ECCV 2026 论文列表

While knowledge distillation (KD) is widely adopted for train-ing lightweight models by leveraging supervision from larger teachermodels, relying solely on output token distributions has proven insuf-ficient for compressing Multimodal Large Language Models (MLLMs).Since output tokens are a byproduct of the model attending to visualinputs, prior works have explored explicitly distilling attention to pro-vide a direct supervisory signal. While promising, the precise utility ofwhich attention signals to distill remains under-explored. In this work, wechallenge the conventional reliance on prompt-to-vision attention by re-vealing that downstream performance correlates strongly with response-to-vision attention similarity to the teacher, but negligibly with that ofprompt-conditioned attention. Furthermore, we observe that attentiondistributions exhibit significant variance across individual tokens, indi-cating that a uniform distillation objective is suboptimal. To this end,we introduce Token-level Response-visual Attention Guidance(TRAG), a distillation objective that 1) shifts the focus to response-to-vision signals and 2) employs token-specific objectives by adaptivelyweighting the Kullback-Leibler divergence based on attention entropy, ef-fectively guiding the student to mirror the teacher’s precise visual focus.Extensive experimental results on multiple benchmarks demonstrate thatTRAG significantly outperforms prior distillation baselines. Our code isavailable at https://github.com/jhjangjh/TRAG.

We present PointGT, a point-based 3D representation thatenables simultaneous editing of object geometry and appearance. Ex-isting reconstruction and view synthesis techniques produce volumetric3D representations that are high-quality and photorealistic, but are dif-ficult to edit. In particular, recent efforts to enable texture editing for3D Gaussian Splatting representations are not compatible with geometryedits and deformations. Our method combines a point-based represen-tation that is well-suited for geometry deformations with a learned UVmapping technique that enables high-resolution texture editing. We showthat PointGT enables fine-grained editing of both geometry and texturein point-based neural representations with high rendering quality.

Sign language translation (SLT) remains challenging due to its high spatio-temporal complexity, long sequences, and the need to model multiple articulators without relying on gloss annotations. Existing approaches are typically tailored to individual datasets or languages and struggle to scale, while overlooking the relationships between sign languages that could inform more effective cross-lingual transfer. We present SIGNET, a framework that enables motion-level knowledge transfer for cross-language sign language translation. Our key insight is that, although sign languages differ in grammar and lexicon, pretrained models capture motion-level visual patterns that can be reused across datasets and languages. SIGNET integrates multiple pretrained sign language backbones through an attention-based, hand-prior aggregation mechanism that guides a gated fusion network in dynamically selecting the most relevant experts. Comprehensive experiments on four benchmarks (How2Sign, Phoenix14T, CSL-Daily, and MeineDGS) demonstrate state-of-the-art translation performance, and SIGNET also surpasses prior methods on WLASL for sign language recognition.

Large vision–language models (LVLMs) incur substantial com-putational and memory overhead due to quadratic attention complexityand enlarged key–value (KV) caches. Training-free vision token prun-ing is promising, yet most methods rely on text-to-vision attention-scoremagnitudes, which often entangle instruction-relevant signals with po-sitional bias and visual attention sinks. To address this issue, we pro-pose DIVA, an instruction-aware, layer-wise vision token pruning methodbased on dual-probe attention discrepancy, with two key modules: DIVA-Pruner and VF-Calibrator. DIVA-Pruner selects important tokens us-ing the discrepancy between attention distributions of contextual andnon-contextual probe sets. A lightweight VF-Calibrator calibrates unim-portant tokens that bypass self-attention, mitigating representation mis-match and enabling effective reuse in subsequent LVLM blocks. On di-verse benchmarks and LVLM baselines, DIVA consistently outperformsprior methods across token budgets, even in highly constrained regimes.Notably, under a 97.2% token reduction, DIVA retains 92.6% of baselineperformance on LLaVA-NeXT-7B with a 10.8× reduction in KV cache.

