ECCV 2026 Papers

Adapting Multimodal Large Language Models (MLLMs) forhour-long videos is bottlenecked by context window limits. Existing effi-ciency heuristics blindly sacrifice fidelity. They frequently discard tran-sient decisive moments, blur fine-grained evidence, and waste represen-tational bandwidth on irrelevant backgrounds. We propose Tempo, anefficient, query-aware framework that compresses long videos for down-stream understanding. Tempo unifies a Small Vision-Language Model(SVLM) to act as a local temporal compressor. It casts visual token re-duction as an early cross-modal distillation process, generating compact,intent-aligned video representations in a single forward pass. To enforcestrict inference budgets without breaking temporal causality, we intro-duce Adaptive Token Allocation (ATA). Exploiting the SVLM’s inher-ent zero-shot relevance prior and empirical semantic front-loading, ATAacts as a training-free, O(1) dynamic router. It allocates dense band-width to query-critical segments while compressing redundancies downto minimal temporal anchors to maintain the global storyline. Exten-sive experiments demonstrate that our compact 6B architecture achieveshighly competitive performance with aggressive dynamic compression(0.5–16 tokens/frame). On the extreme-long LVBench (4101s), Temposcores 52.3 under a strict 8K visual budget, outperforming proprietarymodels like GPT-4o and Gemini 1.5 Pro. Crucially, empirical profiling re-veals that Tempo frequently compresses hour-long videos to token countsbelow theoretical limits, proving that true long-form understanding relieson intent-driven efficiency rather than greedily padded context windows.

Reinforcement learning (RL) has recently emerged as a promis-ing approach for aligning text-to-image generative models with humanpreferences. A key challenge, however, lies in designing effective and in-terpretable rewards. Existing methods often rely on either compositemetrics (e.g., CLIP, OCR, and realism scores) with fixed weights or asingle scalar reward distilled from human preference models, which canlimit interpretability and flexibility. We propose RubricRL, a simple andgeneral framework for rubric-based reward design that offers greater in-terpretability, composability, and user control. Instead of using a black-box scalar signal, RubricRL dynamically constructs a structured rubricfor each prompt—a decomposable checklist of fine-grained visual cri-teria such as object correctness, attribute accuracy, OCR fidelity, andrealism—tailored to the input text. Each criterion is independently eval-uated by a multimodal judge (e.g., o4-mini), and a prompt-adaptiveweighting mechanism emphasizes the most relevant dimensions. This de-sign not only produces interpretable and modular supervision signalsfor policy optimization (e.g., GRPO or DiffusionNFT), but also enablesusers to directly adjust which aspects to reward or penalize. Experimentswith autoregressive and diffusion text-to-image models demonstrate thatRubricRL improves prompt faithfulness, visual detail, and generalizabil-ity, while offering a flexible and extensible foundation for interpretableRL alignment across text-to-image architectures.

Video world models can generate realistic futures from a sin-gle instruction, but they often fail to track the same physical pointsconsistently across time. As a result, the generated videos appear plausi-ble, yet lack the physical grounding required for reliable action execution,such as robot manipulation. We present GEM-4D, a geometry-groundedvideo world model that resolves this limitation by injecting dense 4Dcorrespondence supervision distilled from a pretrained geometry foun-dation model into the video generative backbone during training. Thissupervision enables the video world model to jointly capture appearanceand geometric structure while retaining a single-stream architecture withno additional inference cost. We further introduce an inverse dynam-ics module that converts correspondence-consistent video rollouts intoexecutable robot trajectories, enabling direct deployment in both real-world and simulated manipulation. GEM-4D achieves state-of-the-artperformance on both video prediction and geometric consistency acrossboth simulation and realistic scenarios and improves real-world manipu-lation success from 61% to 81%. Additional results are available at thehttps://gem-4d.github.io/.

Enabling humanoid robots to respond to human speech withsynchronized and semantically meaningful gestures is fundamental tonatural human-robot interaction. However, this task faces three criticalbarriers: the scarcity of semantically rich datasets, the "modality eclipse"where models ignore audio cues in favor of kinematic inertia, and the sim-to-real gap regarding physical safety. We propose RoboGesture, a robot-centric framework that co-designs data, modeling, and control to powera complete interactive human–humanoid system in which the robot lis-tens, responds, and gestures in real time. We first establish the Robo-Gesture dataset featuring over 300 gesture categories and develop anautomated pipeline to synthesize large-scale collision-free, robot-specificaudio-motion pairs. Our architecture features a Hierarchical Semantic-Acoustic Aligner that extracts multi-granular prosodic and semantic cuesdirectly from raw audio tokens. These cues drive a Streaming ConditionalMotion Generator based on a diffusion transformer with conditional flowmatching. To ensure high responsiveness, we introduce Anti-Inertia CFGMasking, which prevents the model from collapsing into repetitive his-torical patterns by compelling it to proactively mine control signals fromthe audio modality. Finally, an MPC-based safety filter ensures real-time, collision-free execution on physical hardware. Experiments on aUnitree G1 humanoid demonstrate that RoboGesture generates safer,more rhythmic, and more semantically appropriate responses comparedto state-of-the-art baselines.

