ECCV 2026 论文列表

The quality of the latent space in visual tokenizers (e.g.,VAEs) is crucial for modern generative models. However, the standardreconstruction-based training paradigm produces a latent space that isbiased towards low-level information, leading to a foundational flaw: bet-ter pixel-level reconstruction accuracy does not lead to higher-qualitygeneration. This implies that pouring extensive compute into visual tok-enizer pre-training translates poorly to improved performance in gener-ation. We identify this as the “pre-training scaling problem” and suggesta necessary shift: to be effective for generation, a latent space must con-cisely represent high-level semantics. We present VTP, a unified visualtokenizer pre-training framework, pioneering the joint optimization ofimage-text contrastive, self-supervised, and reconstruction losses. Ourstudy reveals that perception-oriented tokenizer pre-training un-locks a new scaling law for generation, where generative perfor-mance scales effectively with compute, parameters, and data allocatedto the pre-training of the visual tokenizer. Our large-scale pre-trainingexperiments demonstrate the following results: (1) Without modifyingDiT training specs and FLOPs, solely scaling VTP pre-training con-sistently achieves gains in both ImageNet class-conditional and LAIONtext-to-image generation, while conventional autoencoders stagnate veryearly at 1/10 of the FLOPs. (2) VTP achieves 0.36 rFID while simultane-ously delivering 78.2% zero-shot accuracy and 85.7% linear probing accu-racy, surpassing prior unified tokenizers such as VILA-U and UniTok. (3)Furthermore, the VTP-based diffusion model exhibits exceptionally fastconvergence—reaching 2.03 gFID in only 80 epochs without guidancetricks, outperforming previous methods like VA-VAE and RAE—and ul-timately scales to achieve a remarkable 1.11 gFID on ImageNet 256×256generation. Our code and models are publicly available at VTP.

Visual token pruning, which aims to compress and pruneredundant visual tokens, plays a critical role in efficient inference withlarge vision-language models (LVLMs). However, existing methods failto disentangle intra-modal visual redundancy from cross-modal redun-dancy between vision and language. We show that visual token diversityand task-specific token relevance are two crucial yet orthogonal factorsthat complement each other in conveying useful information and shouldtherefore be treated separately for more effective visual token pruning.Building upon this insight, we design ToDRE, a two-stage and training-free framework that incorporates Token Diversity and task RElevancefor effective token compression and efficient LVLM inference. Instead ofpruning redundant tokens, we introduce a greedy max-sum diversificationalgorithm that selects and retains a subset of diverse and representativevisual tokens after the vision encoder. On top of that, ToDRE leveragesan “information migration” mechanism to eliminate task-irrelevant visualtokens within certain decoder layers of the large language model (LLM),further improving token pruning and LVLM inference. Extensive exper-iments show that ToDRE prunes 90% of visual tokens after the visionencoder as well as all visual tokens in certain LLM decoder layers, leadingto a 2.6× speed-up in total inference time while maintaining 95.0% modelperformance plus excellent model compatibility.

Multi-exposure fusion (MEF) brings the dynamic range ofconventional cameras closer to that of human vision, producing imageswith rich scene content. Given the large variability in scene luminance,exposure strategies often require different numbers of frames to capturethe full radiance range faithfully. However, conventional MEF techniquesare typically designed for a fixed number of inputs, forcing deploymentsystems to maintain separate models for different frame-count require-ments, which undermines deployment efficiency. To address this limita-tion, we propose FreeMEF, the first flexible-frame transformer for MEFthat seamlessly accommodates varying numbers of input exposures with-out retraining or architectural changes. The proposed approach consistsof two key modules. First, we introduce a recurrent state space mod-ule (RSSM) that sequentially fuses features from arbitrary sequencesvia adaptive alignment and state-space recurrent modeling, thereby pro-viding global information guidance for the subsequent restoration. Sec-ond, we devise a global feature guided block (GFGB) incorporating anextremity-aware hybrid attention (EAHA) and an affine-injection feed-forward network (AFFN), which effectively resolves the similarity para-dox while simultaneously optimizing contrast and brightness regulation.Extensive experiments on three benchmark datasets demonstrate the ef-fectiveness of our method, which performs favorably against state-of-the-art methods both quantitatively and qualitatively. The code is availableat https://github.com/qulishen/FreeMEF.

