ECCV 2026 Papers

Diffusion models are powerful deep generative models, butunlike classical models, they lack an explicit low-dimensional latent spacethat parameterizes the data manifold. This absence makes it difficultto perform manifold-aware operations, such as geometrically faithful in-terpolation or conditional guidance that respects the learned manifold.We propose a training-free Riemannian metric on the noise space, de-rived from the Jacobian of the score function. The key insight is thatthe spectral structure of this Jacobian separates tangent and normaldirections of the data manifold; our metric leverages this separation to en-courage paths to stay tangential to the manifold rather than drift towardhigh-density regions. To validate that our metric faithfully captures themanifold geometry, we examine it from two complementary angles. First,geodesics under our metric yield perceptually more natural interpolationsthan existing methods on synthetic, image, and video frame datasets.Second, the tangent–normal decomposition induced by our metric pre-vents classifier-free guidance from deviating off the manifold, improvinggeneration quality while preserving text-image alignment.

The growing availability of trajectory datasets has fueledmajor advances in data-driven motion prediction. Yet, models trained onone dataset often fail to generalize beyond their training domain as a resultof differences in scene layouts, agent behaviors, and sensing conditions. Aframework that learns latent representations of datasets and quantifiestheir similarity using distributional metrics is presented. This large-scalestudy covers 24 major datasets, including the most widely used motion-prediction benchmarks, and shows that the resulting transferability scoresstrongly correlate with cross-dataset model performance. The resultsprovide practical guidance for dataset selection, pretraining, and large-scale foundation models for motion prediction, paving the way towardmore generalizable and robust predictive systems.

How can we tell whether a video has been sped up or sloweddown? How can we generate videos at different speeds? Although videoshave been central to modern computer vision research, little attentionhas been paid to perceiving and controlling the passage of time. In thispaper, we study time as a learnable visual concept and develop modelsfor reasoning about and manipulating the flow of time in videos.We first exploit the multimodal cues and temporal structure naturallypresent in videos to learn, in a self-supervised manner, to detect speedchanges and estimate playback speed. We then show that these learnedtemporal reasoning models enable us to curate the largest slow-motionvideo dataset to date from noisy in-the-wild sources. Such slow-motionfootage, typically filmed by high-speed cameras, contains substantiallyricher temporal detail than standard videos. Using this data, we furtherdevelop models capable of temporal control, including speed-conditionedvideo generation, which produces motion at specified playback speed,and temporal super-resolution, which transforms low-FPS, blurry videosinto high-FPS sequences with fine-grained temporal details. Our findingshighlight time as a manipulable, perceptual dimension in video learning,opening doors to temporally controllable video generation, temporalforensics detection, and potentially richer world-models that understandhow events unfold over time. Project page: https://seeing-fast-and-slow.github.io/.

Recent text-guided image editing (TIE) models have achievedremarkable progress, while many edited images still suffer from issuessuch as artifacts, unexpected editings, unaesthetic contents. Althoughsome benchmarks and methods have been proposed for evaluating editedimages, scalable evaluation models are still lacking, which limits the de-velopment of human feedback reward models for image editing. To ad-dress the challenges, we first introduce EditHF-1M, a million-scale im-age editing dataset with over 29M human preference pairs and 148Khuman mean opinion ratings, both evaluated from three dimensions,i.e., visual quality, instruction alignment, and attribute preservation.Based on EditHF-1M, we propose EditHF, a multimodal large lan-guage model (MLLM) based evaluation model, to provide human-alignedfeedback from image editing. Finally, we introduce EditHF-Reward,which utilizes EditHF as a reward signal to optimize the text-guidedimage editing models through reinforcement learning. Extensive experi-ments show that EditHF achieves superior alignment with human prefer-ences and demonstrates strong generalization on other datasets. Further-more, we fine-tune the Qwen-Image-Edit using EditHF-Reward, achiev-ing significant performance improvements, which demonstrates the abil-ity of EditHF to serve as a reward model to scale-up the image editing.Both the dataset and code will be released in our GitHub repository:https://github.com/IntMeGroup/EditHF.

