ECCV 2026 Papers

We present VIGS-SLAM, a visual-inertial 3D Gaussian Splat-ting SLAM system that achieves robust real-time tracking and high-fidelity reconstruction. Although recent 3DGS-based SLAM methodsachieve dense and photorealistic mapping, their purely visual design de-grades under challenging conditions such as motion blur, low texture,and exposure variations. Our method tightly couples visual and inertialcues within a unified optimization framework, jointly optimizing cam-era poses, depths, and IMU states. It features robust IMU initialization,time-varying bias modeling, and loop closure with consistent Gaussianupdates. Experiments on five challenging datasets demonstrate our su-periority over state-of-the-art methods. The code will be made public.

Transformer-based video diffusion models rely on 3D attention over spatial and temporal tokens, which incurs quadratic time and memory complexity and makes end-to-end training for ultra-highresolution videos prohibitively expensive. To overcome this bottleneck, we propose a pure image adaptation framework that upgrades a video Diffusion Transformer pre-trained at its native scale to synthesize higherresolution videos. Unfortunately, naively fine-tuning with high-resolution images alone often introduces noticeable noise due to the image–video modality gap. To address this, we decouple the learning objective to separately handle modality alignment and spatial extrapolation. At the core of our approach is Relay LoRA, a two-stage adaptation strategy. In the first stage, the video diffusion model is adapted to the image domain using low-resolution images to bridge the modality gap. In the second stage, the model is further adapted with high-resolution images to acquire spatial extrapolation capability. During inference, only the high-resolution adaptation is retained to preserve the video generation modality while enabling high-resolution video synthesis. To enhance fine-grained detail synthesis, we further propose a High Frequency Awareness Training Objective, which explicitly encourages the model to recover high-frequency components from degraded latent representations via a dedicated reconstruction loss. Extensive experiments demonstrate that our method produces ultra-high-resolution videos with rich visual details without requiring any video training data, even outperforming previous state-ofthe-art models trained on high-resolution videos by 0.8 on the VBench benchmark. Code is available here.

Human motion modeling must support generation from text,music, and speech, comprehension through captioning, temporal com-pletion, and multi-person coordination—yet existing methods typicallytackle each scenario with a separate, task-specific system. We arguethat discrete tokenization plus autoregressive (AR) modeling possessesunique compositional advantages for building a unified motion model:all modalities share a single vocabulary and training objective, tasks aredefined by token arrangement rather than architectural changes, andthe same model naturally supports both generation and understand-ing by simply reordering the sequence. However, realizing this potentialhas been blocked by two obstacles. First, current motion tokenizers useroot-relative, non-causal, frame-level encodings that discard world-spaceinformation, break consistency with forward-only AR decoding, and scalepoorly—confining existing discrete AR models to text–motion pairs insingle-human settings. Second, multi-modal motion data with paired textand audio annotations remains scarce. We address both obstacles. Weintroduce VerMoVQ, a causal spatio-temporal VQ-VAE that factorizeseach frame into low-dimensional body tokens while preserving world-space root trajectories, naturally enabling multi-person tokenization viaper-agent concatenation and arbitrary-length encoding through stricttemporal causality. We curate MotionHub, a \sim 380K-clip corpus in unifiedSMPL representation with large-scale multi-modal annotations spanningnine motion tasks. Built on these foundations, we train VersatileMo-tion, a single decoder-only AR Transformer that covers text\leftrightarrow motion,music\leftrightarrow dance, speech

Despite the recent success of multimodal discrete diffusionlanguage models (multimodal dLLMs) in generating text through iter-ative demasking with bidirectional attention, object hallucination re-mains a critical challenge. Existing hallucination mitigation methods aredesigned for autoregressive generation and encounter fundamental mis-matches when applied to multimodal dLLMs, either doubling inferencecost through contrastive decoding or suffering from ambiguous step andposition alignment under bidirectional context. To address this, we pro-pose Unmasking-Time Visual Calibration (UVC), a lightweight, training-free framework that mitigates hallucinations natively within the demask-ing process. UVC extracts multi-granularity contrastive activation shiftvectors offline by comparing model responses to clean and visually de-graded inputs, identifies the most visually informative attention headsvia per-head AUC scoring, and injects the pre-computed calibration sig-nals exclusively at still-masked positions during inference, requiring noadditional forward passes. Extensive experiments on POPE and MMEwith two representative multimodal dLLMs demonstrate that UVC con-sistently reduces hallucinations by a substantial margin while introducingless than one percent inference overhead. Additional evaluation on open-ended captioning with the CHAIR metric further confirms that UVC sub-stantially lowers hallucination rates without affecting generation length,and generalizes effectively across benchmarks without retraining. Codeis available at https://github.com/THU-BPM/UVC.