Recent high-resolution Multimodal Large Language Models(MLLMs) generate thousands of visual tokens per input, leading to avisual token explosion that introduces severe latency bottlenecks. Whiletoken pruning mitigates this issue, state-of-the-art subset-optimizationmethods typically rely on iterative subset construction to jointly cap-ture visual diversity and instruction relevance. As visual token countsscale, this sequential dependency introduces significant selection over-head, severely limiting the translation of theoretical FLOPs reductionsinto actual wall-clock speedups. To address this limitation, we proposeSingle-Forward Pruner (SFPruner), a structural reformulation of visualtoken pruning that embeds redundancy control directly into the scor-ing space, bypassing the need for iterative combinatorial optimization.Our non-iterative framework achieves redundancy-aware importance se-lection in a single forward pass through two complementary mechanisms.First, to attenuate redundancy at the covariance level, we introduce asemantics-guided ridge leverage scheme. By integrating instruction rel-evance and visual saliency, this mechanism suppresses dominant covari-ance directions and mitigates representation bias. Second, ranking-baseddirectional masking resolves residual overlap through asymmetric sim-ilarity competition, where higher-scoring tokens explicitly suppress re-dundant lower-scoring alternatives via parallel tensor operations. Exten-sive evaluations demonstrate that our approach maintains stable selec-tion costs, reducing the token selection process by up to 110 ms (from112.4 ms to just 2.5 ms at 512 tokens in Qwen2.5-VL). This structuralefficiency successfully translates theoretical token reductions into tangi-ble inference speedups while preserving highly competitive performanceagainst state-of-the-art techniques under aggressive compression.

In this paper, we explore dense voxel streaming for accurateand efficient 3D occupancy prediction. While dense voxel representa-tions offer fine-grained spatial details and streaming paradigm enablesefficient temporal processing, naively combining the two introduces keychallenges: (i) warping-induced distortions caused by interpolation usedfor temporal alignment, and (ii) degraded dynamic object representa-tions due to motion misalignment and detail loss in image-to-voxel pro-jection. To address these, we propose StreamOcc, a novel frameworkthat utilizes two aggregation strategies. Specifically, it first refines prop-agated voxel features to reduce warping artifacts before temporal accu-mulation, and then selectively injects instance-level query features en-coding dynamic-object semantics into the corresponding occupied voxelregions, preserving temporally consistent modeling while strengtheningdynamic object representations. Unlocking effective dense voxel stream-ing, StreamOcc achieves state-of-the-art performance on SurroundOcc-benchmark and Occ3D-nuScenes under real-time constraints, outper-forming the prior best methods by +1.3/2.5 and +1.5/2.0 in (over-all/dynamic object) mIoU, respectively, while running at 83.3 msper frame with only 2.8 GB of memory. The project page is available athttps://moonseokha.github.io/StreamOcc/.

Expressive human avatar creation from a single image ishighly challenging due to the inherently ill-posed nature of the prob-lem, as well as the complexities of appearance preservation and dynamichuman modeling. To address these challenges, this work presents a com-prehensive pipeline for training a large feed-forward model that efficientlygenerates fully controllable 3D avatars from single-image. Due to thescarcity of consistent training data, we construct a large-scale 3D Gaus-sian avatar dataset to support model training. To better recover appear-ance details, we propose an input-aware decoding scheme that fully lever-ages information from the input image. Furthermore, to achieve compre-hensive full-body control, we introduce a Gaussian blending–based facialenhancement module and apply Gaussian geometric constraints to sta-bilize expressive avatar generation. Extensive experiments demonstratethat our method enables one-shot reconstruction of photorealistic avatarswith whole-body control, outperforming existing works in terms of hu-man dynamic and appearance details.