Despite the remarkable progress of large vision languagemodels (LVLMs), object hallucination remains a fundamental challengethat hinders their trustworthy deployment. A key finding motivates ourwork: real and hallucinated object tokens are clearly separable in hid-den representations, yet this separability is largely lost at the language-modeling (LM) head. We propose TruthLens, a self-evaluation frame-work that teaches the LM head to expose a per-object truthfulness sig-nal without any auxiliary model or additional inference cost. Concretely,a rarely-used special token is repurposed as a reference token. For eachobject-token position, we extract the log-probability assigned to this spe-cial token by the LM head, and define its difference from a predefinedconstant as the truthfulness score. The model is then fine-tuned withan MSE objective that drives scores toward 1 for real objects and 0for hallucinated ones, while a divergence constraint preserves the origi-nal generation capability. Despite being trained on only a limited set ofobject categories, TruthLens generalizes effectively to benchmarks withsubstantially larger label spaces. Extensive experiments across multipleLVLMs demonstrate state-of-the-art performance; notably, on Qwen2.5-VL-7B, TruthLens outperforms the previous best method on MS-COCOby over 17% in AUROC. Our code is available at https://github.com/wyqstan/TruthLens.

3D Gaussian Splatting achieves impressive novel view syn-thesis for ground-level scenes. However, its performance degrades sig-nificantly in aerial domains due to sparse viewpoint sampling imposedby platform constraints. Existing sparse-view methods typically rely onstructure-from-motion reconstruction, which often yields incomplete ge-ometry due to limited feature correspondences under repetitive texturesand sparse viewpoint overlap. We present Geometry-Propagated Gaus-sian Splatting (GeoProp-GS), which overcomes the limitations of in-sufficient initialization. Our approach introduces two core components:Depth-guided Geometric Initialization (DGI) generates dense point cloudsand extends coverage to unobserved regions; Anchor-constrained Gaus-sian Optimization (AGO) stabilizes under-supervised regions by decom-posing proposal Gaussians into reliable anchors and learnable residuals.Extensive experiments demonstrate that GeoProp-GS achieves state-of-the-art performance and serves as an effective plug-and-play module thatconsistently improves existing methods. Code available at GeoProp-GS .

Implicit neural representation (INR) has achieved remark-able progress in novel view synthesis and image/video coding in recentyears. Compared to conventional end-to-end image codecs, INR-basedcompressors demonstrate significant advantages in decoding complexity.However, their practical application has been hindered by the inferiorencoding speed and underutilized decoding efficiency. In this work, wepropose a feedforward INR image coding architecture, Practical INRImage Codec (PIC), that computes all the necessary information forINR network in a single forward pass, achieving an encoding speed of20 FPS. Additionally, we implement a highly optimized decoder thatreaches 2000 FPS decoding speed, significantly surpassing JPEG’s per-formance at comparable rate-distortion (RD) performance. To the best ofour knowledge, this work presents the first learning-based image codecthat simultaneously outperforms or is comparable with JPEG in bothRD performance and decoding speed while maintaining practical encod-ing speed. Code is available at https://github.com/actcwlf/PIC.

Uncalibrated volumetric video streaming for human reconstruction is essential for holographic communication and AR/VR, yet remains challenging due to the need for temporal consistency and computational efficiency from sparse-view inputs. Existing methods rely on perscene optimization or calibrated cameras, while recent feed-forward models are limited to low-resolution (0.5K) single-frame synthesis. We present HiReFF, a feed-forward method for 2K-resolution 360° human video reconstruction from uncalibrated sparse-view videos. Our framework decomposes the problem into two key tasks: foreground 3D Gaussian reconstruction from sparse-view videos (four views separated by 90°) and computationally efficient high-resolution synthesis. To enable the former, we propose Scale-synchronized Camera Calibration to resolve scale ambiguity for multi-view supervision, and Gaussian-wise Foreground Masking to reconstruct clean foregrounds by modulating Gaussian parameters. For efficient high-resolution synthesis, our High-resolution Side-tuning achieves 2K rendering by augmenting the Gaussian head with supplementary features while keeping the backbone at 0.5K, drastically reducing computational overhead. Experiments demonstrate that HiReFF significantly outperforms existing methods in high-resolution streaming volumetric video reconstruction. https://iridescentjiang.github.io/ HiReFF

Recent diffusion-based models have enabled realistic audio-driven avatar generation in real-time streaming. However, existing ap-proaches struggle to maintain visual temporal consistency and fail toexplicitly perceive user intent in complex interactive streaming scenar-ios. To address these challenges, we propose InteractiveAvatar, a real-time infinite-streaming video generation framework that supports vi-sually consistent avatar video generation and intent-aware interactions.With autoregressive distillation, InteractiveAvatar achieves real-time str-eaming generation of human avatars over arbitrarily long durations. Forvisual consistency, we introduce a Long-Short Visual Memory (LSVM)mechanism that flexibly compresses historical visual information intocompact tokens, preserving both short-range coherence and long-termconsistency. To generate avatars with speeches and actions aligned withuser intent, we propose a Reasoning-Reaction Module (RRM), which in-corporates a State-Cycling strategy and a Cache-Switching mechanism.Extensive experimental results over diverse scenarios demonstrate thatour method achieves state-of-the-art visual consistency in long-durationgeneration, while enabling complex user-avatar interaction in real time.

