ECCV 2026 Papers

Multimodal Large Language Models have achieved signif-icant success in offline video understanding, yet their application tostreaming videos is severely limited by the linear explosion of visual to-kens, which often leads to Out-of-Memory (OOM) errors or catastrophicforgetting. Existing visual retention and memory management methodstypically rely on uniform sampling, low-level physical metrics, or pas-sive cache eviction. However, these strategies often lack intrinsic seman-tic awareness, potentially disrupting contextual coherence and blurringtransient yet critical semantic transitions. To address these limitations,we propose CurveStream, a training-free, curvature-aware hierarchicalvisual memory management framework. Our approach is motivated bythe key observation that high-curvature regions along continuous featuretrajectories closely align with critical global semantic transitions. Basedon this geometric insight, CurveStream evaluates real-time semantic in-tensity via a Curvature Score and integrates an online K-Sigma dynamicthreshold to adaptively route frames into clear and blurred memorystates under a strict token budget. Evaluations across diverse tempo-ral scales confirm that this lightweight framework, CurveStream, consis-tently yields absolute performance gains of over 10% (e.g., 10.69% onStreamingBench and 13.58% on OVOBench) over respective baselines,establishing new state-of-the-art results for streaming video perception.

Cervical cytology is practiced as a sequence of roles: screen-ing decides where to look, careful reading determines what is present,and reporting accounts for the decision. We translate this workflow intoa compact and structurally transparent model. Screening is cast as afree–energy allocation over four explicit, cytology–native concepts (nu-clear–envelope roughness, orientation disorder, local N/C ratio, chro-matin heterogeneity), yielding a closed-form spatial attention as a soft-max of negative energy. Reading is handled by an analytic backbonebuilt from fixed geometric responses, nonnegative near–diagonal chan-nel mixing, and convex residual updates, producing stable and auditablefeatures. A probabilistic readout with a linear head then provides anintrinsic evidence–to–logit accounting without post–hoc explainers. De-spite only 2.2M parameters, the model achieves state-of-the-art accuracyon DSCC (93.2% ACC), competitive results on SIPaKMeD (98.5% ACC)and Herlev (76.8% ACC), and favorable latency. Source code is publiclyavailable at https://github.com/Dichao-Liu/GeoCEAN.

4D understanding and reasoning is a fundamental capabil-ity for embodied AI agents operating in dynamic physical environments.However, existing vision encoders are largely limited to static 2D im-ages or 3D point clouds without temporal modeling, or to 2D videosthat lack accurate geometric depth reasoning. Consequently, current ap-proaches fail to jointly capture spatial structure and motion evolutionin dynamic scenes. We present CL4D, the first foundational 4D visionencoder that directly operates on dynamic point clouds, trained witha contrastive learning objective to align spatio-temporal geometric rep-resentations with natural language descriptions. By learning a shared1 ∗These authors contributed equally. † These authors jointly supervised this work.embedding space between text and 4D scene dynamics, CL4D enableszero-shot motion-to-text and text-to-motion retrieval in dynamic en-vironments and serves as a foundational 4D vision encoder for down-stream 4D vision–language tasks. Building on this encoder, we intro-duce 4DVLM, a 4D vision–language model that conditions languagegeneration on dynamic geometric representations. 4DVLM is the firstVLM designed to operate directly on 4D point clouds without relyingon 2D images, 2D videos, or static 3D point clouds. We train CL4D andsubsequently 4DVLM on a newly constructed dataset termed DynAc-tion4D capturing diverse human motions across varying object inter-actions and scene environments. Extensive experiments across multiple4D human action benchmarks demonstrate that CL4D achieves state-of-the-art performance, with improvements of approximately ∼16.75% overprior methods. Furthermore, 4DVLM outperforms frontier video VLMssuch as Gemini and GPT-5 even when these models are provided withRGB video sequences corresponding to the same scenes represented as4D point clouds for 4DVLM.