Multimodal Large Language Models (MLLMs) have demon-strated immense potential in Earth observation. However, the massivevisual tokens generated when processing Ultra-High-Resolution (UHR) im-agery introduce prohibitive computational overhead, severely bottleneck-ing their inference efficiency. Existing visual token compression methodspredominantly adopt static and uniform compression strategies, neglectingthe inherent Semantic–Geometric Duality in remote sensing interpretationtasks. Specifically, object semantic tasks focus on the abstract seman-tics of objects and benefit from aggressive background pruning, whereasscene geometric tasks critically rely on the integrity of spatial topol-ogy. To address this challenge, we propose DualComp, a task-adaptivedual-stream token compression framework. Dynamically guided by alightweight pre-trained router, DualComp decouples feature processinginto two dedicated pathways. In the object semantic stream, the Spatially-Contiguous Semantic Aggregator (SCSA) utilizes size-adaptive clusteringto aggregates redundant background while protecting small object. Inthe scene geometric stream, the Instruction-Guided Structure Recoverer(IGSR) introduces a greedy path-tracing topology completion mechanismto reconstruct spatial skeletons. Experiments on the UHR remote sens-ing benchmark XLRS-Bench demonstrate that DualComp accomplisheshigh-fidelity remote sensing interpretation at an exceptionally low com-putational cost, achieving simultaneous improvements in both efficiencyand accuracy.

Self-supervised neural garment simulation has become pop-ular due to its computational efficiency, good visual realism, and noreliance on training data. However, existing methods greatly simplifythe mechanical properties of fabrics, ignoring persistent wrinkles causedby plasticity. Although this simplification allows for modeling of purelyelastic material and simple training via energy minimization, the lackof believable wrinkles adversely affects the visual realism. Therefore, weintroduce the first self-supervised neural garment simulator that explic-itly models persistent wrinkles. This is accomplished through a novelphysics-inspired loss function, which turns learning into a moving energyminimization problem to mimic plasticity. However, this requires learn-ing to use a changing loss function, which causes difficulties in trainingbecause the loss function changes during optimization. To this end, wepropose a new physics-inspired curriculum learning scheme where thetarget material for learning gradually changes from pure elasticity toelasto-plasticity, allowing the loss function and the learnable parametersto jointly converge. Through a comprehensive evaluation, we show thatfor the first time, self-supervised learning models can generate naturalpersistent wrinkles, outperforming existing methods on a variety of gar-ments, body shapes, and body motions, according to a range of metrics.Our code is publicly available at https://github.com/realcrane/EPNet

Imitation learning enables robots to acquire manipulationskills from demonstrations by mapping observations to actions. Existingapproaches predict either short-horizon continuous action sequences ordiscrete keyposes. However, continuous prediction methods suffer fromcompounding errors due to short prediction horizons and struggle withmulti-modal action distributions, whereas keypose-based methods ne-cessitate an external planner, constraining real-time applicability. Toaddress these challenges, we introduce SegDiff, a closed-loop visuomo-tor policy that integrates the strengths of both paradigms. SegDiff de-composes demonstrations into motion segments between keyposes andlearns to predict the continuous trajectory from the current state tothe next keypose, enabling long-horizon prediction with real-time refine-ment. Furthermore, we leverage the capability of diffusion models andDDIM inversion to propose a Dynamic Temporal Ensembling mecha-nism, which allows the policy to efficiently respond to dynamic environ-ments and mitigate discontinuities caused by inconsistent multi-modalsampling. SegDiff demonstrates significant performance gains over exist-ing approaches across various simulated and real-world scenarios, indi-cating its strong ability to reason over extended temporal dependencieswhile maintaining real-time adaptability and control stability.

Adversarial examples exhibit cross-model transferability, en-abling threatening black-box attacks on commercial models. Model en-sembling, which attacks multiple surrogates, is a known strategy to im-prove this transferability. However, prior studies typically use small, fixedensembles, which leaves open an intriguing question of whether scalingthe number of surrogate models can further improve black-box attacks.In this work, we conduct the first large-scale empirical study of thisquestion. We show that by resolving gradient conflict with advanced op-timizers, we overcome the quantitative limitations of idealized theoreticalbounds to empirically discover a robust log-linear scaling law, demon-strating that the Attack Success Rate scales linearly with the logarithmof the ensemble size T . We rigorously verify this law across standardclassifiers, SOTA defenses, and MLLMs, and find that scaling distills ro-bust, semantic features of the target class. Consequently, we apply thisfundamental insight to benchmark SOTA MLLMs. This reveals both theattack’s devastating power and a clear robustness hierarchy, as evidencedby achieving an over 80% transfer attack success rate on proprietarymodels like GPT-4o, while also highlighting the exceptional resilienceof Claude-3.5-Sonnet. Our findings urge a shift in focus for robustnessevaluation: from designing intricate algorithms on small ensembles tounderstanding the principled and powerful threat of scaling.