Learning effective robot control policies on physical hard-ware is challenging due to costly data collection and the difficulty ofreward specification. Prior work has incorporated demonstrations intoreinforcement learning (RL), yet existing approaches either require largenumbers of demonstrations or depend on continuous human interventionduring training. To address these limitations, we present AutoSERL, aframework that leverages a single demonstration to fully automate the in-tervention process in real-world robot RL. The framework includes threecomplementary mechanisms to accomplish certain tasks: a sliding win-dow intervention that continuously guides exploration to prevent localoptima and unsafe deviations, a safety recovery mechanism that detectsand corrects failure states via predefined trajectory recovery points, andan intervention termination criterion that automatically disables guid-ance once the policy can independently complete the task, preserving itsexploration advantage. We evaluate AutoSERL on six contact-intensivemanipulation tasks across two robot platforms, spanning insertion, hang-ing, and hinge-based tasks. AutoSERL consistently outperforms SERLinitialized with 20 demonstrations, behavior cloning, and MILES — adedicated one-shot imitation learning baseline — across all tasks whilematching HIL-SERL, achieves 100% success rate on insertion tasks, anddemonstrates improved robustness to positional variations, all from a sin-gle demonstration. Code and videos are available on our project website:https://autoserl.github.io/.

High definition map generation is essential for autonomousdriving, yet remains a labor-intensive process at scale. We present Map-Dreamer, a generative diffusion model that synthesizes lane-level vec-tor maps with explicit topology directly from a single aerial image.MapDreamer learns a compact latent representation of lane centerlinesand their topological relations using a variational autoencoder and pre-dicts graphs with a transformer-based latent diffusion model. To aligngenerated maps with the observed scene, we condition each denoisingstep on dense aerial features injected through cross-attention. To handlethe varying number of lanes across scenes, we propose a lane cardinal-ity module paired with background ghost lane latents, a learned bufferthat prevents slot collapse during diffusion. Furthermore, we introduce asliding-window global graph aggregation strategy that stitches local tilesinto city-scale maps while preserving connectivity through encoded laneboundaries. Experiments on UrbanLaneGraph derived from Argoverse2 show improved geometric and topological fidelity over non-generativebaselines.

High-resolution (HR) image perception remains a key chal-lenge in multimodal large language models (MLLMs), as fine-graineddetails are often lost when the image is processed as a whole. Existingmethods either require training to teach models where to look or heuris-tically divide the image into fixed regions, both of which struggle togeneralize in complex HR scenes. In this work, we propose HierarchicalEntity Exploration (HEE), a training-free and model-agnostic frameworkthat transforms static image understanding into dynamic, query-guidedentity exploration. HEE first evaluates each region using a dual scoringmechanism to determine whether it already contains sufficient evidenceto answer the question. If not, it applies object detection within themost promising region to extract fine-grained entities, clusters them intocoherent subregions, and organizes them into a multi-level semantic hi-erarchy for deeper exploration. When deeper regions still fail to yieldconfident answers, a confidence-guided backtracking mechanism revisitsalternative paths to ensure adaptive perception. Extensive results showthat HEE outperforms training-free methods like ZoomEye and RAPin both accuracy and efficiency on two complex HR benchmarks (VisualProbe and HR-Bench), across different MLLMs such as Qwen2.5-VL andLLaVA-OneVision. Moreover, HEE demonstrates generalization on theMME-RealWorld benchmark.

Dynamical systems theory and reinforcement learning viewworld evolution as the dynamics of latent states driven by actions, withvisual observations providing partial information about those states. Re-cent video world models attempt to learn these action-conditioned dy-namics from data. However, existing datasets rarely meet these require-ments: they typically lack diverse and semantically meaningful actionspaces, and actions are directly tied to visual observations rather thanmediated by underlying states. As a result, actions are often entangledwith pixel-level changes, making it difficult for models to learn structuredworld dynamics and maintain consistent evolution over long horizons. Inthis paper, we propose WildWorld, a large-scale action-conditioned worldmodeling dataset with explicit state annotations, automatically collectedfrom a photorealistic AAA action role-playing game (Monster Hunter:Wilds). WildWorld contains over 108 million frames and features morethan 450 actions, including movement, attacks, and skill casting, togetherwith synchronized per-frame annotations of character skeletons, worldstates, camera poses, and depth maps. We further derive WildBench toevaluate models through Action Following and State Alignment. Exten-sive experiments reveal persistent challenges in modeling semanticallyrich actions and maintaining long-horizon state consistency, highlightingthe need for state-aware video generation.