Accurate motion prediction of surrounding agents and safemotion planning are two closely coupled key tasks for social robot nav-igation in crowded environments. Deploying these systems on resource-constrained edge devices necessitates compact, unified models that canperform both tasks simultaneously. However, within these compact sharedencoders, recent unified models often overlook severe representationalconflicts that arise from the distinct objectives of predicting neighbor be-haviors versus ego-centric safety planning. To address this issue, we firstidentify the Skill Conflict—a phenomenon where overlapping parameterassignments cause distinct tasks to compete for the same weights, pre-venting the model from fully specializing in individual skills. To resolvethis, we propose a novel model-merging-based framework, Disjoint Pa-rameter Training (DPT). DPT mitigates performance degradationcaused by Skill Conflict through distributed parameter learning, whichseparates the key parameter regions of each task while preserving theircore capabilities prior to merging. In addition, we observe that sparsemerging, which selectively integrates only the most influential param-eters for each task rather than combining all task-specific parameters,yields optimal performance by preventing interference among adjacentfeatures and concentrating representational capacity. DPT can be ap-plied in parallel with a variety of merging methods. Evaluated on stan-dard crowd navigation benchmarks (JRDB and JTA), our frameworkdemonstrates superior performance, validating its versatility and effec-tiveness for safe, resource-efficient robot navigation. The project page isavailable at: https://dpt2026.github.io/

Style transfer must match a target style while preservingcontent semantics. DiT-based diffusion models often suffer from con-tent–style entanglement, leading to reference-content leakage and unsta-ble generation. We present UniCSG, a unified framework for content-constrained, style-driven generation in both text-guided and reference-guided settings. UniCSG employs staged training: (i) a latent-space se-mantic disentanglement stage that combines low-frequency preprocess-ing with conditioning corruption to encourage content–style separation,and (ii) a latent-space frequency-aware detail reconstruction stage thatrefines details via multi-scale frequency supervision. We further incorpo-rate pixel-space reward learning to align latent objectives with percep-tual quality after decoding. Experiments demonstrate improved contentfaithfulness, style alignment, and robustness in both settings.

Pretraining 3D encoders through alignment with ContrastiveLanguage–Image Pre-training (CLIP) has emerged as a promising direc-tion to learn generalizable representations for 3D scene understanding. Inthis paper, we propose UniScene3D, a transformer-based framework thatlearns unified 3D scene representations from multi-view RGB–Pointmapby leveraging the priors of a pretrained 2D foundation model. For robustRGB-Pointmap representation learning, we introduce novel cross-viewgeometric alignment and grounded view alignment to enforce geome-try and semantic consistency across views. Extensive low-shot and task-specific fine-tuning across viewpoint grounding, scene retrieval, sceneclassification, and 3D visual question answering achieves state-of-the-artperformance. These results establish UniScene3D as an effective frame-work for unified 3D scene understanding.

Flow-based models learn a target distribution by modelinga marginal velocity field, defined as the average of sample-wise velocitiesconnecting each sample from a simple prior to the target data. However,when sample-wise velocities conflict at the same intermediate state, thisaveraged velocity can misguide samples toward low-density regions, de-grading generation quality. To address this issue, we propose the FlowDivergence Sampler (FDS), a training-free framework that refines in-termediate states before each solver step. Our key finding reveals thatthe severity of this misguidance is quantified by the divergence of themarginal velocity field that is readily computable during inference witha well-optimized model. FDS exploits this signal to steer states towardless ambiguous regions. As a plug-and-play framework compatible withstandard solvers and off-the-shelf flow backbones, FDS consistently im-proves fidelity across various generation tasks including text-to-imagesynthesis, and inverse problems.

We present DeWorldSG, a novel framework that generatesspatio-temporally robust 3D Semantic Scene Graphs from RGB-D se-quences. Existing methods often struggle to construct reliable 3D scenegraphs due to unstable 3D object representations and missing relationscaused by frame-wise inference. DeWorldSG addresses these issues byestimating instance-level geometric 3D Gaussian distributions throughdepth-guided filtering and representing each object as a probabilistic 3Dnode rather than a single projected point. To mitigate relational sparsityfrom frame-wise inference, our framework further aggregates spatiotem-poral evidence across object pairs and refines relations using contextualpriors derived from a world model (V-JEPA 2). Experiments on the3DSSG and ReplicaSSG datasets demonstrate state-of-the-art (SoTA)performance in both object and predicate prediction, while producingtemporally consistent scene structures. In particular, our method im-proves triplet recall by 77.4% and predicate recall by 23.2% over priorSoTA approaches, making it suitable for robotic manipulation and ARapplications. Our code and models are open-sourced.

