ECCV 2026 Papers

Scene understanding requires simultaneous prediction aboutgeometry, appearance, and semantics. However, existing task-specific an-notations are fragmented across incompatible, domain-specific datasets.Current unified systems circumvent this by restricting training to fullyco-annotated data, or by incurring the large computational cost of pseudo-labeling. To mitigate this, we introduce UniD, a unified video model thatjointly predicts eight dense scene properties—depth, surface normals, se-mantic segmentation, boundaries, human parts, albedo, shading, and ma-terials—all learned from disjoint, domain-specific datasets. We propose asimple yet effective distillation step in which per-task experts supervisea unified backbone through lightweight task projectors, eliminating theneed for annotation overlap or pseudo-labeling. Our key insight is thatthe strong visual priors of a pretrained diffusion model are sufficientto bridge the domain gaps introduced by disjoint training sources, en-abling robust generalization to scene-task combinations never seen dur-ing training. UniD achieves competitive performance against per-taskspecialists and multi-task baselines, with strong generalization to out-of-distribution scenarios and enhanced temporal and cross-task consistency.Code and video results are available at https://unid-video.github.io.

Ensuring safe content generation in text-to-image diffusionmodels remains a critical challenge. Existing safety mechanisms focuson model editing or trajectory steering, yet the design of the blacklist,the set of unsafe concepts these methods steer away from, is an equallypivotal but largely under-explored dimension. Through controlled exper-iments on blacklist-guided concept erasure, we uncover three findingsthat overturn common assumptions: longer keyword phrases do not nec-essarily improve safety, enlarging the blacklist yields diminishing returns,and mixing in irrelevant concepts degrades performance. Together theypoint to a single principle: blacklist effectiveness is governed by how wellits keywords cover unsafe regions of the embedding space, not by individ-ual keyword quality or list size. Building on this principle, we introduceMAPS, a framework for automated blacklist optimization. MAPS usesreinforcement learning to discover a prompt token that, when insertedinto an LLM prompt, elicits semantically diverse safety keywords, andthen assembles category-specific blacklists for distinct unsafe concepts.At inference, it selects the most relevant blacklist for each prompt viaembedding similarity, maximizing coverage of the threat while avoidinginterference from unrelated concepts. MAPS acts as a drop-in enhance-ment for existing trajectory-steering methods, reducing attack successrates by 25–52% (relative) across five red-teaming benchmarks on thenudity concept while preserving generation quality, all without modelediting or modifications to the steering procedure. Our code is availableat https://github.com/juyangbai/MAPS.

The rapid expansion of low Earth orbit satellites such asStarlink is increasingly contaminating astronomical surveys. In practice,contaminated images are often identified through inspection. However,modern surveys generate terabytes of data each night, making manualscreening infeasible and necessitating reliable automated methods forsatellite trail removal. Unfortunately, existing general-domain line de-tection methods fail to generalize to astronomical images due to do-main mismatch, which are mostly grayscale with sparse bright stars andhave a low signal-to-noise ratio. Moreover, training new models fromscratch is impractical due to the lack of large-scale annotated astro-nomical datasets. To address these challenges, we introduce STARLINC,the first ML-based framework for satellite trail removal without requir-ing tedious pixel-level annotation of astronomical images. STARLINCcombines synthetic satellite trail generation for training, inter-frame dif-ferential maps from temporally adjacent exposures to highlight transienttrails, and heatmaps to provide additional localization cues for pixel-level segmentation. Extensive experiments on real-world data demon-strate substantial improvements over baselines, establishing STARLINCas a scalable solution for next-generation astronomical surveys. Code isavailable at https://github.com/starioKim/STARLINC.

Implicit Neural Representations (INRs) have been provensuccessful in encoding continuous signals through coordinate-based net-works, yet facing a spectral dilemma: periodic activations capture finedetails but act as all-pass filters that memorise noise, while spatiallycompact activations regularise effectively but suffer from low-frequencybias. Existing attempts to resolve this trade-off introduce computationaloverhead or tuning frailty. We propose to model each neuron’s activationas the steady-state response of a sinusoidally-forced damped harmonicoscillator, whose amplitude naturally governs the network’s spectral se-lectivity during training. By jointly optimising the oscillator parametersalongside the network weights, our method adapts to the target signal’sspectral content without explicit regularisation. Initialised in the stop-band, the network exhibits a coarse-to-fine learning curriculum that pro-gressively expands its spectral gate, capturing low-frequency structuresfirst and high-frequency details only when justified by the reconstructionobjective. Comprehensive experiments show that our approach consis-tently achieves state-of-the-art or competitive results against establishedINRs, while requiring no task-specific tuning of any hyperparameters.Project Page available at https://alex-costanzino.github.io/fdho/.