Single-photon LiDAR (SP-LiDAR) simulators face a dilemma:fast but inaccurate Poisson models or accurate but prohibitively slow se-quential models. This paper breaks that compromise. We present a sim-ulator that achieves both fidelity and speed by focusing on the critical,yet overlooked, component of simulation: the photon count statistics. Ourkey contribution is a Markov-renewal process (MRP) formulation that,for the first time, analytically predicts the distribution of registered pho-ton counts under dead time. To make this MRP model computationallytractable, we introduce a spectral truncation rule that efficiently com-putes the complex covariance statistics. By proving the shift-invarianceof the process, we extend this per-pixel model to full histogram cubegeneration via a precomputed lookup table. Our method generates 3Dcubes indistinguishable from the sequential gold standard, yet is ordersof magnitude faster. This finally enables large-scale, physically faithfuldata generation for learning-based SP-LiDAR reconstruction. Projectpage: https://mars-splidar.github.io/.

Understanding emotions is a fundamental ability for intelligent systems to be able to interact with humans. Vision-language models (VLMs) have made tremendous progress in the last few years for many visual tasks, potentially o!ering a promising solution for understanding emotions. However, it is surprising that even the most sophisticated contemporary VLMs struggle to recognize human emotions or to outperform even specialized vision-only classifiers. In this paper we ask the question ‘Why do VLMs struggle to recognize human emotions?’, and observe that the inherently continuous and dynamic task of facial expression recognition (DFER) exposes two critical VLM vulnerabilities. First, emotion datasets are naturally long-tailed, and the web-scale data used to pre-train VLMs exacerbates this head-class bias, causing them to systematically collapse rare, under-represented emotions into common categories. We propose alternative sampling strategies that prevent favoring common concepts. Second, temporal information is critical for understanding emotions. However, VLMs are unable to represent temporal information over dense frame sequences, as they are limited by context size and the number of tokens that can fit in memory, which poses a clear challenge for emotion recognition. We demonstrate that the sparse temporal sampling strategy used in VLMs is inherently misaligned with the fleeting nature of micro-expressions (0.25–0.5 seconds), which are often the most critical a!ective signal. As a diagnostic probe, we propose a multi-stage context enrichment strategy that utilizes the information from ‘in-between’ frames by first converting them into natural language summaries. This enriched textual context is provided as input to the VLM alongside sparse keyframes, preventing attentional dilution from excessive visual data while preserving the emotional trajectory.

Recent advances in 3D Gaussian Splatting (3DGS) have ach-ieved significant success in high-fidelity Novel View Synthesis (NVS),yet the optimization process still introduces noisy Gaussian primitivesdue to the sparse and incomplete initialization from Structure-from-Motion (SfM) point clouds. Most existing methods focus only on adjust-ing the positions of primitives during optimization, while ignoring theunderlying spatial structure. To this end, we introduce a new perspec-tive by formulating the optimization of 3DGS as a primitive denoisingprocess and propose Denoising-GS, a spatial-aware denoising frame-work for Gaussian primitives by taking both the positions and spatialstructure into consideration. Specifically, we design an optimizer thatpreserves the spatial optimization flow of primitives, facilitating coher-ent and directed denoising rather than random perturbations. Buildingupon this, the Spatial Gradient-based Denoising strategy jointlyconsiders the spatial supports of primitives to ensure gradient-consistentupdates. Furthermore, the Uncertainty-based Denoising module es-timates primitive-wise uncertainty to prune redundant or noisy primi-tives, while the Spatial Coherence Refinement strategy selectivelysplits primitives in sparse regions to maintain structural completeness.Experiments conducted on three benchmark datasets demonstrate thatDenoising-GS consistently enhances NVS fidelity while maintaining rep-resentation compactness, achieving state-of-the-art performance acrossall benchmarks. Source code and models will be made publicly available.(a) (b) (c) (d)Point Cloud Spatial Gradient- Uncertainty-basedPosition DenoisingDenoising based Denoising Denoising↓ optimize and denoise ↓ ↓calculate uncertainty and prune↓Fig. 1: Illustrative comparison of different denoising methods: (a) Point cloud denois-ing, which moves noisy points toward the GT surface; (b) Position denoising of Gaus-sian primitives, which adjusts the mean positions of noisy primitives toward the GTsurface; (c) Spatial gradient-based denoising of Gaussian primitives, which optimizesboth the mean positions and the spatial structure of primitives based on gradients; (d)Uncertainty-based denoising of Gaussian primitives, which estimates the uncertaintyof each primitive and prunes those with high uncertainty.