Traditional glass-based optics are typically optimized for nar-row spectral bands, such as the visible (400–700nm) or shortwave infrared(1000–1800nm). While the emergence of VIS-SWIR sensors (400–1700nm)offers transformative potential, refractive optics struggle to focus thisentire range simultaneously. Mirrors represent a promising achromaticalternative; however, they are often sidelined by field curvature, and off-axis aberrations. This paper introduces Computational Mirrors, aframework that enables high-resolution, full-field-of-view imaging acrossthe complete VIS-SWIR spectrum using a single sensor. Our method isbuilt on the observation that distinct regions of the field of view reachfocus at varying distances from the mirror. By capturing a minimal fo-cal stack (2–4 images), we utilize a computational backend to recovera sharp, all-in-focus image. A key contribution of this paper is Seidel-Conv, a novel, physics-inspired, spatially-varying point spread function(PSF) model designed to accurately characterize and correct the off-axisaberrations inherent in simple concave mirrors. We demonstrate the ef-ficacy of our approach using a first-of-its-kind 50mm F/1 optical systemequipped with a VIS-SWIR sensor. Our system produces sharp imagesacross RGB, NIR, and SWIR wavelengths without requiring refocusing,revealing material details invisible within individual spectral bands. Wefurther validate the scalability of our approach with a 100mm F/2 systemoptimized for long-range imaging.

Recent hybrid Structure-from-Motion (SfM) systems com-bine the robustness of feed-forward 3D reconstruction with the accuracyof traditional bundle adjustment (BA) with pixel matching. They areusually the best performing methods however their scalability and us-ability remains limited since estimating dense correspondences betweenviews is prohibitively costly, especially considering time constraints in-herent to online applications like Visual SLAM (VSLAM). In this paper,we introduce a regularized BA framework that leverages a fast multi-view matcher and monocular priors for initialization and regularization.In contrast to existing systems, our unified approach seamlessly supportsboth online VSLAM and offline reconstruction from unordered image col-lections within the same optimization framework and sharing commonhyperparameters for all tasks. Extensive experiments across both do-mains demonstrate improved performance and speed tradeoffs over tra-ditional, feed-forward, and hybrid baselines. Notably for VSLAM, ouruncalibrated method outperforms all previous calibrated approaches.

Bumblebees play a key role in agriculture, ecology, and bio-inspired robotics, making accurate 3D surface reconstruction essential forquantitative behavioral analysis. However, their dark dense hairs, com-plex structure and flexible limbs prone to visual occlusion and annotationdifficulties, resulting in difficulties in reconstruction. Current template-free or generic template-based models offer limited structural detail andbiological realism, while high-quality bumblebee templates and reliablemorphological priors are lacking. To address these challenges, we pro-pose BioMTBee, a template-based 3D mesh reconstruction frameworkfor bumblebees. We built a high-fidelity articulated mesh template frommicro-CT scans to provide an accurate morphological prior. A multi-view3D pose estimator with spatio-temporal filtering (BPST) is then intro-duced to extract robust 3D keypoints from 2D detections, guiding stablemesh template fitting. Building on this template, we combine 3D poseand shape supervision with biological constraints—bilateral symmetry,kinematic coupling, and temporal smoothness—to achieve anatomicallyconsistent and temporally stable surface reconstructions. Experiments on3D reconstruction from multi-view images show improvements in mor-phological accuracy and biological plausibility over representative base-lines, capturing fast limb movements and fine structural and texturaldetails, and also generalize well to cross-species reconstruction.

Understanding camera motion is a fundamental problem inembodied perception and 3D scene understanding. While visual methodshave advanced rapidly, they often struggle under visually degraded condi-tions such as motion blur or occlusions. In this work, we show that passivescene sounds provide cues complementary to vision for relative camerapose estimation for in-the-wild videos. We introduce a simple but effec-tive audio-visual framework that integrates direction-of-arrival (DOA)spectra and binauralized embeddings into a state-of-the-art vision-onlypose estimation model. Our results on two large datasets show consistentgains over strong visual baselines, plus robustness when the visual infor-mation is corrupted. To our knowledge, this represents the first work tosuccessfully leverage audio for relative camera pose estimation in real-world videos, and it establishes incidental, everyday audio as an unex-pected but promising signal for a classic spatial challenge.