Recent advances in video diffusion models have delivered re-markable generation quality, yet their generation cost remains high. Thisinefficiency is largely driven by the iterative denoising process, with com-putationally expensive per-step architectures. Step distillation enablesfew-step generation, but per-step latency remains high because a large-scale backbone is invoked at each denoising step. Several works seek toreduce per-step latency via block pruning using importance scores aver-aged across denoising steps. However, such averaged scores can obscurestage-critical blocks and bias the pruning process, removing blocks essen-tial for certain stages while leaving redundancy in others. In this paper,we propose a novel pruning paradigm, Stage-Specialized Block Pruning,that partitions the denoising process into multiple stages and assigns atailored pruned variant to each stage. Each variant consists of sharedblocks across multiple stages and stage-specific blocks for its assignedstage. With parameter sharing, the overall parameter budget is smallerthan the original model, while the per-step computational footprint is re-duced for lower latency. To derive stage-specialized variants, we introducea hierarchical selection mechanism consisting of an Inter-Stage VarianceSelector (IVS) and an Intra-Stage Quality Selector (IQS). IVS comparesthe variance of stage-wise block-importance distributions and prioritizesthe stage with higher variance for pruning, indicating greater removableredundancy. IQS mitigates noisy importance estimation by generatingmultiple low-importance pruning candidates within the selected stageand choosing the one with the highest video generation quality. IVS andIQS are applied iteratively until reaching a target latency budget. Fur-thermore, we introduce a two-phase training pipeline to recover the per-formance of these variants and facilitate sequential composition for few-step generation. Extensive experiments on the representative video diffu-sion model Wan2.1-1.3B demonstrate that our method reduces per-stepinference cost while maintaining competitive performance on VBench.

Motion generation for rigged shapes is vital for scalable 4D asset production. However, template-based methods are limited by specific topologies and fail to generalize across diverse morphologies. Conversely, per-case optimization is computationally expensive, susceptible to local optima, and highly sensitive to viewpoint-induced ambiguities. In this paper, we present SkelMo, a diffusion-based framework designed for category-agnostic skeletal animation generation from 2D video guidance. To overcome the scarcity of high-quality training data, we have curated a large-scale dynamic dataset comprising approximately 25,200 diverse 3D animations, each featuring complete textures, skeletal rigging, and a wide array of comprehensive animation sequences. To bridge the kinematic gap between 2D visual motion cues and heterogeneous 3D skeletal structures, we propose a structural-semantic injection mechanism. Our model integrates texture and semantic attributes directly into skeletal joint representations. This allows it to map perceived visual dynamics to specific joint hierarchies and their functional roles. This enables SkelMo to synthesize high-fidelity animations that maintain anatomical consistency across a vast range of unseen categories, from existing biological species to fantastical beings. Extensive experiments demonstrate that our approach significantly outperforms existing methods, setting a new state-of-the-art benchmark for robust and efficient 4D asset generation. Project Page: https://research.davytao.me/skelmo/.

Diffusion-based generative models have transformed visualcontent synthesis, yet they remain vulnerable to unauthorized usage andlack reliable attribution methods. Existing watermarking techniques of-ten treat latent tensors as static spatial feature maps or depend on pixel-domain modification, and most do not explicitly leverage the internalfrequency structure of the latent space for dual-band redundant embed-ding, leaving them susceptible to the stochastic nature of diffusion andregeneration attacks. We introduce BiSLW, a trainable bi-spectral la-tent watermarking framework that jointly embeds aligned identity signalsacross complementary spectral bands of the decoded diffusion latent us-ing learned encoders and decoders, going beyond fixed-pattern frequencyapproaches. We leverage the inherent frequency structure of diffusionlatents to design a dual-band watermarking framework. Low-frequencycomponents encode global semantics, while high-frequency componentscapture fine texture. We exploit this structure to embed watermarksacross complementary spectral bands. The watermark is independentlyinjected into both bands via learned encoders and recombined before de-coding, ensuring it becomes intrinsic to the generative trajectory. Dualspectral decoders recover the watermark from each band, while a cross-band consistency constraint enforces alignment between semantic andtextural embeddings. Experiments show that BiSLW achieves a strongbalance between perceptual fidelity and robustness, improving PSNR byover 3 dB compared to prior latent diffusion watermarking methods whilepreserving near-perfect bit accuracy under aggressive regeneration andcommon distortions, all with negligible computational overhead.

