ECCV 2026 Papers

3D Gaussian Splatting (3DGS) has emerged as a powerfulparadigm for real-time and high-fidelity 3D reconstruction from posedimages. However, recent studies reveal its vulnerability to adversarialcorruptions in input views, where imperceptible yet consistent pertur-bations can drastically degrade rendering quality, increase training andrendering time, and inflate memory usage, even leading to server denial-of-service. In our work, to mitigate this issue, we begin by analyzingthe distinct behaviors of adversarial perturbations in the low- and high-frequency components of input images using wavelet transforms. Basedon this observation, we design a simple yet effective frequency-aware de-fense strategy that reconstructs training views by filtering high-frequencynoise while preserving low-frequency content. This approach effectivelysuppresses adversarial artifacts while maintaining the authenticity ofthe original scene. Notably, it does not significantly impair training onclean data, achieving a desirable trade-off between robustness and per-formance on clean inputs. Through extensive experiments under a widerange of attack intensities on multiple benchmarks, we demonstrate thatour method substantially enhances the robustness of 3DGS without ac-cess to clean ground-truth supervision. By highlighting and addressingthe overlooked vulnerabilities of 3D Gaussian Splatting, our work pavesthe way for more robust and secure 3D reconstructions.

Open-world semi-supervised learning (OWSSL) aims to rec-ognize known and unknown classes from unlabeled samples, where la-beled samples only include known classes. Recent studies introduce class-specific textual descriptions as semantic information and employ learn-able textual prompts to optimize these descriptions, thereby enhancingthe model’s ability to recognize classes. However, these methods oftenoverlook the model bias toward known classes during textual prompttuning, and establish only limited cross-modal connections, leading tosevere sparsity in the graph structure and hindering the propagation ofvisual and semantic information across the graph. These issues collec-tively limit the model’s ability to discover unknown classes, particularlyin scenarios with a large number of classes. To address these issues, wepropose a novel OWSSL method, which consists of two core components:(1) textual prompt optimization aimed at mitigating the model bias to-ward known classes, and (2) bimodal graph enhancement designed toalleviate the sparsity issue in the graph structure. Experimental resultsacross multiple datasets indicate that our method achieves an averageaccuracy 20.7% higher on unknown classes than state-of-the-art OWSSLmethods (relative improvement).

Long-trajectory video generation is a crucial yet challeng-ing task for world modeling, primarily due to the limited scalability ofexisting video diffusion models (VDMs). Autoregressive models, whileoffering infinite rollout, suffer from visual drift and poor controllability.To address these issues, we propose DCARL, a novel divide-and-conquer,autoregressive framework that effectively combines the structural stabil-ity of the divide-and-conquer scheme with the high-fidelity generationof VDMs. Our approach first employs a dedicated Keyframe Generatortrained without temporal compression to establish long-range, globallyconsistent structural anchors. Subsequently, an Interpolation Generatorsynthesizes the dense frames in an autoregressive manner with overlap-ping segments, utilizing the keyframes for global context and a singleclean preceding frame for local coherence. Trained on a large-scale inter-net long trajectory video dataset, our method achieves superior perfor-mance in both visual quality (lower FID and FVD) and camera adher-ence (lower ATE and ARE) compared to state-of-the-art autoregressiveand divide-and-conquer baselines, demonstrating stable and high-fidelitygeneration for long trajectory videos up to 32 seconds in length.

Non-rigid 3D shape matching is a fundamental task in com-puter vision and graphics. In this paper, we propose a hybrid self-supervisedmethod based on a coarse-to-fine strategy, which ensures consistencybetween the coarse mapping and the refined correspondence producedby our refinement module. The architecture features a dual-branch de-sign, consisting of two symmetric functional map learning streams: onebased on the Laplacian basis and the other utilizing the elastic ba-sis. Extensive experiments show that our approach not only maintainscomputational efficiency, but also achieves state-of-the-art performanceacross a variety of challenging scenarios, including non-isometric defor-mations and topological noise. Finally, we rigorously demonstrate thatcontrastive energies promote feature discrimination. Furthermore, inte-grating these energies with existing methods yields consistent improve-ments, validating the overall efficacy of our approach. Our code is avail-able at https://github.com/LuoFeifan77/Coarse-to-Fine-Hybrid-Self-Supervised-Matching.