While text-to-image diffusion models achieve impressive vi-sual quality, they frequently struggle to maintain precise alignment withcomplex compositional prompts. An effective strategy is to improve theinference process of diffusion models, thereby better leveraging their pre-trained priors to address misalignment. Existing training-free methodscan be divided into two categories. The first category focuses on improv-ing the randomly sampled initial noise to obtain an initialization thatencapsulates semantics relevant to the target text. However, existing ap-proaches either perform costly search over noise pools, with no guaranteeof finding a truly prompt-compatible noise within the limited candidates,or manipulate sampled noise without simultaneously ensuring reliablesemantic injection and preservation of the Gaussian distribution. Thesecond category focuses on improving the denoising trajectory. However,these methods lack explicit mechanisms to timely diagnose and correctCorresponding author: Tao Zhang.semantic errors, and therefore fail to prevent error propagation duringgeneration. To address these limitations, we propose AnchorSteer, atraining-free framework that exerts fine-grained control over both ini-tialization and the denoising trajectory. AnchorSteer consists oftwo synergistic components: Semantic Anchoring replaces uninfor-mative Gaussian noise with text-aligned initializations via CLIP-basedprior extraction and a novel Latent-Prior Score Distillation Sampling(LP-SDS) objective. Specifically, LP-SDS distills CLIP visual priors intothe knowledge distribution of diffusion models, mitigating the domaingap between CLIP-based priors and diffusion-based priors. ReflectiveSteering transforms passive denoising with an active Think–Erase–Retouch loop that enables mid-generation self-correction. Think phaseemploys VLM-based dual diagnosis to detect semantic deviations. Erasephase performs a targeted latent rollback with negative-guided inversionto suppress erroneous content, and Retouch phase subsequently appliespositive-guided refinement to recover the missing attributes. Therefore,Reflective Steering enables timely correction of semantic deviationsalong the denoising trajectory, preventing error propagation. Extensiveexperiments on GenEval and T2I-CompBench++ demonstrate that An-chorSteer consistently outperforms existing baselines in text–image align-ment while preserving high visual quality.

With an increasing number of Object Detection (OD) modelsbeing deployed on edge devices, Zero-Shot Quantization for OD (ZSQ-OD) aims to quantize these models when access to the original trainingdata is prohibited. Existing research on Zero-Shot Quantization-AwareTraining (QAT) for OD synthesizes training sets through noise optimiza-tion. However, this approach struggles to maintain performance in low-bit regions. In this paper, we introduce GoodQ (Generative off-the-shelfmodels for object detector Quantization), a QAT pipeline that utilizesoff-the-shelf generative models to construct a training set. We first iden-tify three challenges that arise when introducing a generative model tothe ZSQ-OD task: 1) each image contains dense information with mul-tiple instances, 2) the class-wise distribution in the original dataset isimbalanced, and 3) the pseudo-labels assigned to the generated imagescan potentially act as noisy signals during QAT. GoodQ addresses thesechallenges by 1) introducing an Information-Dense Prompting strategyto generate multi-instance images, 2) applying Intrinsic Distribution-Aware Selection to match the pretrained class distribution, and 3) em-ploying Teacher-guided Adaptive Noise Reduction to mitigate noise aris-ing from the QAT process. Our framework achieves state-of-the-art per-formance in low-bit ZSQ (W4A4) and extends quantization to extremebit-widths (W3A3). Furthermore, we conduct an extensive analysis touncover the underlying factors contributing to the efficacy of GoodQ.The constructed dataset is available at GoodQ.

