ECCV 2026 Papers

We propose UniDynamics, a diffusion-based framework forfuture 4D dynamic scenes (RGB, depth, and optical flow) generationfrom a single event-RGB pair, without requiring long histories or con-trol priors as in existing methods, while explicitly modeling future mo-tion fields. The core idea is to leverage event streams to offer an al-ternative motion prior for single-RGB extrapolation, and to enforce ge-ometric and motion constraints throughout generation via multimodalmodeling. Specifically, we design an Event Latent Enhancement (ELE)module to align and enhance event latents into diffusion-injectable con-ditioning features, providing robust initial motion priors and reliabletexture/structure cues. We further introduce a Perceptual DynamicsSpace (PDS) embedded in the multi-scale U-Net, which decouples andadaptively interacts depth and flow while continuously feeding back con-straints to appearance features, improving geometric-motion consistencyfor physically plausible and spatiotemporally coherent prediction. Exper-iments on VKitti2 and DSEC demonstrate state-of-the-art performance,producing high-quality, temporally coherent, and 4D-consistent futurepredictions, especially under challenging high-speed motion blur.

Ultra-high-resolution image generation is increasingly vital for applications requiring fine-grained textures and global structural fidelity, yet state-of-the-art text-to-image diffusion models such as FLUX and SD3 remain confined to sub 2MP (< 1K × 2K) resolutions due to the quadratic complexity of attention mechanisms and the scarcity of high-quality high-resolution training data. We present UltraImageGen, a novel framework that introduces hierarchical local attention with low-resolution global guidance, enabling efficient, scalable, and semantically coherent image synthesis at ultra-high resolutions. Specifically, high-resolution latents are divided into hardware aligned fixed-size local windows to reduce attention complexity from quadratic to near-linear, while a low-resolution latent with scaled positional embeddings injects global semantics as an anchor. A lightweight LoRA adaptation bridges global and local pathways during denoising, ensuring consistency across structure and detail. To maximize efficiency and achieve scalable ultrahigh-resolution generation, we repermute token sequence in window-first order, so that the GPU-friendly dense local blocks in attention calculation equals to the fixed-size local window in 2D regardless of resolution. Together UltraImageGen reliably scales pretrained models to resolutions higher than 8K with more than 10× speed up and significantly lower memory usage. Extensive experiments demonstrate that UltraImageGen achieves superior quality while maintaining computational efficiency, establishing a practical paradigm for advancing ultra-high-resolution image generation. Our project page and code are available here.

We present a dense RGB-D SLAM system using differen-tiable triangles as the 3D map representation. While 3D Gaussian Splat-ting has emerged as the leading method for novel-view synthesis, trianglesremain the standard primitive for traditional rendering hardware, gameengines, and downstream tasks requiring explicit geometry such as simu-lation, collision, and editing. Recent offline methods have demonstratedthat an unstructured ‘triangle soup’ can be optimised into a photorealis-tic mesh via Delaunay triangulation across a set of posed images. Build-ing upon this insight, we present the first dense SLAM system to employTriangle Splatting to perform both tracking and mapping through on-line differentiable rendering of a triangle soup. The map can be convertedinto a connected mesh on-the-fly via restricted Delaunay triangulation,enabling new online capabilities such as mesh deformation and collisionchecking. On Replica and TUM-RGBD, our system outperforms base-lines on 3D geometry, matches the camera-tracking accuracy, and enablesonline mesh-based scene editing.

Bio-inspired Small Target Motion Detection (STMD) mod-els often suffer “false alarm collapse” due to non-stationary environ-mental interference, failing to adapt to dynamic background statistics.We propose a bio-inspired architecture integrating macro-motion decou-pling with micro-scale Adaptive Gain Control (AGC). Macroscopically, itleverages lobula plate tangential cells (LPTC)-based population codingto eliminate ego-motion-induced baseline drifts in real-time with mini-mal overhead. Microscopically, we introduce an adaptive shunting inhi-bition operator inspired by Natural Scene Statistics (NSS). By quan-tifying local temporal variance, this mechanism maps non-stationaryclutter into a gain control factor to suppress heavy-tailed noise fromhigh-frequency flickers. This versatile plug-and-play module integratesseamlessly into existing STMD frameworks. Extensive experiments onsynthetic and real-world datasets demonstrate that our model signifi-cantly reduces false alarms while maintaining high sensitivity for dimtargets, achieving state-of-the-art precision and robustness in complexdynamic environments. Our code is publicly available at AGC.