Existing pruning methods for 3D Gaussian splatting (3DGS)suffer from either severe quality degradation or prohibitive computa-tional overhead. In this paper, we propose REFINE, a highly acceler-ated 3DGS pruning framework centered on a novel rendering-free primi-tive importance metric. Our approach leverages an analytically approx-imated, rendering-aware Hessian field to quantify the expected percep-tual error induced by the removal of individual primitives. By modelingthe joint modulation of visibility, projection geometry and the contentadaptive hyperparameter, we entirely bypass costly forward renderingpasses and derive an anisotropic perceptual weight field that serves asa high-fidelity proxy for primitive importance. Extensive experimentsacross multiple benchmark datasets demonstrate that REFINE main-tains highly competitive rendering quality while achieving a 3, 000× re-duction in pruning-related computational complexity, translating to apractical ∼ 20× speedup in device latency compared to state-of-the-artpruning methods.

Pre-trained Vision Foundation Models (VFMs) have becomecentral to modern computer vision due to their powerful semantic rep-resentations and strong generalization ability. However, their patchifiedor pooled outputs are inherently low-resolution, limiting their effective-ness in tasks requiring fine-grained, pixel-level reasoning. Existing fea-ture upsampling approaches either degrade semantic fidelity or rely onVFM-specific retraining and heavy architectures, hindering efficiency andscalability. To address these challenges, we propose RaysUp, an ultra-lightweight, task-agnostic, and VFM-agnostic feature upsampling frame-work that reconstructs high-resolution feature maps at arbitrary resolu-tions. Unlike conventional 2D interpolation or attention-based schemes,RaysUp lifts feature reconstruction into a geometry-aware ray domain.Specifically, we introduce a Spatially Decoupled Guidance Encoder fordirection-aware guidance encoding, an Any-Resolution Cross-Attentionmechanism for resolution-flexible reconstruction, and a novel Ray Posi-tional Encoding (RayPE) that injects implicit 3D geometric priors via6D Plücker ray coordinates. Finally, A Geometry-Aware NeighborhoodAttention module further ensures content-adaptive bilateral aggregationwhile preserving geometric consistency. Extensive experiments across di-verse dense prediction tasks demonstrate that RaysUp achieves state-of-the-art performance while using only 16% of the parameters of AnyUpand delivering approximately 7× faster inference. These results highlighta substantially improved accuracy–efficiency trade-off and establish Ray-sUp as a practical and scalable solution for universal feature upsampling.Code is available at https://github.com/MAP-RaysUp/RaysUp.

The development of 0.1◦ global weather forecasting modelsbased on machine learning (ML) is constrained by the limited availabil-ity of high-resolution data, as decades of reanalysis are only availableat 0.25◦ resolution. While existing approaches fine-tune 0.25◦ forecastmodels on limited 0.1◦ samples, we show that this transfer is hinderedby the irreversible information loss inherent in coarse-resolution fore-casting. Therefore, we propose BaguanHR, a framework that shiftsthe focus from transferring models to transferring data. We first showthat super-resolution (SR) has lower conditional entropy and input am-plification than forecasting, making it a more robust vehicle for reso-lution transfer. By leveraging this advantage through variable-wise SR,we synthesize extensive 0.1◦ data from ERA5. BaguanHR’s performanceon the synthetic-plus-real dataset exceeds both ML-based methods andIFS-HRES, achieving superior performance across over 85% of the leadtimes within 72 hours. Furthermore, our findings highlight a power-lawscaling effect, as a twofold increase in data reduces RMSE by 4.6% for 72-hour forecasting and 4.9% for 120-hour forecasting. Our results demon-strate that scaling high-resolution ML-based forecasting is primarily adata bottleneck, and that variable-wise super-resolution provides a sim-ple yet general solution to unlock long coarse-resolution reanalyses forhigh-resolution training.