Action chunking has recently emerged as a standard prac-tice in flow-based Vision-Language-Action (VLA) models. However, thee!ect and choice of the execution horizon—the number of actions tobe executed from each predicted chunk—remains underexplored. In thiswork, we first show that varying the execution horizon leads to sub-stantial performance deviations, with performance initially improvingand then declining as the horizon increases. To uncover the reasons, weanalyze the cross- and self-attention weights in flow-based VLAs and re-veal two key phenomena: (i) intra-chunk actions attend invariantly tovision–language tokens, limiting adaptability to environmental changes;and (ii) the initial and terminal action tokens serve as stable anchors,forming latent centers around which intermediate actions are organized.Motivated by these insights, we interpret action self-attention weights asa proxy for the model’s predictive limit and propose AutoHorizon, thefirst test-time method that dynamically estimates the execution horizonfor each predicted action chunk to adapt to changing perceptual condi-tions. Across simulated and real-world robotic manipulation tasks, Au-toHorizon is performant, incurs negligible computational overhead, andgeneralizes across diverse tasks and flow-based models. Demonstrationvideos are available at this project page.

The inherent complexity of video understanding makes it dif-ficult to determine whether Video-LLM benchmark performance stemsfrom visual perception, linguistic reasoning, or knowledge priors. Whilemany benchmarks have emerged to assess high-level reasoning, sharedcriteria for evaluating video understanding remain largely overlooked.Instead of introducing yet another benchmark, we take a step back tore-examine the criteria for evaluating video understanding. In this work,we introduce Video-Oasis, a sustainable diagnostic suite for systemati-cally auditing existing video understanding benchmarks. This audit re-veals that 55% of existing benchmark samples are solvable without visualinput or temporal context. After filtering these shortcuts, the remainingvideo-native challenges expose a substantial capability gap: state-of-the-art models perform only marginally above random guessing. Building onthese findings, we use the distilled challenges as a testbed to investigatewhich algorithmic design choices contribute to robust video understand-ing. We hope our work provides a practical foundation for constructingrigorous video benchmarks and evaluating future Video-LLMs. Code isavailable at https://github.com/sejong-rcv/Video-Oasis.

The domain gap between abstract sketches and natural im-ages, combined with the challenge of zero-shot generalization, remainsa critical bottleneck in Sketch-Based Image Retrieval (ZS-SBIR). Whilescaling up to large-scale Vision-Language Models offers powerful baselinerepresentations, naive utilization—such as relying solely on the terminal[CLS] token—fails to fully exploit their potential, often discarding thefine-grained structural primitives essential for abstract sketch alignment.In this paper, we propose a robust framework that bridges this modalitygap by decoupling and composing visual states and semantics. First, ourMulti-Layer Feature Aggregation (MLFA) acts as an anti-compressionmechanism, explicitly reviving low-level geometric primitives alongsidehigh-level semantic objects. Second, a Phased Cross-Modal Interaction(PCMI) serves as a dynamic compositional engine: learnable promptsprogressively query and bind these structural states with global seman-tics for precise coarse-to-fine alignment. Extensive experiments demon-strate our framework achieves state-of-the-art results on several bench-marks and remains competitive on QuickDraw Ext.

Despite the remarkable prowess of Vision-Language Models(VLMs) in general multimodal tasks, they remain fundamentally “flat”when reasoning about the physical world. We argue that this spatialbottleneck stems from a profound dimensional mismatch: while VLMsare trained to interpret 2D projections, true spatial reasoning demandsthe recovery of latent 3D geometry and temporal continuity. To conquerthis high-dimensional complexity, we advocate a shift from monolithiclearning to a “divide and conquer” paradigm. We present FactoSR, afactorized reinforcement learning framework that explicitly interpret the* Equal Contribution, † Project Lead, B Corresponding Authordimensions collapsed by visual projection. At its core, FactoSR decom-poses the monolithic problem of world-consistent reasoning into three or-thogonal, geometric sub-objectives: planar correspondence (XY ), depthconsistency (Z), and temporal reversibility (T ). By optimizing these ver-ifiable constraints within a unified policy learning mechanism, we effec-tively transform an ill-posed projection recovery problem into a series oftangible reasoning steps. Extensive evaluations on multi-view and videobenchmarks demonstrate that this elegant decomposition yields substan-tial gains in 3D and 4D reasoning, achieving a 5.9% boost on VSI-Benchand 4.5% on All-Angles-Bench. Our findings suggest that reinforcing ex-plicit, factorized 4D consistency is a critical step toward evolving VLMsinto robust, world-aware reasoners.Code: https://github.com/scott-yjyang/FactoSR