True video intelligence demands more than recognizing whatis visible: it requires reasoning about why events unfold, predicting whatwould change under different conditions, and deciding what to do next.We refer to this full progression—from perception through causal reason-ing and simulation to strategic planning—as Strategic Video Intelligence(SVI). No existing benchmark evaluates this capability stack: in-the-wildvideos lack verifiable ground truth for causal and strategic questions,while synthetic environments sacrifice the complexity of real multi-agentsystems. To bridge this gap, we introduce SVI-Bench, a large-scale bench-mark that leverages team sports as a dynamic microworld, a domain thatuniquely combines the complexity of real-world multi-agent interactionwith the verifiability of explicit rules and definitive outcomes. SVI-Benchcomprises ∼35K hours of broadcast video, ∼15M annotated actions,∼15K hours of expert commentary, ∼23K game reports, and ∼103Kstructured statistical records across basketball, soccer, and hockey, allconstructed via a data engine that transforms raw game data into adense, cross-referenced corpus. We organize evaluation into 9 tasks span-ning a progressive four-pillar hierarchy: Dynamic Scene Understanding,Causal Reasoning, Strategic Simulation, and Agentic Synthesis. Evaluat-ing strong multimodal and agentic baselines, we find a capability cliff :models perform competently on perceptual tasks (achieving ∼74% onfine-grained action QA) but degrade sharply at higher levels of the stack.Agentic tasks prove hardest of all: the strongest model achieves only 5%accuracy when required to autonomously gather and integrate evidenceacross a corpus of 1.8M clips. We release the full benchmark to catalyzeprogress toward AI systems capable of strategic intelligence in complex,dynamic multi-agent environments.§ Code github.com/texaser/svi-benchõ Data huggingface.co/mvp-group/svi-bench€ Website svi-bench.github.ioA Extended Paper svi-bench.github.io/svi_bench_extended.pdfPillar 1: Perceptionparsing who is where, doing what and whenPillar 2: Reasoningexplaining why something happenedPillar 3: Simulationgenerating plausible alternative futuresPillar 4: Agencyautonomous analysis over multimodal evidenceThe Strategic Video Intelligence StackFig. 1: Overview of SVI-Bench, illustrated through a single play from the 2022NCAA Final Four. SVI-Bench is the first large-scale video benchmark evaluatingthe full SVI stack: Perception (describing what happens), Reasoning (explaining why),Simulation (generating plausible alternatives), and Agency (autonomous analysis).

Skeleton-based Video Anomaly Detection (VAD) offers a ro-bust, privacy-preserving solution for identifying abnormal behaviors. Tomodel the distribution of normal static and moving poses, recent meth-ods train Energy-Based Models (EBMs) via Denoising Score Matching(DSM). However, directly injecting noise, required for training, into rawjoint coordinates creates physically impossible poses, and this structuralcollapse severely worsens as the temporal window expands. To addressthis, we introduce STEP, a simple framework that utilizes PrincipalComponent Analysis (PCA) to project pose sequences into a compact,whitened PC-space. Learning the data density within this well-behavedPC-space ensures that the injected noise translates into physically plau-sible variations, which allows the model to process longer video sequenceswithout the performance collapse of raw coordinate baselines. Addition-ally, to mitigate inherent pose estimation inaccuracies arising from occlu-sions or motion blur, we integrate a sequence-level weighting mechanismbased on the estimator’s confidence scores. Operating at real-time com-putational efficiency, our simple and lightweight framework outperformsthe previous skeleton-based state-of-the-art by 12.2% (90.1% AUROC)on the challenging UBnormal dataset and achieves highly competitiveresults by improving on the ShanghaiTech benchmark.

Multi-shot video generation is crucial for long narrative storytelling, yet current bidirectional architectures suffer from limited interactivity and high latency. We propose ShotStream, a novel causal multi-shot architecture that enables interactive storytelling and efficient on-the-fly frame generation. By reformulating the task as next-shot generation conditioned on historical context, ShotStream allows users to dynamically instruct ongoing narratives via streaming prompts. We achieve this by first fine-tuning a text-to-video model into a bidirectional nextshot generator, which is then distilled into a causal student via Distribution Matching Distillation. To overcome the challenges of inter-shot consistency and error accumulation inherent in autoregressive generation, we introduce two key innovations. First, a dual-cache memory mechanism preserves visual coherence: a global context cache retains conditional frames for inter-shot consistency, while a local context cache holds generated frames within the current shot for intra-shot consistency. And a RoPE discontinuity indicator is employed to explicitly distinguish the two caches to eliminate ambiguity. Second, to mitigate error accumulation, we propose a two-stage distillation strategy. This begins with intra-shot self-forcing conditioned on ground-truth historical shots and progressively extends to inter-shot self-forcing using self-generated histories, effectively bridging the train-test gap. Extensive experiments demonstrate that ShotStream generates coherent multi-shot videos with sub-second latency, achieving 16 FPS on a single GPU. It matches or exceeds the quality of slower bidirectional models, paving the way for real-time interactive storytelling. Code and models will be released.

Recent pre-trained text-to-image flow models have enabledremarkable progress in text-based image editing. Mainstream approachesadopt a corruption-then-restoration paradigm, where the source im-age is first corrupted into an editable “intermediate state” and then re-stored to the target image under the prompt guidance. However, currentmethods construct this intermediate state in a target-agnostic manner,i.e., they mainly focus on realizing source image reconstruction whileneglecting the semantic gaps towards the specific editing target. Thisdesign inherently results in limited editability or inconsistency when thedesired modifications substantially deviate from the source. In this pa-per, we argue that the intermediate state should be target-aware, i.e.,selectively corrupting editing-relevant contents while preserving editing-irrelevant ones. Thus, we propose FlowCycle, an inversion-free and flow-based editing framework that parameterizes corruption with learnablenoises and optimizes them through a cycle-consistent process. By itera-tively editing the source to the target and recovering back to the sourcewith dual consistency constraints, FlowCycle learns to produce a target-aware intermediate state, enabling faithful modifications while preservingsource consistency. For efficiency, we further accelerate the optimizationby dynamically adjusting the sampling steps. Extensive ablations demon-strated that FlowCycle achieves superior editing performance.