Dense feature matching aims to estimate all correspondencesbetween two images of a 3D scene and has recently been established asthe gold standard due to its high accuracy and robustness. However,existing dense matchers still fail or perform poorly for many hard real-world scenarios, and high-precision models are often slow, limiting theirapplicability. In this paper, we attack these weaknesses on a wide frontthrough a series of systematic improvements that together yield a sig-nificantly better model. In particular, we construct a novel matchingarchitecture and loss, which, combined with a curated diverse trainingdistribution, enables our model to solve many complex matching tasks.We further make training faster through a decoupled two-stage matching-then-refinement pipeline, and at the same time, significantly reduce re-finement memory usage through a custom CUDA kernel. Finally, weleverage the recent DINOv3 foundation model along with multiple otherinsights to make the model more robust and unbiased. In our extensiveset of experiments, we show that the resulting novel matcher sets a newstate-of-the-art, being significantly more accurate than its predecessors.ω

With the rapid advancement of vision–language models, anincreasing number of studies have explored their potential for SVG gener-ation tasks. Although existing approaches improve performance by con-structing large-scale SVG datasets and introducing SVG-specific tokens,they still suffer from limited generalization, redundant paths in code out-puts, and a lack of explicit reasoning. In this work, we present CTRL-S(Chain-of-Thought Reinforcement Learning for SVG), a unified frame-work that introduces a chain-of-thought mechanism to explicitly exposethe model’s reasoning process during SVG generation. To support thisstructured reasoning, we construct SVG-Sophia, a high-quality datasetcontaining 145K samples across SVG code refinement, Text-to-SVG, andImage-to-SVG tasks. By training the model to generate group-level struc-tured SVG code, CTRL-S significantly improves structural coherenceand visual fidelity. Furthermore, we adopt the GRPO algorithm and de-sign a multi-reward optimization framework, incorporating DINO, im-age–text similarity, format, and code efficiency rewards. Through jointmulti-reward optimization and multi-task training, our approach system-atically enhances overall generation capabilities. Extensive experimentsshow that CTRL-S outperforms existing methods, achieving higher tasksuccess rates, superior SVG code quality, and exceptional visual fidelity.Our code is available at https://github.com/hmwang2002/CTRL-S.

3D Gaussian Splatting (3DGS) has emerged as a powerfultechnique for real-time, high-fidelity novel view synthesis. Despite theseadvantages, its multi-million point representation leads to large stor-age footprints that hinder asset distribution and on-device deployment.Existing compression approaches typically rely on access to training im-ages, camera poses, or the rendering pipeline for retraining or iterativefine-tuning. However, in realistic post-training and archival scenarios,often only the optimized model parameters are available. To addressthis limitation, we propose a render-agnostic, post-training compressionpipeline for 3DGS operating strictly under a model-only constraint. Wefirst introduce a camera-independent, multi-criteria importance score toprune redundant Gaussians based on a combination of geometric, spa-tial, and appearance-related cues. Following this pruning, a lightweight,training-free refinement step conservatively restores local coverage. Tofurther reduce the storage footprint, we apply degree-wise vector quanti-zation to the spherical harmonic coefficients and compress the remainingattributes via entropy coding. The compression performance is signifi-cantly enhanced by a custom spatial reordering to maximize local dataredundancy. Experiments on standard benchmarks show that our ap-proach achieves an average of 30× compression rate while maintainingcompetitive visual quality against existing methods.