Automated segmentation of cryo-electron tomograms rou-tinely produces masks that are voxel-accurate but topologically broken:membranes fragment, organelles merge into one another and enclosedcavities collapse. Existing topology-aware losses reduce these violationsbut cannot eliminate them because topology is encouraged through gradi-ent pressure rather than structurally enforced. We introduce TopoFuse,which reframes topology as a differentiable projection operator rather thana loss penalty. At each forward pass, the projection operator ProjT (a PH-guided sparse edit) identifies the critical voxels responsible for topologicalviolations via bottleneck matching and applies sparse edits to satisfy aspecified topology target (diagram feature counts and lifetime budgets)for dimensions d ∈ {0, 2}. If the projection converges, the output satisfiesthose constraints on the downsampled grid (s = 2), when it does not, arepair certificate exposes this explicitly enabling downstream filtering.A topology prior head predicts the correction target directly from inputfeatures, removing any dependence on ground-truth topology at inference.Across three cryo-ET benchmarks, TopoFuse reduces Betti Number Errorby 54% over the strongest soft-loss baseline (p < 0.001), improves Diceby 4.6 pp and edits only 3.1% of voxels to achieve this. Code and SYNdataset are available at: https://github.com/rohitsalla/TopoFuse

Active Speaker Detection determines whether a visible per-son in a video is speaking at each moment. While recent audio–visualfusion methods perform well on clean data, they degrade under real-world corruptions such as background noise, occlusion, or simultaneousmodality degradation. We attribute this limitation to the absence of ex-plicit consistency constraints that promote robust, semantically alignedrepresentations across modalities. Without such guidance, models tendto learn fragile modality-specific shortcuts that fail under corrupted con-ditions. We propose C 3 ASD, a multi-level consistency-driven frameworkwith three complementary constraints: embedding-level inter-modalityconsistency aligns audio-visual representations during speech; sequence-level intra-modality consistency separates speaking and non-speakingclusters via track-aware contrastive learning; and prediction-level con-sistency stabilizes fusion through knowledge distillation. Extensive ex-periments demonstrate significant improvements under diverse audio, vi-sual and joint corruptions, while maintaining competitive performanceon clean data.

As an important 3D sensing method, photometric stereo estimates surface normals from multiple photos taken under varying lighting. The surface normal quality highly relies on large number of input images, which requires long capturing time, high data burden and limits high-speed applications for photometric stereo. Recently, event-based photometric stereo has emerged as an efficient and rapid approach, but the low resolution of event sensors and the noise of event signals limit its high-quality application. In this paper, we propose DualResPS, a novel photometric stereo pipeline utilizing a frame-event hybrid camera. We design a lighting and capturing strategy tailored to a hybrid-camera setup to utilize observations of different spatial and temporal resolutions. The algorithm explicitly models non-ideal factors including shadow effects and specular reflection to achieve high-resolution, high-quality normal reconstruction. Our proposed DualResPS is validated on both semi-real datasets from the DiLiGenT, DiLiGenT-Pi, and real captured data. The results demonstrate that DualResPS surpasses its frame-based counterpart with 18.1% data bandwidth.

Connector-based video unified models have demonstratedstrong capability in instruction-grounded video synthesis, but integratinga large high-fidelity generator into the unified training loop is computa-tionally prohibitive, limiting achievable visual quality. We therefore pro-pose Lumos-Nexus, a training-efficient unified video generation frame-work that facilitates the development of strong reasoning-driven gen-eration capabilities while significantly enhancing visual fidelity. Lumos-Nexus adopts a two-stage design: 1) During training, only a lightweightgenerator is aligned with the understanding block to learn to take inreasoning-driven semantic control. 2) During inference, we introduceUnified Progressive Frequency Bridging (UPFB) to progressively handoff generation to a high-capacity pretrained generator in the shared la-tent space, enabling coarse-to-fine refinement and producing high-fidelityvideos without compromising reasoning quality. To fill the gap in reasoning-driven video generation benchmarks, we introduce VR-Bench, which as-sesses a model’s capability to translate inferred intent into coherent andsemantically aligned video content. Extensive experiments demonstratethat Lumos-Nexus achieves substantial gains in visual realism and tem-poral coherence on VBench, while exhibiting strong reasoning-based gen-erative performance on VR-Bench.

