improvements. To understand this phenomenon,through comprehensive analysis, we show that with more modalities, thecomplex interactions between modalities make the model hard to paymore attention to tokens with more information, especially in deeperlayers, leading to improper attention allocation and the observed contra-dictory behavior. Based on these insights, we propose a training-freeattention rectification method that leverages structured shallow-layerattention as a prior to regularize entangled final-layer attention dur-ing inference, without introducing additional parameters or modifyingthe backbone model. Extensive experiments on nine datasets in MER-UniBench demonstrate that our method achieves state-of-the-art per-formance, consistently outperforming both reallocation directions acrossdiverse multimodal scenarios. Our code is available at https://github.com/yzl77/ReAR.

Human trajectory forecasting is crucial for safe navigationin crowded environments, requiring models that balance accuracy withcomputational efficiency. Efficiently modeling social interactions is key toperformance in dense crowds. Yet, most recent methods rely on attentionmechanisms, which are effective at capturing complex dependencies, butincur quadratic computational costs that scale poorly with the grow-ing number of neighbors. Recently, Selective State-Space Models haveprovided a linear-time alternative; however, their inherently sequentialdesign is misaligned with the unstructured and dynamic nature of socialinteractions. To address this challenge, we propose Social-Mamba, a fore-casting architecture that reformulates social interactions as structuredsequential processes. At its core is the Cycle Mamba block, a novel mod-ule that enables continuous bidirectional information flow. Social-Mambaorganizes agents on an egocentric grid and introduces social triplet fac-torization, which decomposes interactions into temporal, egocentric, andgoal-centric scans. These are dynamically integrated through a learnablesocial gate and global scan to generate accurate and efficient trajectorypredictions. Extensive experiments on five trajectory forecasting bench-marks show that Social-Mamba achieves state-of-the-art accuracy whileoffering superior parameter efficiency and computational scalability. Fur-thermore, embedding Social-Mamba into a flow-matching framework fur-ther enhances both accuracy and efficiency, establishing it as a flexibleand robust foundation for future trajectory forecasting research. Thecode is publicly available: https://github.com/vita-epfl/Social-Mamba.

Recent co-speech gesture generation methods often overlookcultural differences, limiting their effectiveness in human–agent inter-action. Moreover, culture-conditioned models are rarely evaluated un-der speaker-disjoint splits, so apparent “cultural” behavior may be con-founded with speaker-specific gesturing style. We introduce SICAGE, amodular framework for culture-aware co-speech gesture generation thatconditions motion synthesis models on speaker-independent cultural rep-resentations. SICAGE learns these representations from audio and textby treating each speaker as a separate domain while imposing invari-ance across speakers. This encourages representations to remain culture-discriminative while reducing dependence on speaker identity. The re-sulting cultural embeddings condition a multimodal generator to pro-duce culturally appropriate gestures. We instantiate this idea with twodomain generalization approaches: adversarial learning and Fishr reg-ularization. We further introduce ALaDiT, a real-time diffusion-basedgesture generator designed to efficiently incorporate the learned culturalembeddings. To validate our method, we built TED4C-L, a 106-hourmultimodal dataset of 764 TED speakers from four cultural groups. Ex-periments show that SICAGE improves motion realism, diversity, beatsynchronization, semantic relevance, and cultural consistency.

Modern Visual Place Recognition (VPR) methods excel onstandard benchmarks yet remain brittle in feature-poor environments.By treating each query image in isolation, they discard the sequentialcontext in any real trajectory. We formalize a task that exploits thiscontext: given a query sequence, localize the final image against an un-ordered reference database—which, unlike sequence-to-sequence meth-ods, requires no sequential structure in the database. We propose TRAIL(TRajectory-Aware Image Localization), a principled framework basedon Conditional Random Fields (CRF) that combines learned functionsfor visual similarity and for camera-motion consistency, refining a dis-tribution over candidate references as each query arrives. A lightweightpost-processing layer atop any pre-trained VPR backbone, TRAIL im-proves a state-of-the-art baseline by up to 8.3 percentage points on ourprimary benchmark, transfers to unseen datasets without retraining, anddelivers its largest gains where visual cues are scarce.