While single-view 3D reconstruction has seen significant pro-gress, extrapolating complex 3D structures from inherently ambiguous2D observations remains fundamentally ill-posed, particularly in the crit-ically underexplored data-scarce regime. To address this challenge, wepropose Point Diffusion Mamba (PDM), a method that integrates thegenerative power of diffusion models with the efficiency of state-spacemodel for single-view 3D reconstruction under data-scarce conditions.Specifically, PDM employs a lightweight reconstruction module tailoredto handle unordered point-cloud inputs effectively. By combining a Lo-cal Geometric Aggregation module with Mamba blocks, our approachjointly models global geometric structures and local details. In 3D re-construction, each point in the initial noisy input requires a precise pre-diction, yet the high-level features extracted by the Mamba module cap-ture only abstract semantic information from sparse points. To bridgethis gap, we introduce the Hierarchical Feature Integration Network,which fuses high-level semantic and local geometric features for eachpoint, overcoming the limitations of token-based point-cloud reconstruc-tion. Furthermore, we propose a Dynamic Weighted Sampling strategythat adaptively unifies 3D generation with single-view reconstruction byleveraging generative priors to enhance reconstruction quality. Exper-imental results on the ShapeNet and Pix3D benchmarks demonstratethat PDM outperforms state-of-the-art methods, providing an effectivesolution for 3D reconstruction under data-scarce settings. Code is avail-able at: https://github.com/NWUzhouwei/PDM.

This paper addresses the fundamental challenge of accelerating the inherently autoregressive sampling process in Flow Matching (FM) models through a numerical systems perspective. We introduce ParaFlow, a training-free framework that recasts sampling as a system of Triangular Nonlinear Equations (TNEs) to enable step-level parallelism. Unlike existing parallel samplers for diffusion models that require 2–3× more function evaluations (NFEs), we propose a velocity approximation scheme that leverages the temporal smoothness of FM trajectories. This allows ParaFlow to perform parallel sampling on a single GPU while requiring less total NFE count to the standard sequential sampler. We theoretically guarantee that our method converges to the exact autoregressive trajectory and that our approximation maintains negligible error bounds. Extensive experiments on Stable Diffusion 3 and Flux demonstrate that ParaFlow achieves up to 4.3× wall-clock speedup with negligible impact on quality, making high-performance parallel sampling truly practical for resource-constrained settings. The source code is released in the supplementary materials. Code is available at https://github.com/Jianrong-Lu/ParaFlow.

Interactive segmentation of 3D Gaussians offers a compellingopportunity for real-time manipulation of 3D scenes, thanks to the real-time rendering capability of 3D Gaussian Splatting (3DGS). However, ex-isting methods require a time-consuming per-scene setup—typically tensof seconds or even minutes—before interactive segmentation can beginon a raw 3DGS scene. This setup involves multi-view mask preparation,mask lifting, and feature distillation, creating a major bottleneck for on-line applications. To address this limitation, we aim to completely elimi-nate the setup stage for interactive 3DGS segmentation while keeping thesegmentation time practical (under 1 second). In this work, we presentSAGO (Segment Any Gaussians Online), a novel setup-free frameworkfor interactive 3DGS segmentation. By introducing virtual drones, ourmethod reframes the 3D segmentation problem as an online Next-Best-View (NBV) planning task formulated within a Markov process. Exten-sive experiments demonstrate that SAGO can extract clean 3D assetsdirectly from 3D Gaussians with sub-second latency, thereby enablinga broad range of downstream applications such as object manipulationand scene editing. Moreover, our method achieves over a 50× speedupcompared to the previous setup-free 3DGS segmentation frameworks.

X-ray computed tomography (CT) reconstruction is inher-ently an ill-posed inverse problem, particularly with sparse-view pro-jections. Unlike traditional frame-based sensors that acquire projectionsat fixed intervals, neuromorphic event cameras operate asynchronouslyat the pixel level and trigger events only when the log-intensity changeexceeds a threshold. This asynchronous sensing naturally captures high-frequency angular variations during object rotation, providing comple-mentary information between sparsely sampled projections. However,the application of event cameras to CT reconstruction remains largelyunexplored. To bridge this gap, we introduce a Neuromorphic X-rayCT framework. Specifically, we propose an Event-enhanced Neural At-tenuation Field (ENAF) that models the volumetric attenuation fieldby jointly leveraging sparse projections and event streams. To supportand validate this framework, we construct synthetic neuromorphic CTdatasets and acquire a real-world dataset. Experiments on both syn-thetic and real-world datasets demonstrate that ENAF consistently out-performs state-of-the-art frame-only methods while maintaining train-ing efficiency. To the best of our knowledge, this work presents thefirst application of neuromorphic sensing to X-ray CT and establishesthe feasibility of event-enhanced sparse-view CT reconstruction. Thecode and datasets are available at https://wanghongjian98.github.io/projects/neuroxct/.