Vision-language models (VLMs) excel at many tasks, yetcontinue to struggle with spatial reasoning, problems where the key in-formation is not directly observable in the input. Many spatial questionsrequire imaginative perception: simulating an unseen viewpoint, tracinga trajectory through an occluded space, or integrating partial views intoa coherent spatial map. Humans naturally support this kind of reason-ing through imagination. Prior work has introduced intermediate visualrepresentations (e.g., visual thoughts, depth, or box tokens), but theseintermediates often refine structure already visible rather than predict-ing the missing spatial structure implied by the evidence. We introduceImaginative Perception Tokens (IPT), intermediate perceptual rep-resentations that externalize what a VLM would perceive under an alter-native spatial configuration while remaining consistent with the observedinput. To study this capability, we formulate three tasks that requireimaginative perception: Perspective Taking (PET), Path Tracing(PT), and Multiview Counting (MVC). For each task, we constructdatasets of →20K examples spanning simulated and real-world settings,paired with ground-truth intermediate imaginations, final answers, andcurated evaluation benchmarks. Using the unified VLM BAGEL [12] asour backbone, IPT supervision improves spatial reasoning across severalsettings and often outperforms textual chain-of-thought training, evenwhen no image is generated at inference time. For example, on MVC,IPT improves accuracy by 3.4% and achieves performance competitivewith strong closed-source models on Path Tracing. We also find thatmixed training with IPT and label-only data can further improve perfor-mance. In contrast, textual chain-of-thought can be detrimental on thesetasks, substantially degrading performance in some cases, highlighting amodality mismatch when forcing spatial computation through language.Overall, IPT provides a principled supervision signal for reasoning overunobserved structure, yielding stronger spatial generalization and a moreinterpretable intermediate aligned with the underlying geometry of thetask. Code will be released at the project page.Path Tracing (11k)Question: As you move from waypoint 1 to 2. This is what you see looking forward from Question: Which object can you seepoint 1: <image_2>. This is what you see looking forward from point 2: <image_3>. on your right side at waypoint M?Which object can you see on your left side at waypoint M?Options: Imagination1 M 2A. A dining table 2 22B. A fridge MC. A countertop 1D. A sofa 1 13Perspective Taking (55k)Question: After moving to the marked ‘X’ position on the ground while keeping my Question: Will the sideboard appearcurrent orientation, then turning left for 90 degrees, will the sideboard appear closer closer or further away?or further away?Options:A. CloserB. Further XX1 ImaginationMultiview Counting (17k)Question: What is the total number of office chairs in this scene? Question: How many refrigerator(s)are in this area?Answers: 8 Imagination1 23 4Fig. 1: Overview of the three spatial imagination tasks. The left columns showtraining examples with ground-truth imaginative perception; the right columns showevaluation examples.

3D representations are fundamental for scene rendering, un-derstanding, and interaction. Approaches like Neural Radiation Fieldsand 3D Gaussian Splatting achieve impressive photorealistic novel-viewsynthesis, but lack the ability to easily decompose scene elements intoobject level, requiring additional segmentation or grouping. We presentMLP-Splatting, a method that enables scene decomposition via a fewexpressive light-field primitives while providing photorealistic novel-viewsynthesis.MLP-Splatting models each primitive as an independent compact MLPwith localized spatial support that predicts radiance and opacity. In con-trast to low-level Gaussian primitives or a single global radiance field,our neural primitives provide greater expressive capacity while remain-ing spatially localized. Rendering is performed through efficient sparsevolumetric compositing over ray–primitive interactions.Our primitives are supervised using RGB supervision alone, which yieldsprimitives that represent local scene regions often corresponding to ob-jects or object parts, enabling interactive object-level editing withoutsegmentation masks by selecting a handful of primitives. Our method,augmented with optional semantic feature distillation, enables open-vocabulary scene interaction and open-set instant segmentation. Com-pared to state-of-the-art semantic 3DGS methods, we achieve substan-tially lower memory usage (1/7×) and faster rendering (5×).Project Page: https://shinjeongkim.com/mlp-splatting