Real-world image restoration (IR) remains challenging dueto complex and coupled degradations. While recent agentic IR frame-works leverage Large Language Models for flexible tool planning, theyface two critical limitations. First, from a search scheme perspective,excessive reliance on greedy strategies fails to balance exploration andexploitation. Second, existing agentic systems underutilize information,exhibiting episodic amnesia. To address these challenges, we proposeSelf-Evolving Agentic Image Restoration (SEAR), which formu-lates restoration as a sequential decision-making problem. Inspired by thedual-process theory, SEAR comprises an Intuitive Executor and a Delib-erate Planner, respectively following the fast-thinking System 1 and slow-thinking System 2 principles. The Deliberate Planner employs Pruning-Aware Monte Carlo Tree Search for long-horizon reasoning, utilizing ahybrid no-reference reward and a Multimodal Large Language Model(MLLM)-based tournament to prevent metric exploitation. Complemen-tarily, the Intuitive Executor leverages a self-evolving episodic memoryindexed by degradation-aware state fingerprints. This mechanism dis-tills expensive search trajectories into adaptive expertise, overcomingepisodic amnesia while progressively amortizing cold-start explorationcosts through memory reuse. Extensive experiments on synthetic andreal-world benchmarks demonstrate its strong perceptual and quantita-tive performance.

Recent advancements in 3D Gaussian Splatting (3DGS) haveachieved significant progress in novel view synthesis (NVS), offering com-putationally efficient and photorealistic rendering, even in the few-shotsetting where input training views are sparse. However, several repre-sentative 3DGS-based methods for few-shot NVS face a critical issue ofsparsity discrepancy, i.e., Gaussian initialization leverages dense fusedstereo points from extended views, creating an inconsistency with sparsetraining views. To resolve this issue, we propose ReInGS, a novel methodfeaturing a two-stage coarse-to-fine 3D Gaussian re-initialization pro-cess designed explicitly for few-shot NVS. In the coarse-level initializa-tion stage, an expansion-based hybrid point sampling (EHPS) strategyis deployed to generate dense points and effectively capture the scene’sglobal geometric structure. In the fine-level re-initialization stage, a view-dominant details-aware point sampling (VDPS) strategy, which com-prises two sequential sub-strategies on intra-view point sampling andcross-view point sampling, is adopted to reconstruct the scene’s localgeometric details with high fidelity. Comprehensive experiments on twobenchmarking datasets demonstrate that ReInGS achieves state-of-the-art performance across 2-view, 3-view, and 4-view settings, especiallydelivering PSNR gains of up to 2.92 dB, 1.92 dB, and 1.38 dB, respec-tively, while synthesizing high-quality novel views without data leakagefrom extended views.

We study sinusoidal recurrence as an iterative mechanism forharmonic spectral enrichment in implicit neural representations (INRs).Our analysis reveals that sinusoidal activations induce a harmonic linespectrum, providing a spectral account of how recurrent unrolling en-riches the effective spectral support. We realize this principle with ashared sinusoidal block that iteratively refines the latent representation.We empirically validate the resulting spectral behavior against feed-forward INRs, non-sinusoidal recurrent variants, and equilibrium-stylesinusoidal models. Complementing this analysis, we evaluate the pro-posed architecture across image and 3D representation tasks. On RGBimage benchmarks, our method achieves higher fidelity than feed-forwardbaselines with fewer parameters and fewer optimization steps, and it fur-ther transfers favorably to super-resolution, NeRF, and SDF tasks.