Multimodal large language models (MLLMs) have demon-strated strong performance in offline video understanding tasks, but mostremain constrained to offline inference or exhibit weak online reasoningability, rendering online multi-turn interaction over continuously arriv-ing video streams challenging. Existing streaming approaches adopt aninterleaved perception-generation paradigm, which precludes concurrentperception and generation and induces early memory decay with grow-ing streams, degrading long-range dependency modeling. We proposeThink While Watching, a memory-anchored streaming video reason-ing framework that maintains continuous segment-level memory duringmulti-turn interaction. We construct a three-stage, multi-round, chain-of-thought (CoT) dataset with a stage-matched training strategy while en-forcing strict causality in streaming reasoning via a segment-level stream-ing causal mask and streaming positional encoding. At inference, we de-sign an efficient pipeline that overlaps watching and thinking and adap-tively selects the optimal attention backend. We evaluate our methodunder single-round and multi-round streaming input protocols. Basedon Qwen3-VL, we improve single-round accuracy by 2.6% on Streaming-Bench and 3.79% on OVO-Bench. In the multi-round protocol, we main-tain performance while reducing output tokens by 56%. Code is avail-able at: https://github.com/wanglu-cs/Think_While_Watching.

Autoregressive video diffusion models have enabled the gener-ation of arbitrarily long videos by removing conditioning on future frames,thus greatly improving computational efficiency. Yet, they suffer from er-ror accumulation over time, as the denoised sequence gradually drifts awayfrom the conditioning distribution seen during training. Recent advancesattempt to reduce this error by anchoring each generated frame to thelearned manifold of real ones. However, even when all generated individualframes lie close to the real manifold, there are trajectories which the modellacks sufficient knowledge to continue without exiting it, thus reachinga terminal point. To prevent the model from being trapped in terminalpoints, we start from the hypothesis that for well-modeled future trajec-tories the distribution of the predicted noise should match the one of theforward noising process. To enforce such a prior at test time, we introduceTerminal points Avoidance through Noise Guided Optimization (TANGO),which uses the diffusion model as a critic of its own outputs, by predictingone step forward and requiring an isotropic Gaussian noise prediction. Weuse the deviation from this expected noise distribution to search for an al-ternative trajectory that does not lead to a terminal point. Our approachachieves a 3.1% absolute improvement on VBench over state-of-the-art,while reducing Fréchet Video Distance by 28.3% on average across 15 svideos. Our code is available on https://mever-team.github.io/tango.

Accurate landslide segmentation from high-resolution remote sensing imagery remains challenging due to weak spectral contrast and strong dependence on terrain structure. Unlike appearance-defined objects, landslides follow geomorphological patterns governed by slope continuity and elevation discontinuities. However, most existing segmentation methods treat landslides as generic semantic regions and incorporate Digital Elevation Models (DEM) through simple feature fusion without explicitly enforcing terrain-consistent spatial reasoning. To address this limitation, we reformulate landslide segmentation as a terrainconditioned structured prediction problem and propose TerrainGraphNet, a terrain-aware graph reasoning framework that embeds geomorphological constraints into representation learning and spatial propagation. The proposed method introduces terrain-modulated feature interaction, where elevation structure adaptively regulates visual representations, and terrain-aware graph construction, where spatial connectivity is defined jointly by feature similarity and slope continuity. This formulation can be interpreted as learning a terrain-weighted smoothness prior that encourages predictions to respect geomorphological coherence. Extensive experiments on three benchmark datasets demonstrate consistent improvements over strong CNN and transformer baselines. In addition to higher IoU and F1 scores, TerrainGraphNet significantly improves boundary accuracy and topological consistency, highlighting the importance of terrain-constrained reasoning for reliable landslide delineation.