Image restoration under diverse degradations remains chal-lenging for unified all-in-one frameworks due to feature interference andinsufficient expert specialization. We propose SLER-IR, a sphericallayer-wise expert routing framework that dynamically activates special-ized experts across network layers. To ensure reliable routing, we in-troduce a Spherical Uniform Degradation Embedding with contrastivelearning, which maps degradation representations onto a hypersphereto eliminate geometry bias in linear embedding spaces. In addition, aGlobal–Local Granularity Fusion (GLGF) module integrates global se-mantics and local degradation cues to address spatially non-uniformdegradations and the train–test granularity gap. Experiments on three-task and five-task benchmarks demonstrate that SLER-IR achievesconsistent improvements over state-of-the-art methods in both PSNRand SSIM. Code and models will be available at https://github.com/PSR666/SLER-IR.

Robust and efficient local feature matching remains a bot-tleneck for real-time visual localization, especially in environments withlow texture, repetitive structures, or illumination changes. Most priorapproaches model either appearance or geometry in isolation, failing toexploit the semantic–geometric interactions crucial for disambiguatingchallenging matches. To address this limitation, we propose SemLight,built on the insight that compact semantic cues can serve as a lightweightprior for resolving matching ambiguity under challenging conditions.Specifically, high-level semantic priors are first distilled into compact,matching-oriented descriptors via a task-aware student network. Thesedescriptors then conditionally re-weight appearance and geometric chan-nels locally, enabling efficient disambiguation with minimal computa-tional overhead. Extensive experiments on relative pose estimation, ho-mography estimation, and visual localization demonstrate state-of-the-art performance. Under comparable computational cost to XFeat, ourmethod achieves a 17.8% relative improvement in AUC@5 over XFeat(50.2 vs. 42.6).

Pretrained video generative models are promising backbonesfor visuomotor control, but their imagined futures often drift from taskintent and are not reliably action-conditional. As a result, these mod-els can be difficult to use for planning or policy extraction. To addressthese limitations, we propose RoboTALES, a single-stage framework thatlearns task-aligned simulated futures and uses them to train robot poli-cies. Our approach introduces two key innovations: (1) a hierarchicalLLM-based planner that breaks complex tasks into a sequence of sub-goals to guide the model’s imagination; and (2) a VLM-based critic thatevaluates these “imagined” futures and uses reward-based feedback tokeep the model’s internal representations focused on the goal. By an-choring the video generator in abstract reasoning, we produce temporallyconsistent rollouts and more coherent actions. We evaluate RoboTALESon diverse manipulation tasks from RoboCasa and LIBERO10, and showthat our method consistently outperforms existing methods, especiallyin long-horizon tasks. Our code and models are publicly available athttps://github.com/hananshafi/RoboTALES.

Humans learn locomotion through visual observation, inter-preting visual content first before imitating actions. However, state-of-the-art humanoid locomotion systems rely on either curated motion capturetrajectories or sparse text commands, leaving a critical gap between visualunderstanding and control. Text-to-motion methods suffer from semanticsparsity and staged pipeline errors, while video-based approaches onlyperform mechanical pose mimicry without genuine visual understanding.We propose RoboMirror, the first retargeting-free video-to-locomotionframework embodying “understand before you imitate”. Leveraging VLMs,it distills raw egocentric/third-person videos into visual motion intents,which directly condition a diffusion-based policy to generate physicallyplausible, semantically aligned locomotion without explicit pose recon-struction or retargeting. Extensive experiments validate RoboMirror’seffectiveness, it enables telepresence via egocentric videos, drasticallyreduces third-person control latency by 80%, and achieves a 3.7% highertask success rate than baselines. By reframing humanoid control aroundvideo understanding, we bridge the visual understanding and action gap.