Boundary representations (B-Reps) are fundamental to com-puter aided design (CAD), as they encode editable parametric surfacestogether with the precise topological structure required by downstreamengineering workflows. In contrast, modern 3D generative models pro-duce meshes or implicit fields, leaving a gap between high-quality geom-etry synthesis and CAD-native representations. Existing reconstructionpipelines treat B-Rep recovery as deterministic segmentation and fit-ting on discrete samples, despite the inherent ambiguity of CAD-styleface decompositions, often resulting in unstable topology recovery. Weintroduce BRepFacetGen, a generative framework that formulates CADstructure inference as a latent variable problem conditioned on geometryrepresenting continuous surfaces. Our representation couples a pretrainedgeometry latent set with a geometry-conditioned label latent space thatmodels multiple plausible dense surface segmentations, naturally han-dling permutation ambiguity and structural non-uniqueness. Joint de-coding yields a faceted mesh where each segment maps to a B-Rep facethat can be recovered by surface fitting and topology reconstruction. Ex-periments show improved segmentation and more reliable downstreamB-Rep recovery, as well as strong performance in image-conditioned andpoint-conditioned B-Rep generation.

Open-vocabulary 3D scene understanding aims to segment3D scenes beyond predefined categories by transferring semantic knowl-edge from vision-language models. Existing methods have advanced thistask by lifting language-aligned 2D features into 3D, yet they oftenrely on context-independent semantic representations, leaving object re-lationships underexplored for contextual refinement. We propose Rel-GraphOV, a relationship-aware framework that uses 3D scene graphs toenhance open-vocabulary 3D understanding. Our method constructs re-lational scene graphs from multi-view observations by leveraging vision-language reasoning to infer object relationships and prune geometricallyimplausible connections, without manual relationship annotations. Toaggregate relational context while avoiding feature interference, we in-troduce an Adaptive Gated Dual-Stream Contextual GAT that separatesdense geometric features and semantic CLIP embeddings, performs edge-guided message passing, and adaptively fuses complementary semantics.A hierarchical contrastive objective further promotes instance-level con-sistency and category-level discrimination. Experiments on ScanNetV2,ScanNet200, ScanNet++, and Replica demonstrate strong performanceand generalization ability. Project Page: cxavireh.github.io/relgraphov-projectpage

3D visual grounding aims to localize language-referred ob-jects in 3D scenes. However, existing methods rely on limited annotateddatasets under controlled environments and thus struggle to generalizein the wild. Therefore, the scarcity of scalable 3D grounding supervisionremains a major challenge. In this paper, we propose AnyGround3D,a scalable framework that learns to ground arbitrary 3D objects in thewild via 2D-to-3D lifting. Instead of requiring manual 3D annotations,our approach lifts diverse Internet images into structured 3D supervisionthrough multi-level 2D-to-3D lifting. First, we perform scene-level 2D-to-3D lifting to reconstruct metric point clouds from monocular images,providing geometric context. Second, we conduct object-level 2D-to-3Dlifting to recover complete 3D object representations with fine-gained 3Dbounding boxes and neural language descriptions. To further enhancethe representation robustness and open-vocabulary generalization of thepoint encoder, we introduce feature-level lifting that distills semantic rep-resentations from 2D foundation models into the 3D encoder. Together,these lifting processes enable large-scale synthetic 3D grounding supervi-sion across diverse real-world scenes. Extensive experiments demonstrateimproved in-the-wild generalization across multiple 3D grounding bench-marks, without relying on human-annotated 3D grounding data.

Vector Neuron Networks (VNNs) are widely used in 3D tasksfor their data efficiency and strong generalization from equivariance.However, their rigid equivariance constraints hinder handling symmetry-breaking, where low-symmetry outputs must be inferred from highlysymmetric inputs. In this paper, we reveal the representation collapse is-sue in VNNs and propose Ada-VNNs, a vector-neuron architecture thatcan adaptively relax equivariance constraints according to the symmetrylevel of the data. Ada-VNNs introduce an Adaptive Gating Unit (AGU)to control equivariance relaxation in linear layers, and design a resid-ual pathway that transforms hard structural constraints into adjustablesoft priors. We further provide a theoretical characterization that linksthe learned equivariance behavior to intrinsic symmetry breaking in thedata. Experiments on pose estimation across 26 categories with differ-ent degrees of symmetry breaking demonstrate that Ada-VNNs enabledata-driven adaptive equivariance adjustment, alleviate representationcollapse, and significantly improve symmetry-breaking capability overVNN and VN-Transformer, achieving higher performance with almostno additional overhead.