Atmospheric turbulence severely degrades thermal infraredimaging quality, yet most existing mitigation methods focus solely on im-age restoration while ignoring underlying physical characteristics. In thisstudy, we construct a large-scale infrared turbulence imaging dataset andproposed the Turbulence-oriented Mutual Progressive Layered Extrac-tion (TurboMPLE), a joint network for infrared turbulence mitigationand physical fields estimation. Through Mutual PLE Blocks, TurboM-PLE achieves progressive feature-level collaboration, introducing turbu-lence priors into the mitigation process while incorporating richer visualinformation into physical fields estimation. In addition, embedding theLambert–Beer law and turbulent physics constraints enables more in-terpretable results. With a lightweight and computationally efficient de-sign, TurboMPLE achieves state-of-the-art performance on both tasksand demonstrates strong generalization capability on real-world data.The code is available at https://github.com/Ayt777/TurboMPLE.

Concept erasure techniques (CETs) edit text-to-image diffusion models to erase undesired targets (e.g., NSFW content or copyrighted styles) while preserving model utility on image generation for benign concepts. Current CETs often face a trade-off between erasure robustness and model utility: stronger edits erase target more reliably, but degrade model utility on non-target concepts and vice-versa. A key source of this trade-off is how existing methods define what to erase and what to preserve. Many CETs rely on static concept banks, where erase or retain concepts are manually specified, generated by Large Language Models (LLMs), or selected using CLIP-based image–text similarity. However, such static banks do not explicitly model how prompts steer the diffusion model during denoising, leaving edited models vulnerable to triggers that can reintroduce the target while unintentionally suppressing nearby benign concepts. We present Preservation aware Adaptive Ranked Subspace Expansion (PARSE), a training-free framework for robust concept erasure in latent diffusion models. Given a target, PARSE queries the diffusion model using classifier-free guidance to dynamically discover target-inducing erase concepts and nearby retain concepts from the model vocabulary. PARSE then edits the cross-attention value space with a preservation-aware projection that removes target directions while leaving retain directions intact. To address triggers beyond the vocabulary-indexed search space, PARSE iteratively searches for re-emergence triggers using textual inversion and adaptively expands the erased subspace only when a new trigger direction does not conflict with retain semantics. We further introduce Balanced Erasure Utility Score (BEUS), a balanced erasure utility score that combines robustness (ASR under multiple attacks) and utility preservation (FID) via bounded monotone transforms and harmonic mean aggregation. Extensive experiments across NSFW, artistic style, and object erasure, together with a large-scale robustness utility analysis over many CET baselines, show that PARSE achieves robust erasure across multiple concepts without sacrificing post-edit model utility. Project page: https: //shaswati1.github.io/parse/

Recent breakthroughs of transformer-based diffusion models, particularly with Multimodal Diffusion Transformers (MMDiT) driven models like FLUX and Qwen Image, have facilitated thrilling experiences in visual generation. However, these models rely only on the interactions between textual conditions and visual features to produce semantically aligned images. Once the interactions fail to reflect the nuanced compositional structure of the prompt, the generated images might be unsatisfactory. Thus, a comprehensive understanding of how different blocks and their interactions with textual conditions is crucial for better understanding the intrinsic attributes and for enhancing their interactions accordingly to strengthen the prompts adherence. In this paper, we first develop a systematic pipeline to comprehensively investigate each block’s functionality by removing, disabling, and enhancing textual hidden-states at corresponding blocks. Our analysis reveals that 1) semantic information appears in earlier blocks and finer details are rendered in later blocks, 2) removing specific blocks is usually less disruptive than disabling text conditions, and 3) enhancing textual conditions in selective blocks improves semantic attributes. Building on these observations, we propose TexTailor, a novel inference-time method for tailoring block-wise textual guidance. Our approach not only improves text-image alignment but also enables a range of downstream applications, including precise editing and inference acceleration. Extensive experiments demonstrated that our method outperforms various baselines and remains flexible across text-to-image generation, image editing, and inference acceleration. Our method improves T2I-Combench from 56.92% to 63.00% and GenEval from 66.42% to 71.63% on SD3.5, without sacrificing synthesis quality. These results advance understanding of MMDiT models and provide valuable insights to unlock new possibilities for further improvements. Our code is available at https://github.com/phil329/TexTailor.