We present RiO-DETR: DETR for Real-time OrientedObject Detection, the first real-time oriented detection transformer tothe best of our knowledge. Adapting DETR to oriented bounding boxes(OBBs) poses three challenges: semantics-dependent orientation, angleperiodicity that breaks standard Euclidean refinement, and an enlargedsearch space that slows convergence. RiO-DETR resolves these issueswith task-native designs while preserving real-time efficiency. First, wepropose Content-Driven Angle Estimation by decoupling angle from po-sitional queries, together with Rotation-Rectified Orthogonal Attentionto capture complementary cues for reliable orientation. Second, Decou-pled Periodic Refinement combines bounded coarse-to-fine updates witha Shortest-Path Periodic Loss for stable learning across angular seams.Third, Oriented Dense O2O injects angular diversity into dense super-vision to speed up angle convergence at no extra cost. Extensive ex-periments on DOTA-1.0, DIOR-R, and FAIR-1M-2.0 demonstrate RiO-DETR establishes a new speed–accuracy trade-off for real-time orienteddetection. Code is available at https://github.com/RicePasteM/RiO-DETR.

Egocentric devices, such as wearable front-facing cameras,provide a unique perspective for capturing the continuous interactionbetween a human viewer and the surrounding environment. A holisticand efficient multimodal model capable of reconstructing this 4D rep-resentation is therefore highly desirable. However, existing approachesoften rely on auxiliary inputs such as pre-computed camera trajecto-ries, treat scene perception and human ego-motion modeling as separateproblems despite their strong interdependency, and suffer from slow in-ference time. To address these limitations, we present ReViV, the firstunified framework for holistic egocentric 4D reconstruction that extractsboth viewer and view dynamics from a single monocular RGB video.We formulate the task as learning the full joint probability distribu-tion over multimodal signals, including RGB video, camera trajectory,gaze direction, full-body motion, hand motion, and depth. Powered bya Masked Generative Egocentric Transformer, ReViV operates within asingle feed-forward architecture to simultaneously reconstruct the tem-porally consistent 4D reconstruction across the viewer and the view withfast inference speed. Extensive experiments on diverse benchmarks, in-cluding HoloAssist, HOT3D, ARCTIC, Aria Digital Twin, and TACO,demonstrate that ReViV achieves state-of-the-art accuracy and efficiencyacross holistic ego-body, hand, and gaze reconstruction, camera track-ing, while maintaining highly competitive egocentric depth estimation,without relying on heavy task-specific priors. Code and models are fullyopen-sourced: https://reviv4d.github.io/.

Monocular depth foundation models achieve remarkable gen-eralization by learning large-scale semantic priors, but this creates a criti-cal vulnerability: they hallucinate illusory 3D structures from planar/low-curvature but perceptually ambiguous inputs. We term this failure the3D Mirage. This paper introduces a novel end-to-end framework toprobe, score, and tame this under-quantified safety risk in monoculardepth under context variation. To probe, we present 3D-Mirage, thefirst benchmark to combine context variation and precise annotation forreal-world illusions with real object exclusions, multi-surface support;purpose-built to stress-test monocular depth on real-world illusions. Toscore, we propose a second-order magnitude-based evaluation with twometrics: the Deviation Composite Score (DCS) for high second-order 3D structure and the Confusion Composite Score (CCS) forcontextual instability. To tame this failure, we introduce GroundedSelf-Distillation, a parameter-efficient strategy on Depth-Anything-V2baseline that surgically targets and resolves hallucination on illusion ROIswhile preserving background knowledge, avoiding catastrophic forgetting.Our work provides an innovative pipeline for diagnosing and addressingthis phenomenon, urging a necessary shift in the evaluation of MDE frompixel-wise accuracy to structural and contextual robustness.Code: https://github.com/hdnndh/The- 3D- Mirage- Probing- and-Taming-3D-HallucinationsDataset: https://huggingface.co/datasets/3dmirage/3D-Mirage

Visual navigation requires agents to reach goals in complexenvironments through perception and planning. World models addressthis task by simulating action-conditioned state transitions to predictfuture observations. Current navigation world models typically learnstate evolution under actions within the compressed latent space of aVariational Autoencoder, where spatial compression often discards fine-grained structural information and makes precise action-conditioned pre-diction more difficult. To better understand the propagation character-istics of different representations, we conduct a linear dynamics probeand observe that dense DINOv2 features exhibit stronger linear pre-dictability for action-conditioned transitions. Motivated by this obser-vation, we propose the Representation Autoencoder-based NavigationWorld Model (RAE-NWM), which generatively models navigation dy-namics in a dense visual representation space. We employ a Condi-tional Diffusion Transformer with a Decoupled Diffusion Transformerhead (CDiT-DH) to model continuous transitions, and introduce a sep-arate time-driven gating module for dynamics conditioning to regulateaction injection strength during generation. Extensive evaluations showthat modeling sequential rollouts in this space improves structural stabil-ity and action accuracy, benefiting downstream planning and navigation.Code is available at https://github.com/20robo/raenwm.