Online vectorized map construction provides autonomous ve-hicles with essential, real-time semantic and geometric scene understand-ing. While leveraging prior maps presents an opportunity to improveaccuracy, they are often biased due to outdated information or tem-porary road changes. Existing methods, which typically treat priors asfully reliable or only consider limited noise patterns, suffer severe perfor-mance degradation when these priors are imperfect. To address this, wepropose PriorMaskMap, a robust framework for prior-guided vectorizedmap construction. Its first innovation is a Prior-Aware Confidence Esti-mator (PACE) that evaluates the input rasterized map against currentvisual observations. It assigns confidence scores to regions classified asreliable, erroneous, or blank, creating a foundational mask that guidestwo subsequent branches. In one branch, the Dual-Prior Query Genera-tor (DPQG) selectively filters and embeds both rasterized and vectorizedpriors to initialize informative queries from trustworthy map structures.In the other, the Confidence-Guided Prior Fusion (CGPF) module spa-tially propagates and refines the mask, using the refined version to gov-ern the fusion of BEV features with the prior map, thereby providingrich contextual representations for the map decoder. Evaluation on thenuScenes dataset shows that PriorMaskMap achieves superior mappingprecision over existing approaches and exhibits marked robustness whenprovided with priors containing apparent errors. Our code is released athttps://github.com/healenrens/PriorMaskMap.

Object detection in remote sensing images (RSIs) is chal-lenged by the coexistence of geometric and spatial complexity: targetsmay appear with diverse aspect ratios, while spanning a wide range ofobject sizes under varied contexts. Existing RSI backbones address thetwo challenges separately, either by adopting anisotropic strip kernelsto model slender targets or by using isotropic large kernels to capturebroader context. However, such isolated treatments lead to complemen-tary drawbacks: the strip-only design can disrupt spatial coherence forregular-shaped objects and weaken tiny details, whereas isotropic largekernels often introduce severe background noise and geometric mismatchfor slender structures. In this paper, we extend PKINet, and presenta powerful and efficient backbone that jointly handles both challengeswithin a unified paradigm named Poly Kernel Inception Network v2(PKINet-v2). PKINet-v2 synergizes anisotropic axial-strip convolutionswith isotropic square kernels and builds a multi-scope receptive field, pre-serving fine-grained local textures while progressively aggregating long-range context across scales. To enable efficient deployment, we furtherintroduce a Heterogeneous Kernel Re-parameterization (HKR) Strategythat fuses all heterogeneous branches into a single depth-wise convolutionfor inference, eliminating fragmented kernel launches without accuracyloss. Extensive experiments on four widely-used benchmarks, includingDOTA-v1.0, DOTA-v1.5, HRSC2016, and DIOR-R, demonstrate thatPKINet-v2 achieves state-of-the-art accuracy while delivering a 3.9×FPS acceleration compared to PKINet-v1, surpassing previous remotesensing backbones in both effectiveness and efficiency.

In animation production, paint-bucket colourisation for hand-drawn animation is a labour-intensive procedure that assigns each en-closed region in line sketches a colour from reference design sheets. Re-cent automatic paint-bucket colourisation pipelines mirror this workflowvia region correspondence, but correspondences can be brittle when re-gions are ambiguous fragments without proper context. In this paper,we propose Palette Context Assisted (PeCA), a new training-free, plug-and-play framework for animation video colourisation that aims to closethis gap at test-time via reasoning over spatial and temporal contexts.Extensive experiments on existing benchmarks and a newly introducedlong-video test case show consistent performance boosts.

Robust 3D object detection under adverse weather remains acritical hurdle for autonomous driving. Despite progress with LiDAR–4Dradar fusion, most methods are constrained by a closed-world assump-tion, implicitly requiring training and test weather to align in both typeand severity. This premise fails in practice: the open-ended nature ofweather, and even variations within a single type like rain, cause dramat-ically different LiDAR degradation patterns, leading to significant per-formance drops in unseen conditions. To address this, we present Dual-Critic Guided Diffusion Alignment (DCDA), a weather-agnostic frame-work that learns to recover degraded LiDAR features toward a cleanmanifold. Rather than modeling specific weather types, DCDA employsa 4D radar-conditioned diffusion process to progressively refine features,guided by two complementary critics. (i) A detection-guided critic, an-chored by a pre-trained clean-weather model, ensures that the refinedfeatures retain object-level discriminability and localization accuracy. (ii)A weather adversarial critic enforces holistic distributional consistencywith clean-weather representations. By aligning features through seman-tic and distributional constraints rather than explicit weather model-ing, DCDA generalizes effectively to unseen weather types and severitieswithout requiring paired data or weather labels. We further introduce astructured open-weather benchmark with held-out type–severity combi-nations and extensive experiments verify DCDA’s advantages. Code isavailable at https://github.com/Mangonn/DCDA.