Volume denoising is a foundational problem in computa-tional imaging, as many 3D imaging inverse problems face high levels ofmeasurement noise. Inspired by the strong 2D image denoising propertiesof Field of Junctions (ICCV 2021), we propose a novel, fully volumetric3D Field of Junctions (3D FoJ) representation that optimizes a junctionof 3D wedges that best explain each 3D patch of a full volume, while en-couraging consistency between overlapping patches. In addition to directvolume denoising, we leverage our 3D FoJ representation as a structuralprior that: (i) requires no training data, and thus precludes the risk of hal-lucination, (ii) preserves and enhances sharp edge and corner structuresin 3D, even under low signal to noise ratio (SNR), and (iii) can be used asa drop-in denoising representation via projected or proximal gradient de-scent for any volumetric inverse problem with low SNR. We demonstratesuccessful volume reconstruction and denoising with 3D FoJ across threediverse 3D imaging tasks with low-SNR measurements: low-dose X-raycomputed tomography (CT), cryogenic electron tomography (cryo-ET),and denoising point clouds such as those from lidar in adverse weather.Across these challenging low-SNR volumetric imaging problems, 3D FoJoutperforms the evaluated classical denoisers, untrained neural denois-ers, and denoisers trained only on noisy examples. Code is available athttps://github.com/voilalab/3D-Field-of-Junctions.

Generating precise multi-entity 3D trajectories is fundamen-tal for interactive video generation and embodied world models. However,existing text-to-trajectory mappings face two bottlenecks: severe featureleakage from coupled multi-modal inputs, and physical collisions causedby blind end-to-end 3D coordinate regression. We propose ProgTraj-Director, a framework driven by physical property disentanglement andhierarchical spatial planning. To resolve the “who does what” attribu-tion ambiguity, our Structured Vision-Language Alignment (StructVLA)module disentangles static identity masks from dynamic spatial motionflows, binding them via orthogonal factorized tensor fusion in the latentspace. This formulation encourages entity-specific semantic separationand mitigates cross-entity feature leakage. To bridge the “where to go”semantic-physical gap, we formulate trajectory generation as hierarchicalspatial planning. The model first constructs a Geometry-Aware Bird’s-EyeView (BEV) to resolve anti-collision topological boundaries. Guided bythis structural prior, 3D trajectories are derived via progressive spatialdenoising to improve geometric and temporal consistency. To supportthis research, we introduce Stepwise-ME, a large-scale dataset of over31,000 interactive video clips (8M frames) with fine-grained stepwiseannotations. Extensive experiments show that our approach enhancesaction attribution and spatial coherence, providing a reliable structuralprior for downstream video generation. Code and dataset are available athttps://github.com/wzbos-token/Who-Does-What-and-Where-to-Go.

Large-scale vision generative models, including diffusion andflow models, have demonstrated remarkable performance in visual gen-eration tasks. However, transferring these pre-trained models to down-stream tasks often results in significant parameter redundancy. In thispaper, we propose EntPruner, an entropy-guided automatic progressivepruning framework for diffusion and flow models. First, we introduceentropy-guided pruning, a block-level importance assessment strategyspecifically designed for generative models. Unlike discriminative models,generative models require preserving the diversity and condition-fidelityof the output distribution. As the importance of each module can varysignificantly across downstream tasks, EntPruner prioritizes pruning ofless important blocks using data-dependent Conditional Entropy Devia-tion (CED) metric. CED quantifies how much the distribution divergesfrom the learned conditional data distribution after removing a block.Second, we propose a zero-shot adaptive pruning framework to auto-matically determine when and how much to prune during training. Thisdynamic strategy avoids the pitfalls of one-shot pruning, mitigating modecollapse, and preserving model performance. Extensive experiments onDiT and SiT models demonstrate the effectiveness of EntPruner, achiev-ing up to 2.22× inference speedup while maintaining competitive gener-ation quality on ImageNet and three downstream datasets.

Synthesizing native 2K multi-garment virtual try-on is aformidable frontier in digital fashion, critically bottlenecked by two fun-damental limitations: the O(N 2 ) memory explosion induced by 2k con-ditions, and the spectral bias of diffusion models that over-smooths high-frequency fabric details. We present WearWow, an end-to-end, mask-freegenerative framework that pioneers ultra-high-resolution multi-garmentsynthesis. To mitigate the memory explosion , we propose Adaptive 2DToken Packing (ATP). ATP leverages inherent garment sparsity to al-gorithmically pack heterogeneous items onto a unified 2D canvas andprune uninformative background tokens, minimizing the effective se-quence length and subsequent memory overhead while rigorously pre-serving 2D spatial priors. To rectify texture degradation, we introducethe Multi-dimensional Try-on Reward (MTR) system. MTR synergizes aSemantic Guidance Reward to explicitly drive tactile restoration with aCloth Distribution Reward to implicitly anchor the physical distribution,a joint formulation that effectively mitigates the severe reward hacking.Furthermore, we curate WearWow-2K, an extreme-quality dataset com-prising native 2K triplets, providing physically correct spatial interac-tions that naturally empower the model’s mask-free generation. Exten-sive experiments demonstrate that WearWow establishes a new state-of-the-art, exceeding existing commercial baselines in native 2K multi-garment synthesis.