Split inference (SI) offloads computation from edge devicesand is often considered privacy-friendly since raw inputs remain local.However, recent data reconstruction attacks (DRA) show that interme-diate features can be inverted to recover sensitive content. Existing de-fenses suppress feature information through pruning, noise injection, ordecorrelation, yet strong reconstructions persist. In this paper, we ar-gue that reconstruction vulnerability is not solely determined by mutualinformation between inputs and representations, but also by the localorganization of representations that facilitates reliable inversion in prac-tice. Therefore, we propose ReTarget, a lightweight transformation thatpreserves task-discriminative representations while perturbing inversion-consistent features. ReTarget combines adversarial reconstruction, taskpreservation, and semantic regularization training to substantially de-grade inversion performance in practice without sacrificing utility. Ex-periments on CLIP-ViTB/16, CLIP-RN50 and DINOv2 across multiplesplit points and datasets show substantial degradation in reconstructionquality. Under both state-of-the-art and adaptive attacks, particularlyat shallow split points where prior defenses fail, our method remainseffective, where achieve up to 35% improvement while maintaining com-petitive task accuracy.

While diffusion models excel at generating images with con-ventional dimensions, pushing them to synthesize ultra-high-resolutionimagery at extreme aspect ratios (EAR) often triggers catastrophic struc-tural failures, such as object repetition and spatial fragmentation. Thislimitation fundamentally stems from a lack of robust spatial priors, asstatic text-to-image models are primarily trained on image distributionswith conventional dimensions. To overcome this bottleneck, we presentScrollScape, a novel framework that reformulates EAR image synthe-sis into a continuous video generation process through two core inno-vations. By mapping the spatial expansion of a massive canvas to thetemporal evolution of video frames, ScrollScape leverages the inherenttemporal consistency of video models as a powerful global constraint toensure long-range structural integrity. Specifically, Scanning PositionalEncoding (ScanPE) distributes global coordinates across frames to actas a flexible moving camera, while Scrolling Super-Resolution (ScrollSR)leverages video super-resolution priors to circumvent memory bottle-necks, efficiently scaling outputs to an unprecedented 32K resolution.Fine-tuned on a curated 3K multi-ratio image dataset, ScrollScape ef-fectively aligns pre-trained video priors with the EAR generation task.Extensive evaluations demonstrate that it significantly outperforms ex-isting image-diffusion baselines by eliminating severe localized artifacts.Consequently, our method overcomes inherent structural bottlenecks toensure exceptional global coherence and visual fidelity across diverse do-mains at extreme scales.

Online 3D scene graph generation builds a persistent, struc-tured representation of a scene by incrementally fusing 2D observationsinto a global 3D graph. Existing online methods treat this fusion asa fully deterministic pipeline, where we identify three sources of un-certainty that are overlooked: observation, 2D model, and 3D repre-sentation. We propose PUF: a Plug-and-play, Uncertainty-aware, andtraining-free Fusion framework. Scene graph node association is refor-mulated as a probabilistic likelihood over semantic and spatial factors,replacing binary accept/reject gates. Dirichlet evidence accumulation dis-tributes class and relationship evidence across plausible candidates pro-portional to association likelihood. An optional class-conditional priorcompletes edges for sparsely or never co-observed object pairs. We instan-tiate PUF with both a 3D Gaussian and a 3D voxel backend and observeconsistent improvements, demonstrating its ability to generalize acrossdifferent representations. Experiments on the 3DSSG and ReplicaSSGbenchmarks show that our method substantially outperforms existingapproaches while maintaining real-time latency. These results establishuncertainty-aware fusion as a principled and effective paradigm for on-line 3D scene understanding. The source code is publicly available athttps://github.com/yyyyangyi/PUF.