In active reconstruction, an embodied agent must decidewhere to look next to efficiently acquire views that support high-qualitynovel-view rendering. Recent work on active view planning for neuralrendering largely derives next-best-view (NBV) criteria by backpropa-gating through radiance fields or estimating information entropy over3D Gaussian primitives. While effective, these strategies tightly cou-ple view selection to heavy, representation-specific mechanisms and failto account for the computational and resource constraints required forlightweight online deployment. In this paper, we revisit active recon-struction from a renderability-centric perspective. We propose R3CON,a radiance-fields-free active reconstruction framework that induces an im-plicit, pose-conditioned renderability field over SE(3) from a lightweightvoxel map. Our formulation aggregates per-voxel online observation statis-tics into a unified scalar renderability score that is cheap to update andcan be queried in closed form at arbitrary candidate viewpoints in mil-liseconds, without requiring gradients or radiance-field training. This ren-derability field is strongly correlated with image-space reconstruction er-ror, naturally guiding NBV selection. We further introduce a panoramicextension that estimates omnidirectional (360◦ ) view utility to acceleratecandidate evaluation. In the standard indoor Replica dataset, R3CONachieves more uniform novel-view quality and higher 3D Gaussian splat-ting (3DGS) reconstruction accuracy than recent active GS baselineswith matched view and time budgets. .

Concept erasure and machine unlearning have emerged asa critical paradigm for enabling models to remove unwanted knowledgefrom learned representations at the level of classes, attributes, or higher-level semantics, motivated by the safety concerns surrounding moderngenerative AI models. However, existing approaches face three persis-tent challenges: (1) they often degrade utility on the remain concepts,(2) they often leave residual feature-level signals that enable recoveryof the forget concepts via _x001C_ne-tuning or representation probing, and (3)they impose substantial computational or memory overhead, limitingscalability. Prior work, such as Erasing Space Concept (ESC), moves to-ward structured removal in representation space, but achieving robusterasure with low overhead remains a challenge. To address this issue, wepropose Robustly Erasing Space Concepts (R-ESC), a framework for ro-bustly erasing space concepts via stochastic feature remapping. R-ESCintroduces three components: (i) prototype-orthogonal projection thatreduces interference by decorrelating forget and remain prototypes priorto erasure, (ii) stochastic multi-target remapping that blends forget pro-totypes into multiple remain prototypes to prevent re-separation andhinder recovery, and (iii) activation-mean prototypes that compress theerasure procedure to a single forward pass, yielding linear-time computa-tion with constant additional memory. Across extensive experiments, R-ESC maintains remain data utility, strengthens resistance to feature-levelrecovery, and scales to large models and datasets, providing a practicalpathway toward robust and e_x001E_cient concept erasure.

Most end-to-end autonomous driving methods rely solely oninstantaneous sensor observations, limiting them to reactive behaviorwithout the anticipatory foresight human drivers employ through priorexperience. We introduce geospatial visual priors, street-level visual con-text anchored to the intended driving route, providing visual-spatialforesight independent of real-time sensors. We propose a memory aug-mentation module featuring a dual-memory architecture and an adap-tive memory gate, which can be easily integrated into existing end-to-end approaches. This design pairs a contextual memory for retrievedpriors with a persistent fallback memory, and dynamically regulatesthe influence of memories based on current state compatibility. Eval-uated on the NAVSIM-v2 benchmark, our approach consistently im-proves performance across diverse end-to-end baselines. Furthermore,because these priors are independent of onboard sensors, our method in-herently improves robustness against sensor corruption, while the dual-memory design ensures safe fallback when the retrieved priors them-selves become unreliable. Our project page is available at https://ori-mrg.github.io/PriorEye.

Image-to-point cloud registration aims to estimate the 6-DoFcamera pose of a query image with respect to a 3D point cloud, whichremains challenging due to the heterogeneous nature of dense visual ap-pearance and sparse geometric structure. Most existing learning-basedmethods follow an overlap–then–registration paradigm, where pose esti-mation is restricted to a predicted overlapping region. Such a sequentialdecomposition is inherently fragile, as inaccurate overlap prediction maydiscard geometrically consistent regions and irreversibly constrain poseoptimization. In this work, we reformulate image-to-point cloud registra-tion as a pose–geometry coupled reasoning problem, where camera poseand scene geometry are treated as interdependent variables. Based onthis formulation, we propose PGCR, a unified framework that estimatescamera pose by explicitly reasoning over dense cross-modal geometry.PGCR jointly predicts camera pose together with dense geometric rep-resentations and grounds pose inference in global geometric consistency.To enable effective pose–geometry coupling, we introduce a Geometry-aware Pose Refinement mechanism to enforce prediction-level consistencybetween pose and geometry, and a Pose-guided Progressive Refinementstrategy to adapt cross-modal interaction according to intermediate poseestimates. Extensive experiments on two widely used outdoor and twoindoor benchmarks demonstrate that PGCR consistently outperformsprior state-of-the-art methods across all evaluation metrics while remain-ing computationally efficient.