Flow-based generative methods offer a simple and effectiveframework for high-fidelity generation, yet pretrained flow models arerarely used in their vanilla conditional form: in image generation, sam-ples without guidance often appear diffuse and lack fine-grained detail.Existing guidance techniques such as classifier-free guidance (CFG) im-prove fidelity but reduce sample diversity. We introduce MomentumGuidance (MG), a guidance method that improves sample quality byextrapolating the current velocity away from an exponential moving av-erage of past velocities along the ODE trajectory, while preserving thestandard one-evaluation-per-step cost. MG provides gains beyond CFG,improving the precision–recall Pareto frontier. Experiments demonstratethe effectiveness of MG across benchmarks. On ImageNet-256, MG im-proves FID by 36.54% without CFG and 25.42% with CFG on averageacross sampling settings, attaining an FID of 1.553 at 16 sampling steps.Evaluations on large flow-based models, including Stable Diffusion 3 andFLUX.1-dev, further confirm improvements across standard metrics.

Vision-Language Action (VLAs) models promise to extendthe remarkable success of vision-language models (VLMs) to robotics. Yet,unlike VLMs in the vision-language domain, VLAs for robotics requirefinetuning to contend with varying physical factors like robot embodiment,environment characteristics, and spatial relationships of each task. Wepropose leveraging few-shot demonstrations of tasks to better capture thisphysical variability. Overfitting to the trained task is yet another problemof existing methods. Inspired by functional specificity, we hypothesizethat finetuning only task-specific sparse model representations is bothmore effective and better retains pretrained model generality. In thiswork, we introduce Robotic Steering, a finetuning approach groundedin mechanistic interpretability that leverages few-shot demonstrations toidentify and selectively finetune task-specific attention heads aligned withthe physical, visual, and linguistic requirements of robotic tasks. Throughcomprehensive on-robot evaluations with a Franka Emika robot arm, wedemonstrate that Robotic Steering outperforms LoRA while achievingsuperior robustness under task variation, reduced computational cost,and enhanced interpretability for adapting VLAs to diverse robotic tasks.

Video frame interpolation (VFI) remains a challenging task,particularly when dealing with large, non-linear motions and complexocclusions. While flow-based methods are prevalent, they often strugglewith ambiguous correspondences. Recent VFI methods based on selectiveState Space Models (SSMs) are still limited by static grid-based scanningthat misaligns with physical motion. In this paper, we propose Motion-Aligned Selective Scan (MASS), a novel framework that reformulates fea-ture scanning from static spatial grids to dynamic motion trajectories.MASS builds a feature sequence along each pixel’s flow-guided trajectoryand aggregates it with an SSM. Specifically, we introduce a learnable non-linear path integration to approximate complex curved trajectories viaresidual velocity updates, and a velocity-aware SSM that dynamicallyadjusts the sampling budget and step size based on motion magnitude.This adaptive strategy allocates denser sampling to fast-motion regionswhile keeping static regions efficient. Furthermore, the aggregated statesguide a refinement module to rectify intermediate flows and masks in anend-to-end manner. Extensive experiments indicate that MASS achieveshighly competitive overall performance on standard benchmarks, estab-lishing state-of-the-art results particularly in challenging scenarios withlarge displacements and complex dynamics.

Synthesizing high-fidelity and controllable 4D LiDAR datais crucial for creating scalable simulation environments for autonomousdriving. This task is inherently challenging due to the sensor’s uniquespherical geometry, the temporal sparsity of point clouds, and the com-plexity of dynamic scenes. To address these challenges, we present LiS-TAR, a novel generative world model that operates directly on the sen-sor’s native geometry. LiSTAR introduces a Hybrid-Cylindrical-Spherical(HCS) representation to preserve data fidelity by mitigating quanti-zation artifacts common in Cartesian grids. To capture complex dy-namics from sparse temporal data, it utilizes a Spatio-Temporal At-tention with Ray-Centric Transformer (START) that explicitly modelsfeature evolution along individual sensor rays for robust temporal co-herence. Furthermore, for controllable synthesis, we propose a novel 4Dpoint cloud-aligned voxel layout for conditioning and a correspondingdiscrete Masked Generative START (MaskSTART) framework, whichlearns a compact, tokenized representation of the scene, enabling effi-cient, high-resolution, and layout-guided compositional generation. Com-prehensive experiments validate LiSTAR’s state-of-the-art performanceacross 4D LiDAR reconstruction, prediction, and conditional genera-tion. Our method achieves substantial quantitative gains: improving re-construction IoU by up to 85%, lowering prediction L1 Med by up to64%, and reducing generation MMD by 61%. This level of performanceprovides a powerful new foundation for creating realistic and control-lable simulations for autonomous driving systems. Project Page: https://ocean-luna.github.io/LiSTAR.github.io/.