Pretrained Vision Transformers (ViTs) such as DINOv2 andMAE provide generic image features that can be applied to a varietyof downstream tasks such as retrieval, classification, and segmentation.However, such representations tend to focus on the most salient visualcues in the image, with no way to direct them toward less prominentconcepts of interest. In contrast, Multimodal LLMs can be guided withtextual prompts, but the resulting representations tend to be language-centric and lose their effectiveness for generic visual tasks. To addressthis, we introduce Steerable Visual Representations, a new class of vi-sual representations, whose global and local features can be steered withnatural language. While most vision-language models (e.g., CLIP) fusetext with visual features after encoding (late fusion), we inject text di-rectly into the layers of the visual encoder (early fusion) via lightweightcross-attention. We introduce benchmarks for measuring representationalsteerability, and demonstrate that our steerable visual features can focuson any desired object in an image while preserving the underlying rep-resentation quality. Our method also matches or outperforms dedicatedapproaches on anomaly detection and personalized object discrimination,exhibiting zero-shot generalization to out-of-distribution tasks.Project Website: jonaruthardt.github.io/project/SteerViTPrompt CLS AttentionSteerable Visual RepresentationsGoal: control what vision features encode New Pareto FrontierEval: retrieving images w/ prompted objects Previous SoTAUse: task-specific adaptationFig. 2: SteerViT produces high-quality visual representations that can besteered by text. Left: Traditional (non-steerable) representations like DINOv2 tendto focus on the dominant object in an image and retrieve images with the same object.SteerViT can adapt to a text prompt, enabling retrieval of images even with smallobjects of interest. Right: We compare SteerViT to prior work in terms of its ability toadapt to text (measured by text-guided image retrieval (cf. Sec. 4.1)) and the qualityof the visual representation (measured by the accuracy of linear probing for the CLSfeature and semantic segmentation for patch features). While models typically tradeoff steerability for representation quality, SteerViT preserves both. By modulating agating factor (Eq. (2)), SteerViT achieves a new Pareto frontier.

Cross-attention is a crucial component in learning-based imageto-point-cloud (I2P) registration. Although existing cross-attention mechanisms have achieved promising progress, attention ambiguity remains a fundamental challenge that hinders the learning of discriminative 2D- 3D correspondences. To address this problem, we revisit cross-attention and establish ordinary differential equations (ODEs) to model the ideal I2P feature interaction. Based on this formulation, we develop an ODEdriven cross-attention (OCA) module that refines feature representations and attention matrices through ODEs. In practice, OCA can be seamlessly integrated into existing I2P registration frameworks. To validate its effectiveness, we incorporate OCA into five state-of-the-art baselines and evaluate on four public benchmark datasets. Experimental results demonstrate that OCA improves registration recall by up to 5%, 9%, and 15% under the standard, fine-tuning, and zero-shot settings, respectively. Code is released at github.com/anpei96/oca-i2p-demo.

Tokenizers are a key component of state-of-the-art generativeimage models, extracting the most important features from the signalwhile reducing data dimension and redundancy. Most current imagetokenizers are based on KL-regularized variational autoencoders (KL-VAE), trained with reconstruction, perceptual and adversarial losses.Diffusion decoders have been proposed as a more principled alternativeto model the distribution over images conditioned on the latent. However,matching the performance of KL-VAE still required adversarial losses,as well as a higher decoding time due to iterative sampling. To addressthese limitations, we introduce a new pixel diffusion decoder architecturefor improved scaling and training stability, benefiting from transformercomponents and GAN-free training. We use distillation to replicate theperformance of the diffusion decoder in an efficient single-step decoder.This makes SSDD the first GAN-free single-step diffusion decoder toreach state-of-the-art continuous tokenization, with higher reconstructionquality and faster sampling than KL-VAE. In particular, SSDD improvesreconstruction FID from 0.87 to 0.46 with 1.4× higher throughput andpreserves generation quality of DiTs with 3.8× faster sampling. As such,SSDD can be used as a drop-in replacement for KL-VAE, and for buildinghigher-quality and faster generative models.

Drone image–text retrieval (DITR) aims to associate aerialimages with textual descriptions in complex low-altitude environments.However, intricate spatial layouts and long-range structural dependen-cies in drone imagery challenge conventional Euclidean embedding mod-els, often leading to cross-modal misalignment. To address this issue, wepropose SARA, a structure-aware Riemannian-guided alignment frame-work that incorporates Riemannian manifold modeling to complementsemantic embeddings. Specifically, the Manifold Structural Feature Ex-traction (MSFE) module captures second-order structural dependenciesby encoding local feature correlations as symmetric positive definite rep-resentations. The Log-Euclidean Structural Alignment (LESA) moduleperforms geometry-consistent alignment in the tangent space, integratingsemantic correspondence with structural coherence through distributionalignment and feature fusion. We further provide theoretical analysisdemonstrating convergence and minimization of an upper bound on theexpected alignment error. Experiments on ERA and UDV benchmarksdemonstrate that SARA consistently improves retrieval performance overstate-of-the-art methods, particularly in structurally complex or seman-tically ambiguous drone scenarios.