Building one-shot 3D animatable head avatars is an impor-tant yet challenging problem. Existing methods generally collapse underlarge camera pose variations, compromising the realism of 3D avatars.In this work, we propose a new framework to tackle the novel settingof one-shot 3D full-head animatable avatar reconstruction in a singleforward pass via inpainted UV-space Gaussian modeling, enabling 360◦rendering views and real-time animation. To facilitate efficient animationcontrol, we model 3D head avatars with Gaussian primitives embeddedon the surface of a parametric face model within the UV space, andproject the input image features to the UV space, resulting in incom-plete local UV feature maps. To inpaint the missing regions, we obtainknowledge of full-head geometry and textures from rich 3D full-head pri-ors within a pretrained 3D generative adversarial network (GAN) forglobal full-head feature extraction and multi-view supervision. Specifi-cally, to enhance the fidelity of 3D reconstruction during inpainting, wetake advantage of the symmetric nature of the UV space and humanfaces to fuse incomplete yet detailed local UV feature maps with theextracted global full-head textures, resulting in inpainted UV Gaussianattribute maps for avatar modeling. Extensive experiments demonstratethat our method is the first to achieve high-quality 3D full-head animat-able avatar modeling, significantly improving side and back views whileoutperforming state-of-the-art animation approaches, thereby improvingthe realism of 3D animatable avatars. The project page is available athttps://shaelynz.github.io/fhavatar/.

Robotic Vision-Language-Action (VLA) models generalizewell for open-ended manipulation, but their perception is fragile undersensing-stage degradations such as extreme low light, motion blur, andblack clipping. We present E-VLA, an event-augmented VLA frameworkthat improves manipulation robustness when conventional frame-basedvision becomes unreliable. Instead of reconstructing images from events,E-VLA directly leverages motion and structural cues in event streamsto preserve semantic perception and perception-action consistency un-der adverse conditions. We build an open-source teleoperation platformwith a DAVIS346 event camera and collect a real-world synchronizedRGB-event-action manipulation dataset across diverse tasks and illumi-nations. We also propose lightweight, pretrained-compatible event in-tegration strategies and study event windowing for stable deployment.Experiments show that even a simple parameter-free fusion, i.e., over-laying accumulated event maps onto RGB images, could substantiallyimprove robustness in dark and heavy-blur scenes: on Pick-Place at 20lux, success increases from 0% (image-only) to 60% with overlay fusionand to 90% with our event adapter; under severe motion blur (1000 ms-exposure proxy), Pick-Place improves from 0% to 20-25%, and Sortingfrom 5% to 32.5%. Overall, E-VLA provides systematic evidence thatevent-driven perception can be effectively integrated into VLA mod-els, pointing toward robust embodied intelligence beyond conventionalframe-based imaging. Code and dataset will be available at E-VLA.

Training-free guidance (TFG) steers a pretrained diffusionmodel toward a desired attribute at inference. To be effective, this guid-ance must be applied from the earliest, high-noise steps of sampling.Because its objective (a classifier or energy) is defined on clean images,ϵ- and v-prediction models must first estimate the clean image x̂ fromthe noisy state at each step, and the accuracy of that estimate deter-mines how easily guidance drifts off the data manifold. x-prediction,a recent alternative, outputs the clean image directly, removing thissource of error even at high noise. This is our motivation. We providea theoretical analysis of how each prediction target shapes this accu-racy, and introduce guided-class FID (Child FID), a metric that exposesthe manifold damage standard evaluation misses. Experiments on anew fine-grained bird benchmark and on style transfer confirm that x-prediction keeps guided samples on the manifold most reliably, making itthe strongest foundation for training-free guidance. Code is available athttps://github.com/ManLuML/on-manifold-tfg.