Multi-expert models have become the dominant paradigmfor long-tailed learning, largely attributed to their presumed ability tobenefit from expert diversity. However, we revisit this central assump-tion and reveal that diversity induced by logit adjustment or explicitregularizers does not guarantee better ensemble accuracy. Our work sug-gests that multi-expert models benefit more from variance reduction thandiversity maximization. We introduce VICAL, a VIcinal ConsistencyALignment framework that improves long-tailed recognition not by en-forcing expert diversity, but by reducing prediction variance. Specifically,our approach comprises two key components: Self-Consistency Learn-ing and Deep Ensemble Distillation. Self-Consistency Learning discour-ages reliance on unstable high-frequency information, smoothing the lo-cal loss landscape and mitigating overfitting, especially for tail classes.Deep Ensemble Distillation promotes cross-expert low-frequency seman-tic agreement using a low-resolution view, thereby sidestepping opti-mization conflicts with established knowledge. Extensive experiments onCIFAR-LT, ImageNet-LT, and iNaturalist 2018 show that VICAL consis-tently outperforms state-of-the-art methods, validating the effectivenessof our consistency-driven design.

Models trained via Contrastive Language-Image Pretrain-ing (CLIP) serve as the foundational vision encoders for many modernLarge Vision Language Models (LVLMs). Despite their widespread adop-tion, CLIP models exhibit a critical yet underexplored failure mode:irrelevant text appearing within images confounds visual representa-tions, biasing them toward lexical meaning rather than true visual se-mantics. This robustness issue, commonly described as a TypographicAttack (TA), exposes a vulnerability that poses a significant risk tosafety-critical applications such as autonomous driving. To achieve in-terpretable and effective robustness against TA, we propose a novel,training-free mechanistic interpretability method. Our method providessampling-based interpretations of hidden state representations and quan-titatively attributes semantic versus lexical focus to individual attentionheads. Through probabilistic analysis and circuit mining, we isolate spe-cific Vision Transformer (ViT) components that disproportionately en-code lexical information, thereby identifying the mechanistic source ofTA. We further show that simple interventions applied directly to theidentified circuits, without any additional training, can substantially im-prove robustness against Typographic Attacks in object classification.These interventions, such as selective adjustment of attention weights,outperform both supervised and training-free defense methods. Our ex-periments also demonstrate that applying the proposed intervention tothe vision encoders of several state-of-the-art LVLMs yields substantialgains in Visual Question Answering accuracy under Typographic At-tack interference on RIO-Bench. These results confirm both the efficacyand the generalizability of our mechanistic approach. Code is released athttps://github.com/Liu-524/SamplingTAR.

Tactile graphics are a primary medium for blind and low-vision (BLV) individuals to access non-textual information. However,they are difficult to scale or personalize. While recent generative modelshave revolutionized visual content creation, they are optimized for screen-based visual realism and fail to satisfy the haptic perceptual and physicalfabrication constraints required for touch. We present the first integratedgenerative system that produces fabrication-ready 2.5D tactile graphicsdirectly from natural language prompts, jointly generating global basegeometry, fine-grained tactile surface textures, and standard-compliantbraille within a unified 3D-printable representation. Our approach intro-duces fabrication-aware techniques, including template-guided relief gen-eration, a fast diffusion-based text-to-texture module for high-resolutiontileable normal maps, and strict base flattening to ensure tactile readabil-ity and printability, while supporting both automatic generation and in-teractive texture control. Extensive evaluations, together with in-personuser studies with BLV participants and blindfolded sighted participantsusing physically 3D-printed outputs, show that participants consistentlyprefer our results over baselines. By extending generative graphics be-yond screens to touchable reliefs, our work broadens access to generativeAI for the BLV community and beyond. The code and project webpageare available at https://ruihangao.github.io/Text2TactileGraphics/.