We introduce Synthetic Visual Genome 2 (SVG2), alarge-scale panoptic video scene graph dataset. SVG2 contains over 636Kvideos with 6.6M objects, 52.0M attributes, and 6.7M relations, pro-viding an order-of-magnitude increase in scale and diversity over priorspatio-temporal scene graph datasets. To create SVG2, we design a fullyautomated pipeline that combines multi-scale panoptic segmentation,online–offline trajectory tracking with automatic new-object discovery,per-trajectory semantic parsing, and GPT-5-based spatio-temporal re-lation inference. Human verification of SVG2 annotation accuracy con-firms its reliability (objects: 93.8%, attributes: 88.3%, relations: 85.4%).Building on this resource, we train TraSeR, a trajectory-groundedvideo scene graph generation model. TraSeR augments VLMs with atrajectory-aligned token arrangement mechanism and new modules: anobject-trajectory resampler and a temporal-window resampler to convertraw videos and panoptic trajectories into compact spatio-temporal scenegraphs in a single forward pass. The temporal-window resampler bindsvisual tokens to short trajectory segments to preserve local motion andtemporal semantics, while the object-trajectory resampler aggregates en-tire trajectories to maintain global context for objects. On PVSG, VIPSeg,VidOR, and SVG2test , TraSeR outperforms the strongest open-sourcebaselines by 15∼20 points in relation detection, 20∼40 points in objectprediction, and 13 points in attribute prediction. It also surpasses GPT-5by 13 points in object prediction and 3 points in attribute prediction.When TraSeR’s generated scene graphs are sent to a VLM for videoquestion answering, it delivers a +1.5∼4.6 absolute accuracy gain overusing video alone or video augmented with Qwen2.5-VL’s generated scenegraphs, demonstrating the utility of explicit spatio-temporal scene graphsas an intermediate representation5 .

While humans naturally gesture during speech, only a sparse subset of these co-speech gestures are visually depictive and semantically linked to specific spoken words. In this paper, we introduce a large-scale dataset – Gesture Recognition in the Wild (GRW), comprising co-speech gestures corresponding to a diverse vocabulary of 155 words. GRW contains 140k manually annotated video clips where the word is spoken, with 17k instances of semantic co-speech gestures including their frame-level temporal boundaries. The video clips are collected ‘in the wild’ from public-facing discourse, including lectures, talk shows, and interviews, covering a diverse range of speakers and visual conditions. We also introduce video models to: (a) classify gestures as semantic or not; (b) recognize the word corresponding to a co-speech gesture; and (c) temporally localize the gesture. These models are trained and evaluated on the GRW dataset and compared against a range of strong baselines, establishing benchmark results for all three tasks. The dataset, annotations, and trained models are publicly available on the project website.