In autonomous driving scenarios, the collected LiDAR pointclouds can be affected by occlusion and long-range sparsity, limitingthe perception of autonomous driving systems. Scene completion meth-ods can infer the missing parts of incomplete single LiDAR scans. Re-cent methods adopt point-level denoising diffusion probabilistic modelsthat rely on approximations to handle scene-scale data, leading to amismatch between training and inference initial distributions. We pro-pose LiFlow, the first point-level flow matching method for 3D LiDARscene completion. LiFlow improves upon existing diffusion-based meth-ods by directly aligning single LiDAR scans to complete scenes, en-suring consistent initial distributions between training and inference.LiFlow introduces nearest neighbor flow matching and Chamfer distancematching to enhance both local structure and global coverage in thealignment of point clouds. Compared to existing diffusion-based meth-ods, LiFlow reduces the number of inference steps, enabling more ef-ficient scene completion while maintaining stability and high genera-tion quality. Extensive experiments demonstrate that LiFlow achievesstate-of-the-art performance across multiple metrics. Code is availableat https://github.com/matteandre/LiFlow.

Video understanding has become more and more importantwith the growth of Artificial Intelligence (AI) for video generation. Re-cently, Multimodal Large Language Model (M-LLM) has shown its ca-pability in video understanding. Video summarization, a specific domainof video understanding, has proven its importance for efficient naviga-tion and retrieval. Both video understanding and video summarizationrequire a good selection of key frames in a video. Current video summa-rization methods heavily focus on the selected key frames and correlatedsegment captions. However, existing approaches overlook the perspectiveof treating the importance of the frames globally. We argue that usingdiscrete selected frames for summarization will not only reduce the un-derstanding coherence, but also lost important information in the video,as well as wasting the original capacity of the MLLMs. In this paper, wepropose HAS, a Highlight-guided Attention Steering method for videosummarization. We consider a challenging but practical setting wherethe video given to MLLMs for summarize should be continuous but withhighlight guidance. HAS mainly consists of two parts: The first part isto find a continuous frame-level highlight distribution for the video glob-ally. The second part is to apply the highlight distribution as an attentionsteering vector for the MLLM, targeting a better understanding of thevideo, and thus during the model inference time, putting more attentionon the highlighted frames, while avoiding lost entire information on lesshighlighted frames through putting less attention instead of forgettingthem. We evaluated HAS on a variety of benchmarks, and it has shownconvincing performance in video summarization.

Point cloud segmentation is critical for 3D scene understand-ing. However, sparse and irregular point distributions provide limitedappearance evidence, making geometry-only features insufficient to dis-tinguish objects with similar shapes but distinct appearances (e.g., color,texture, and material). We propose Gaussian-to-Point (G2P), which trans-fers Gaussian attributes from 3D Gaussian Splatting to point clouds formore discriminative and appearance-consistent segmentation. Our G2Paddresses the misalignment between optimized Gaussians and originalpoint geometry by establishing point-wise correspondences. By distillingopacity-derived visibility cues, we mitigate the geometric ambiguity thatlimits existing models. Additionally, Gaussian scale attributes enableprecise boundary localization in complex 3D scenes. Extensive experi-ments demonstrate that our approach achieves competitive performanceon standard benchmarks and shows notable improvements on geomet-rically challenging classes, without pretrained 2D features or languagesupervision in our segmentation pipeline.

3D Gaussian Splatting (3DGS) achieves state-of-the-art rendering quality at real-time speeds but suffers from “model bloat”—a large number of redundant, low-opacity Gaussians that inflate memory usage and training costs. This inefficiency stems from the standard “densifythen-prune” paradigm, which expands the model aggressively before relying on pruning to achieve compactness. To mitigate this problem, we present an efficient training framework that builds an intrinsically compact representation, replacing the conventional densify-then-prune cycle. Our method leverages a synergistic design: an L reconstruction loss to 2 provide error-proportional gradients that stabilize optimization, and a novel Polarized Opacity Prior (POP) to actively manage the Gaussian population. POP steers informative primitives toward full opacity and uninformative ones toward transparency, enabling natural pruning and accelerating rendering through Early Ray Termination. Experiments on three public datasets demonstrate that our approach consistently achieves accelerated 3DGS training with significantly fewer Gaussians while maintaining comparable visual reconstruction quality. These results show that the proposed framework provides a simple and effective path toward fast and inherently compact 3DGS training.