360 video object segmentation (360VOS) aims to predicttemporally-consistent masks in 360 videos, offering full-scene coverage,benefiting applications, such as VR/AR and embodied AI. Learning360VOS model is nontrivial due to the lack of high-quality labeled dataset.Recently, Segment Anything Models (SAMs), especially SAM2 – withits design of memory module – shows strong, promptable VOS capabil-ity. However, directly using SAM2 for 360VOS yields implausible resultsas 360 videos suffer from the projection distortion, semantic inconsis-tency of left-right sides, and sparse object mask information in SAM2’smemory. To this end, we propose PanoSAM2, a novel 360VOS frame-work based on our lightweight distortion- and memory-aware adaptationstrategies of SAM2 to achieve reliable 360VOS while retaining SAM2’suser-friendly prompting design. Concretely, to tackle the projection dis-tortion and semantic inconsistency issues, we propose a Pano-AwareDecoder with seam-consistent receptive fields and iterative distortionrefinement to maintain continuity across the 0°/360° boundary. Mean-while, a Distortion-Guided Mask Loss is introduced to weight pixelsby distortion magnitude, stressing stretched regions and boundaries. Toaddress the object sparsity issue, we propose a Long–Short Mem-ory Module to maintain a compact long-term object pointer to re-instantiate and align short-term memories, thereby enhancing temporalcoherence. Extensive experiments show that PanoSAM2 yields substan-tial gains over SAM2: +5.6 on 360VOTS and +6.7 on PanoVOS, show-ing the effectiveness of our method. Code is available at https://github.com/Eric-Bumaro/PanoSAM2.

As Vision–Language Models (VLMs) increasingly power au-tonomous agents—from coding assistants that interpret terminal outputand execute shell commands, to Computer-Use Agents with full desktopcontrol, to HR screening and loan evaluation bots—these agents act di-rectly over visual artifacts from untrusted sources. While prior work hasstudied direct multimodal jailbreaks, text-based indirect prompt injec-tion, and HTML-embedded attacks, the security implications of Multi-modal Indirect Prompt Injection (M-IPI) in realistic agentic workflowsremain insufficiently understood. We ask a central question: does pro-cessing an artifact visually, rather than as text, degrade an agent’s abil-ity to resist embedded attacks? To study this, we introduce the M-IPIBenchmark—a suite of 2,600 high-fidelity visual artifacts encompass-ing two attack families: (1) Technically-Framed Attacks, where maliciouscommands are interwoven with genuine debugging workflows in terminalscreenshots, and (2) Ethics-Framed Attacks, a novel vector where adver-saries exploit alignment priors such as fairness mandates to override eval-uation policies—and use it to systematically analyze modality-dependentfailures. We evaluate five model families in paired vision and text-onlyconfigurations that receive identical content, isolating the modality effect.Our results reveal that VLMs consistently underperform text-only coun-terparts at detecting embedded attacks—a “visual authority” effect whererendered presentation suppresses critical evaluation. For end-to-end at-tack success, both modalities show comparable susceptibility (∼50% fortechnical, >80% for ethics-framed), with qualitatively distinct failureprofiles: VL models disproportionately trust visually-rendered commandswhile text-only models are more vulnerable to natural-language direc-tives. These findings indicate that the vulnerabilities are rooted in theshared language backbone and that alignment priors represent a dis-tinct, exploitable attack surface. Our benchmark artifacts are releasedunder gated, research-only access at https://github.com/adityakm24/OpenYourEyes to support future defensive research.