Existing methods for adapting 2D foundation models such asSAM to 3D volumes either process slices independently—ignoring inter-slice context—or require substantial architectural changes and retraining.In this paper, we present SAM+D, a parameter-efficient framework thatlifts SAM-family models by one spatial dimension—enabling 3D volu-metric segmentation from 2D SAM and, for the first time via parameter-efficient fine-tuning, end-to-end 4D (3D+T) spatiotemporal segmenta-tion from video-based SAM2—while keeping the vast majority of pre-trained parameters frozen. SAM+D introduces two lightweight, model-agnostic modules into frozen transformer blocks: (1) Depth-RoutedLoRA (DRLoRA) experts with learned routing for spatially adaptivelow-rank updates, and (2) Depth Shift Modules (DSM) for cross-slice feature exchange at zero additional parameter cost. Together, theyprovide volume-level context while tuning only ∼2.8% of parameters forSAM and ∼3.7% for SAM2. We evaluate SAM+D in two distinct set-tings, each lifting the base model by one spatial dimension: 3D segmenta-tion, where SAM (2D → 3D) is evaluated on four CT benchmarks (KiTS,Pancreas, LiTS, Colon), and 4D segmentation, where SAM2 (2D+T →3D+T) is evaluated on a cell tracking challenge (CTC) dataset (Fluo-N3DH-SIM+). In both settings SAM+D achieves competitive or superiorresults under the single-point prompt setting while using fewer trainableparameters than existing methods, demonstrating that SAM+D general-izes across SAM-family architectures, target dimensionalities (3D, 4D),and domains spanning medical imaging and bio-scene understanding.Code is publicly available at https://github.com/JerrySongCST/SAM-Plus-D.

Micro-expressions are brief, involuntary facial movementsthat reveal subtle affective states. While most computational work fo-cuses on micro-expression recognition, which classifies clips into dis-crete emotion categories, Micro-Expression Action Unit (ME-AU) de-tection offers a more theory-consistent formulation grounded in the Fa-cial Action Coding System. Despite its finer granularity, ME-AU de-tection remains comparatively under-explored due to the limited avail-ability of high-quality data for efficient automatic models. Recent volu-metric micro-expression datasets provide temporally coherent 3D facialrecordings with reliable AU labels, creating new opportunities for fine-grained analysis. However, their adoption remains limited due to largedata volumes, heterogeneous formats, and the absence of standardizedbenchmarks, which together hinder reproducibility and cross-study com-parison. Moreover, because of the small spatio-temporal spanning of theAUs, designing models that effectively combine volumetric representa-tions with temporal dynamics remains a non-trivial research challenge.To address these barriers, we introduce an extended and standardizedrelease of the 4DME dataset with cleaned data and normalized format-ting, accompanied by a compressed version to reduce storage overhead.We further design VoluME, a family of dual-stream optical-flow-basedarchitecture for volumetric ME-AU detection. To determine appropriatesettings for compression, we combine state-of-the-art compression base-lines, perceptual evaluation, and downstream task-specific benchmark-ing. As the first work dealing with volumetric ME-AU detection undercompression, our results show that a compression ratio of approximately3000× can be achieved with ≈ 1% F1 degradation. The extended releaseadditionally introduces explicit two-way cultural balancing, supportingfuture cross-cultural micro-expression studies.

We address the problem of discovering repeated elementsfrom a single image. In contrast to existing approaches that depend onlarge annotated datasets, curated multi-image collections, or object seg-mentation masks, we show that a single image can suffice to learn ameaningful object model in a completely bottom-up fashion, without anyprior knowledge beyond a coarse scale prior. Our method learns a tunableimage-space prototype of the repeated elements through a reconstructionobjective, enabling the model to identify and synthesize consistent objectinstances within the same image. Experiments on 116 real images fromthe FSC-147 dataset demonstrate that our method successfully learnscoherent element models and captures intra-category variation on chal-lenging images. Qualitative results reveal superior reconstructions andinterpretable decompositions compared to classical decomposition, jointalignment, and 3D object modeling methods, while maintaining a simple2D formulation. These results suggest that meaningful object discoverycan emerge from single-image learning alone.