Current image-to-video models struggle to maintain appear-ance consistency such as the back of a garment. Lacking complete obser-vations, they hallucinate missing views, leading to visual inconsistencies,limiting deployment in applications like e-commerce, where strict visualfidelity is required. In this paper, we propose the Multiview EnhancedImage-to-Video Generation Model (MVI2V), which introduces additionalmulti-view images of a person or garment as reference inputs. Concretely,MVI2V adds a structurally identical forward stream for reference images,upgrading single- or dual-stream baselines into dual- or triple-stream ar-chitectures, and performs cross-stream fusion via self-attention to enablebidirectional information flow among token types. To better exploit thereferences, we adopt an inpainting sub-task that randomly masks theperson region in the conditioning image, forcing the model to rely moreon the reference views. We further design a data curation pipeline thatselects videos with large viewpoint changes and extracts diverse multi-view reference frames. Extensive experiments on single-stream Wan2.1and our in-house dual-stream model demonstrate that MVI2V effectivelyleverages multi-view references while preserving the base models’ single-image to video capability, validating our method’s generalizability todifferent architectures.

Neural video codecs have surpassed classical codecs in codingefficiency but remain impractical for deployment due to cross-platformincompatibility and high computational cost. Existing quantization-basedsolutions fail to produce deterministic results across diverse hardwareplatforms, leading to catastrophic decoding failures. We introduce MLVC,a hardware-robust neural video codec designed for practical cross-platforminference. The key idea is to explicitly transmit scale parameters throughthe hyperprior, which guarantees entropy coding consistency across de-vices without requiring bit-exact arithmetic. While this increases bitrateoverhead, we recover most of the coding efficiency through architecturalimprovements (gated memory, ReGLU activation), a long-term refer-ence recovery mechanism, and domain-specific perceptual training. Onthe VCD video conferencing benchmark, MLVC achieves >70% BD-rate (MOS) improvement over hardware HEVC, the strongest deployablebaseline, while reaching subjective quality competitive with DCVC-RT,which cannot operate across diverse platforms. Both the encoder anddecoder run at 100 FPS on average on commodity NPUs from Apple,Intel, and Qualcomm. MLVC is the first neural video codec to com-bine competitive compression performance, real-time speed, and cross-platform robustness across diverse consumer devices, making it suitablefor widespread deployment. Code will be released.

We present MagnetGS-Mesh, an integrated framework thatdirectly recovers high-quality, object-wise surface meshes from 3D Gaus-sian Splatting (3DGS) while preserving photorealistic novel-view syn-thesis. Prior 3DGS-to-mesh methods generate meshes by considering allGaussians without distinguishing objects from the background, whichinevitably yields excessive and unnecessary vertices in irrelevant regions.To address this efficiency bottleneck, we adopt an object-wise reconstruc-tion approach. However, decomposing 3DGS representations into indi-vidual objects introduces distinct geometric challenges: numerous out-lier Gaussians appear far from true surfaces, and the remaining surfaceGaussians often exhibit highly non-uniform density. Since robust meshextraction fundamentally requires a uniform distribution of Gaussiansto prevent surface holes and artifacts, these irregularities severely de-grade geometric fidelity. To resolve these issues, we propose a geometry-aligned two-stage training procedure. In Stage 1, we employ EnhancedOccupancy Learning to produce crisp boundaries, ensuring accurate 3Dsegmentation. In Stage 2, to address the uneven distribution, we in-troduce the Adaptive Local Normal Consistency (ALNC) Loss.This mechanism acts as a “magnet,” strategically relocating outlier Gaus-sians to sparse surface regions to ensure uniform coverage. By adap-tively aligning these relocated Gaussians with local surface normals,ALNC effectively fills geometric gaps while maintaining structural co-herence. Experimentally, MagnetGS-Mesh achieves competitive or supe-rior mesh quality compared to state-of-the-art methods while simultane-ously reducing model size by substantial margins. The code is availableat https://github.com/MinsuPark0752/MagnetGS-Mesh.