Multimodal large language models (MLLMs) often halluci-nate, producing fluent but visually ungrounded outputs, partly becausedirect preference optimization (DPO) overfits to superficial linguisticcues under static preference supervision. We propose TARS, a token-adaptive preference strategy that reformulates DPO as a min-max op-timization problem: the inner maximization perturbs visual-agnostic to-kens to induce worst-case distributional shifts, while the outer minimiza-tion enforces alignment with causal visual signals rather than surface-level patterns. A spectral alignment loss further regularizes hidden rep-resentations in the frequency domain via the Fast Fourier Transform(FFT). With only 4.8k preference samples and no expert feedback, TARShalves the hallucination rate (26.4% → 13.2%) and reduces the cognitionscore from 2.5 to 0.4, outperforming standard DPO. It also surpasses 5×larger LLM-based data augmentation (28.8k samples; Hal-Rate 16.0% vs.13.2%) and narrows the gap to GPT-4o, suggesting that token-adaptiveoptimization improves data efficiency beyond simply scaling augmentedpreference data.

Modern high-resolution mobile sensors often adopt clusteredcolor filter array (CFA) layouts such as Quad and Hexadeca Bayer,while some premium sensors use quad-pixel phase-detection or 2 × 2 on-chip-lens (OCL) architectures with shared microlenses for dense phase-detection autofocus (PDAF). While effective for PDAF, this structurecan introduce grid-aligned phase artifacts during demosaicing, which aredifficult to model without calibrated sub-aperture streams, PSF super-vision, or sensor-specific optical calibration. We propose Synthetic Sub-Aperture Phase Augmentation (SPA), a physics-inspired data genera-tion framework for artifact-aware demosaicing. SPA uses compact phasekernels to approximate focus-dependent sub-aperture shifts in syntheticRAW, requiring no measured point spread functions (PSFs) or sensor-specific calibration. Experiments on synthetic and real RAW data fromHexadeca and Quad Bayer sensors show consistent suppression of zipper-ing, color bleeding, and grid-aligned artifacts. We also introduce spatiallattice imbalance (SLI) and Nyquist energy ratio (NER) for 2× 2 artifactanalysis and validate robustness to spatial phase variations.

Recent Multimodal Large Language Models (MLLMs) strug-gle to bridge the representational gap between 2D semantic understand-ing and 3D spatial geometry. Existing 3D-aware models either rely oncostly 3D-specific data or utilize RGB-only inputs with heuristic sam-pling and monolithic, shallow fusion, which respectively disrupt essentialspatiotemporal connectivity and induce modality contention across di-verse spatial tasks. To overcome these bottlenecks, we introduce SpaR3D-MoE, an end-to-end framework that enables adaptive spatial reasoningby equipping MLLMs with geometry-aware capabilities from only sparseRGB inputs. First, we propose an adaptive spatiotemporal manifoldsampling mechanism that constructs a geometry-aware spatiotemporalgraph to extract informative keyframes, effectively mitigating sequenceredundancy while preserving the scene’s topological connectivity. Second,we introduce the heterogeneous geometry-inductive Mixture-of-Expertsdriven by an instruction-pose aware router, which adaptively routes mul-timodal tokens to specialized experts, resolving the cross-modal con-tention inherent in monolithic fusion. Extensive experiments on VSI-Bench, ScanQA, and SQA3D demonstrate that our method achievesstate-of-the-art performance. Notably, SpaR3D-MoE achieves the highestaverage score of 63.5 on VSI-Bench, outperforming the strongest base-line by 7.8 absolute points, alongside relative improvements of 35.4% and51.4% in Route Plan and Relative Direction tasks, respectively.