Test Time Training (TTT) is a mechanism in which a modeladapts to an incoming test-sample by performing some self-supervised(SSL) task and updating its weights even during inference. This proce-dure does not require labels at test-time. This paper focuses on TTT forlong-videos. A major concern with existing approaches is: 1) they per-form TTT updates using a sliding window containing frames in the past,whose compute increases linearly with the size of window. This becomescomputationally intractable when the videos are hours long. 2) TTT isperformed even when temporally close frames look similar, thereby con-suming a lot of compute.We present the Frame Forgetting Network (FFN) that: 1) operates ononly three frames within the sliding window, namely the frame that ex-its, the current frame and the frame after that. The model still managesto retain temporal context and work for hours long-videos; 2) mathemat-ically define a ‘surprise’ metric: how much ‘new information’ the incom-ing frame contains with respect to the past seen frame. This facilitatesdetermining how to modify the effective window size during TTT andconstitutes the core mechanism of an adaptive windowing algorithm. Ad-ditionally, we curate a dataset EpicTours containing up to 3 hour longvideos of walking city-tours, whereas earlier datasets on this problemwere only 5 min long. We demonstrate FFN’s empirical effectiveness ondense-segmentation, video classification tasks, generalization to depth-estimation, and multi-hour long videos. The project page can be foundat https://github.com/rajatmodi62/ffn.

Diffusion Transformers (DiTs) have significantly advancedaudio-driven portrait animation, but their high computational cost leadsto substantial inference latency. Although training-free diffusion cachingaccelerates inference significant, existing methods are primarily devel-oped for text-conditioned generation and overlook the spatial and modal-ity imbalances inherent in audio-driven portrait animation. In this pa-per, we propose SyncCache, a training-free caching acceleration methodtailored for DiT-based portrait animation that explicitly exploits asym-metric dynamics. Specifically, high-frequency dynamics driven by audioconditions and concentrated in human regions are more challenging andcritical to cache and reuse than the low-frequency visual background inportrait animation. First, we introduce Spatially-Asymmetric Probingto prioritize error sensitivity in dynamic human region. Second, throughModality-Decoupled Caching, we bypass heavy DiT block by reusingstable inter-block residuals, while continuously recomputing lightweightaudio blocks to preserve precise lip synchronization. Furthermore, weintroduce a cache ratio to control cache capacity and formulate memory-adaptive cache selection as an offline dynamic programming problemwithout online overhead. Extensive experiments demonstrate that Sync-Cache achieves superior speed–quality trade-offs, delivering up to 4.12×acceleration on HunyuanVideo-Avatar and 3.75× on Wan-S2V with near-lossless visual fidelity and precise audio alignment.

Feedforward reconstruction is crucial for autonomous drivingapplications, where rapid scene reconstruction enables efficient utiliza-tion of large-scale driving datasets in closed-loop simulation and otherdownstream tasks, eliminating the need for time-consuming per-scene op-timization. We present StreetForward, a pose-free and tracker-free feed-forward framework for dynamic street reconstruction. Building uponthe alternating attention mechanism from Visual Geometry GroundedTransformer (VGGT), we propose a simple yet effective temporal maskattention module that captures dynamic motion information from im-age sequences and produces motion-aware latent representations. Staticcontent and dynamic instances are represented uniformly with 3D Gaus-sian Splatting, and are optimized jointly by cross-frame rendering withspatio-temporal consistency, allowing the model to infer per-pixel ve-locities and produce high-fidelity novel views at new poses and times.We train and evaluate our model on the Waymo Open Dataset, demon-strating superior performance on novel view synthesis and depth estima-tion compared to existing methods. Furthermore, zero-shot inference onCARLA validates the generalization capability of our approach.

We propose a novel training-free method for inpainting with off-the-shelf text-to-image models. While guidance-based methods in theory allow generic models to be used for inverse problems such as inpainting — in practice their effectiveness is limited, leading to the necessity of specialized inpainting-specific models. In this work, we argue the missing ingredient for training-free generic model usage is proper optimization of the initial noise sample. We optimize the initial noise to approximately reproduce the unmasked image, in as few as tens of optimization steps, then use it with a conventional training-free inpainting method. Critically, we propose two core ideas that make this possible: (i) we perform linear approximation that avoids the costly and often impractical unrolling required to relate the initial noise sample to model output—which potentially is why this relationship was previously overlooked; and (ii) perform spectral preconditioning by optimizing the initial noise sample in the spectral domain with Adam, which stabilizes the optimization. We demonstrate our method on various inpainting tasks, outperforming the state of the art. Project website: https://ubc-vision.github.io/sonic/