Self-supervised learning (SSL) has produced a diverse land-scape of vision transformers (ViTs) whose pretrained representations sup-port a wide range of downstream tasks. Towards a better understandingof these models, a body of works has assessed the mechanics of their self-attention as well as which types of information they capture across theirrepresentations, revealing, e.g., stark differences between models trainedwith contrastive learning (CL) vs. masked image modeling (MIM). How-ever, the total of these advances on model understanding has to datenot yet fully permeated a larger community, where, e.g., insights thatare specific to CL models are still at times generalized to MIM models.To make model understanding straightforward and intuitive for a broadcommunity, we propose a simple and easily interpretable visualizationprotocol. Our protocol is based on visualizing unsupervised semanticsegmentation results – yet by no means do we focus on top segmenta-tion performance. Instead, our protocol allows us to easily convey modelbehavior that consistently emerges across images. Benchmarked on a di-verse set of SSL models across layers and representations, our protocolallows us to gain novel insights into distinct positional biases and scalingbehaviors, including, e.g., strong boundary artifacts in DINOv3-Largemodel tokens. These novel insights come on top of more easily conveyinga range of previous findings. Our protocol further allows us to clearly vi-sually convey and distinguish between positional effects and the closelyrelated yet distinct locality bias, the latter being much more extensivelystudied in the literature so far. Our protocol is publicly available6 , serv-ing to catalyze further model understanding for a broad community.

Sparse-view Computed Tomography (CT) reconstructs im-ages from a limited number of X-ray projections to reduce radiation andscanning time, which is an ill-posed inverse problem. Existing methodsachieve high-fidelity reconstructions but overfit to a fixed acquisitionsetup, failing to generalize well across sampling rates. For example, con-volutional neural networks (CNNs) use the same kernels across resolu-tions, leading to artifacts when data resolution changes. This is a criticallimitation in clinical practice, where acquisition sampling settings varyacross organs and diagnostic protocols.We propose Computed Tomography neural Operator (CTO), the firstneural operator (NO) framework for CT reconstruction. CTO extendslearning from fixed discretized grids to continuous function space, en-abling a single model to generalize across measurement sampling rateswithout retraining. We also propose new NO architectural designs forCT: (i) a dual-domain NO architecture in both sinogram and imagespaces, capturing complementary spatial–frequency information, and (ii)rotation-equivariant DIScrete–COntinuous convolutions (DISCO) thatexploit the rotational structure inherent in tomographic acquisition. Em-pirically, CTO outperform CNNs (> 3.4dB PSNR gain) and other base-lines in multi-resolution settings across multiple CT datasets. Comparedto state-of-the-art di!usion methods, CTO has 500→ faster inferencewith an average 3dB gain. CTO further demonstrates strong out-of-distribution robustness, maintaining gains under cross-dataset transferand noisy sinogram conditions. Ablation studies also validate each de-sign choice. CTO establishes neural operators as a principled and practi-cal paradigm for flexible, discretization-agnostic CT reconstruction. Ourcode is available at https://github.com/neuraloperator/sparse_ct.

RGBT tracking on edge devices is often constrained by lim-ited computational power and data transmission bandwidth, necessitat-ing the use of low-resolution inputs in real-world deployments. However,such resolution reduction typically leads to severe semantic loss and per-formance degradation. In this paper, we propose GAP-Track, an efficientframework that bridges the resolution gap by enabling high-precisiontracking of low-resolution inputs. Specifically, we introduce a hierarchicalknowledge distillation strategy that guides the low-resolution student bymulti-level teacher supervision. To recover the missing semantic details,a generative reconstruction module trained by a random masking strat-egy is introduced to strengthen the student’s feature representation byforcing the backbone to recover fine-grained structural information fromlow resolution data. Furthermore, we adapt a polar-geometric sensitivityloss that transforms the bounding box regression into a polar coordinatesystem, effectively mitigating localization ambiguities in extremely low-resolution scenarios by providing higher sensitivity to center-point driftand shape deformations. Extensive experiments on the comprehensiveLasHeR and RGBT234 datasets demonstrate that our method achievesstate-of-the-art performance. Notably, GAP-Track delivers superior pre-cision at 1/2 resolution and continues to significantly outperform thebaseline even at 1/4 resolution, ensuring a robust balance between track-ing accuracy and inference efficiency. The source code will be availableat https://github.com/shijiahahaha/GAPTrack.