We introduce the Large Sparse Reconstruction Model tostudy how scaling transformer context windows affects feed-forward 3Dreconstruction. Although recent object-centric feed-forward methods pro-duce robust, high-quality reconstructions, they still lag behind dense-view optimization in recovering fine-grained texture and appearance. Weshow that expanding the context window—by substantially increasingthe number of active object and image tokens—narrows this gap andenables high-fidelity 3D object reconstruction and inverse rendering. Toscale effectively, we adapt native sparse attention [68] for 3D reconstruc-tion with three key contributions: (1) an efficient coarse-to-fine pipelinethat focuses computation on informative regions by predicting sparsehigh-resolution residuals; (2) a 3D-aware spatial routing mechanism thatestablishes accurate 2D-3D correspondences using explicit geometric dis-tances rather than standard attention scores; and (3) a custom block-aware sequence-parallel strategy with an All-gather-KV protocol to bal-ance dynamic, sparse workloads across GPUs. As a result, LSRM handles20× more object tokens and >2× more image tokens than prior state-of-the-art (SOTA) methods. Extensive evaluations on standard novel-viewsynthesis benchmarks show substantial gains over the current SOTA,yielding >2.4 dB higher PSNR and >40% lower LPIPS. Furthermore,when extending LSRM to inverse rendering, qualitative and quantitativeevaluations on widely used benchmarks demonstrate consistent improve-ments in texture and geometry details, achieving an LPIPS that matchesor exceeds that of SOTA dense-view optimization methods. Code andmodel weights are available on our project page.

Diffusion models have shown strong potential for multi-modalplanning in end-to-end autonomous driving. However, most existing meth-ods confine diffusion to the planning module, conditioning on fixed out-puts from separate discriminative perception networks. This decoupleddesign propagates perception errors to the planner, increasing optimiza-tion difficulty and reducing robustness. To overcome these limitations, wepropose UniTeD, a Unified Temporal Diffusion framework that jointlymodels perception and planning through iterative denoising in a sharedgenerative space. By enabling bidirectional information exchange, theframework facilitates mutual refinement between tasks and improves ro-bustness via noise-conditioned multi-task training. We further extendthis unified diffusion paradigm to a streaming setting by incorporat-ing temporal context. A Temporal Transition Module (TTM) is intro-duced to resolve the noise-level mismatch between historical and currentframes. In addition, we propose an Anchor Refresh Strategy (ARS) toalleviate the training–inference distribution shift commonly observed insparse diffusion-based end-to-end driving frameworks. Without bells andwhistles, UniTeD achieves state-of-the-art performance across multiplebenchmarks, surpassing both recent discriminative end-to-end methodsand diffusion-based planning approaches.

Situated reasoning often relies on active exploration, yet inmany real-world scenarios such exploration is infeasible due to phys-ical constraints of robots or safety concerns of visually impaired users.Given only a limited observation, can an agent mentally simulate a futuretrajectory toward a target situation and answer spatial “what-if” ques-tions? We introduce WanderDream, the first large-scale dataset designedfor the emulative simulation of mental exploration, enabling models toreason without active exploration. WanderDream-Gen comprises 15.8Kpanoramic videos across 1,088 real scenes from HM3D, ScanNet++,and real-world captures, depicting imagined trajectories from currentviewpoints to target situations. WanderDream-QA contains 158K ques-tion–answer pairs, covering starting states, paths, and end states alongeach trajectory to comprehensively evaluate exploration-based reason-ing. Extensive experiments with world models and MLLMs demonstrate(1) that mental exploration is essential for situated reasoning, (2) thatworld models achieve compelling performance on WanderDream-Gen, (3)that imagination substantially facilitates reasoning on WanderDream-QA, and (4) that WanderDream data exhibit remarkable transferabilityto real-world scenarios. The source code and all data are released athttps://github.com/RuipingL/WanderDream.

Deep neural networks remain vulnerable to imperceptible ad-versarial perturbations, yet efficient robustness certification for seman-tic segmentation is largely unexplored. We propose the first real-timecompatible certifiably robust semantic segmentation framework basedon networks with built-in Lipschitz constraints. Our method achievescompetitive pixel accuracy on Cityscapes while providing determinis-tic robustness guarantees, without the computational overhead of ran-domized smoothing or formal verification. We introduce a generalizedcertification framework for segmentation that computes worst-case per-formance under ℓ2 attacks of radius ϵ across diverse evaluation metrics.Crucially, our approach is approximately 600× faster than randomizedsmoothing at inference on an NVIDIA A100 GPU, delivering comparablecertificates. We validate our worst-case bounds against state-of-the-artadversarial attacks, demonstrating that Lipschitz-constrained architec-tures offer a practical and scalable path toward trustworthy dense pre-diction in safety-critical applications.