Fine-tuning pre-trained point-cloud backbones typically up-dates all parameters, resulting in substantial computation and memoryoverhead. More importantly, modern point backbones rely on aggressivetokenization and downsampling, which yields compact global tokens butirreversibly discards fine-grained local geometry, an inherent bottleneckfor parameter-efficient adaptation. Consequently, existing PEFT meth-ods that operate only on these coarsened tokens can modulate globalsemantics but struggle to recover the missing multi-scale locality. Wepresent Point Ladder Tuning (PLT), a locality-aware PEFT frameworkthat performs hierarchical, instance-conditioned adaptation while keep-ing the backbone frozen. PLT forms a lightweight closed loop: (i) a Hier-archical Ladder Network (HLN) constructs a multi-resolution local fea-ture pyramid directly from raw points; (ii) a Local–Global Fusion (LGF)aligns and fuses local pyramids with intermediate backbone semantics;and (iii) a Dynamic Prompt Generator produces instance-aware multi-scale prompts to modulate the frozen backbone effectively. For denseprediction, we further introduce a lightweight segmentation head thatprogressively upsamples fused features and leverages backbone priors torefine fine structures. Extensive experiments on classification and denseprediction show that PLT consistently surpasses prior PEFT baselineswith minimal tunable parameters. PLT achieves state-of-the-art perfor-mance using only 2.71% trainable parameters for classification and 7.69%for dense prediction, and scales favorably to larger backbones, requiringmerely 0.36% parameters on PointGPT-L.

State-of-the-art super-resolution (SR) transformers stack 16–30+ blocks of small shifted-window attention. Information between dis-tant positions must therefore traverse multiple attention layers, eachbounded by the local window—an indirect path that grows with net-work depth. We show that direct wide attention over large position-fixedwindows (64×64 at training, up to 128×128 at inference) can matchor surpass these deep narrow-window designs with only 12 blocks. Akey enabler is 2D Rotary Position Embeddings (RoPE), which encoderelative positions through query-key rotations rather than additive biasmatrices, enabling memory-efficient attention over thousands of tokens.To fully exploit this large spatial context, we replace the standard linearQKV projection with a rich nonlinear projection module that producesmore expressive features for each attention operation; sharing a singlemodule across all blocks keeps the model at 20M parameters while pro-viding the representational capacity that would otherwise require 115M.The shared module jointly produces spatial and channel QKV througha unified output. RoPE further enables resolution-adaptive inference viaYaRN-style scaling [22] with an inverted temperature correction tailoredfor SR. With only 20M parameters, our model, WARP, achieves PSNRcompetitive with or superior to state-of-the-art methods on standardbenchmarks.

We present PAI-Studio, a new reference-conditioned videosynthesis task that addresses a long-standing challenge in cinematic back-ground replacement: generating dynamic backgrounds aligned with fore-ground motion while preserving foreground identity, matching referencescene appearance, and achieving globally consistent illumination with re-alistic foreground relighting. Existing open-source systems and commer-cial APIs cannot simultaneously ensure motion-consistent backgroundgeneration, high-fidelity foreground relighting and foreground identitypreservation, often resulting in static backgrounds, inconsistent bound-aries, and noticeable compositing artifacts. To bridge this gap, we buildupon a Diffusion Transformer video backbone and reformulate the prob-lem as an in-context conditional generation task. Through bidirectionalattention, our model jointly captures foreground dynamics and back-ground reference information within a unified architecture. We furtherconstruct a 30K-scale dataset sourced from high-quality films and on-line videos to support this task. Extensive evaluations demonstrate thatour method significantly outperforms existing open-source and commer-cial API solutions. Project page: https://github.com/showlab/PAI-Studio.

We present a real-time, task-oriented 3D instance segmenta-tion framework for unposed monocular video, enabling embodied agentsto task-adaptively perceive and interact with objects in open-world scenes.Unlike most previous bottom-up segmentation paradigms that segmentbefore recognition, we adopt a task-oriented segmentation approach.Specifically, objects are decoupled within each frame using an open-vocabulary detector combined with a prompt-based 2D segmentationmodel, while the 3D underlying geometry of the scene is simultane-ously being reconstructed using a modern dense SLAM system, whosecontinuously re-optimized camera poses and depth are unsuitable forincremental mask association. Guided by the SLAM-derived pose graph,we selectively associate multi-view masks and reuse the dense correspon-dences provided by the SLAM system, incrementally converting theminto geometric association scores with minimal additional computation.By incorporating semantic similarity and mutual exclusivity metrics, wedesign a priority-ordered mask clustering algorithm for efficient onlinemulti-view mask matching and merging. Evaluations on open-vocabulary3D instance segmentation benchmarks show that our method effectivelymitigates the performance degradation of existing approaches when usingdense SLAM reconstructions instead of depth-sensor point clouds. Onthe Replica dataset, using only unposed images, it even achieves resultscomparable to methods leveraging ground-truth depth and poses. Moreimportantly, qualitative results show that our method achieves more reli-able task-oriented 3D object disentanglement than previous bottom-upsegmentation paradigms.