EgoExo proficiency estimation aims to assess action qual-ity by integrating fine-grained motion cues from egocentric (1st-person)views with spatial context from multiple exocentric (3rd-person) views.Simply adding more exocentric views degrades EgoExo performance, asredundant or noisy perspectives dilute useful motion cues. Our anal-ysis identifies two key causes: (1) Multiview redundancy — Fromthe data perspective, certain views provide limited or noisy informa-tion, diluting discriminative cues; (2) Overfitting — From the featureperspective, conventional fusion increases representational complexity,causing the model to memorise view-specific patterns rather than learngeneralisable representations. To address these issues, we propose twocomplementary modules. AdaMVS adaptively identifies and fuses themost informative view tokens under weak supervision from the data per-spective, while VIB-GB combines Gradient Blending and VariationalInformation Bottleneck regularisation from the feature perspective tocompress redundant signals and suppress overfitting during training.Experiments on EgoExo-4D and EgoExo-Fitness demonstrate that ourmethod learns both which view to look at and how to fuse them,achieving new state-of-the-art results. Our source code is available athttps://github.com/dx199771/AdaMVS

Unified multimodal models (UMMs) strive to consolidate visual understanding and visual generation within a single architecture. However, prevailing training paradigms independently optimize understanding via sparse text signals and generation through dense pixel objectives. Such a decoupled strategy yields misaligned representation spaces, isolating visual understanding from generation and hindering their mutual reinforcement. This work presents the first systematic investigation into generative post-training, where we formulate hierarchical visual tasks as generative proxies to bridge the isolation in UMMs. Our empirical investigation reveals that high-level semantic tasks, particularly image segmentation, serve as optimal proxies. Unlike low-level tasks that distract models with texture details, segmentation provides structural semantics that significantly enhance both vision-centric perception and generative layout fidelity. Building upon these insights, we introduce Semantic Generative Tuning (SGT), a novel paradigm that leverages segmentation as a generative proxy to align and synergize multimodal capabilities. Mechanistic analyses further demonstrate that SGT fundamentally improves feature linear separability and optimizes visualtextual attention allocation pattern. Extensive evaluations show that SGT consistently improves both multimodal comprehension and generative fidelity across mainstream benchmarks. Our code is available on the Project Page.

Despite rapid advances in chest X-ray (CXR) foundationmodels, most radiology report generation (RRG) systems still rely onheavily downsampled inputs (e.g., 256×256) due to the fixed visual tokenbudgets of pretrained vision encoders, suppressing subtle yet clinicallyimportant cues present in native-resolution images. However, enablinghigh-resolution (high-res) perception remains challenging: naïve tilingcauses prohibitive token inflation, while global compression suppressessubtle lesions and degrades diagnostic fidelity. Inspired by radiologists’workflow, localizing suspicious regions before detailed high-res assess-ment. We propose Lesion-Aware High-Resolution Patch Discovery andFusion for Chest X-ray Reporting (LePaX), the first RRG frameworkthat enables efficient high-res CXR perception (up to 1920×1920) with-out increasing the vision-token count. LePaX formulates high-res percep-tion as a constrained spatial resolution allocation problem under a fixedtoken budget and introduces two key components: Learnable Spatial Res-olution Allocation (LSRA), which learns a spatial utility map that adap-tively allocates limited high-res capacity to diagnostically relevant re-gions, enabling targeted extraction of high-res patches from native CXRs;and Global–Regional Fusion (GRF), which performs token-preservingregion-to-global refinement by projecting high-resolution regional evi-dence back onto the global feature grid through spatially aligned resolu-tion write-back, avoiding token inflation. Experiments on multiple CXRbenchmarks demonstrate that LePaX consistently improves both clinicaland linguistic metrics while enabling native-resolution CXR perceptionwith over 10× fewer visual tokens than naïve high-res tiling.

Cross-modal super-resolution (SR) on real-world misaligneddata is challenging, as only unlabeled low-resolution (LR) source andhigh-resolution (HR) guide images with complex spatial misalignmentare available. Previous methods either rely on simulated training dataor adopt suboptimal alignment strategies that overlook cross-modal de-pendencies, limiting their practical performance. To address these is-sues, we propose RobSelf, a self-supervised model that jointly optimizesa misalignment-aware feature translator and a content-aware referencefilter online. The translator resolves unsupervised cross-modal and cross-resolution alignment via weakly-supervised, misalignment-aware trans-lation, yielding an aligned guide feature. Guided by this feature, thefilter performs reference-based discriminative self-enhancement on thesource, enabling SR prediction with high resolution and high fidelity.Experiments on synthesized data and collected real-world data demon-strate that RobSelf achieves state-of-the-art performance, outperformingexisting self-supervised and supervised methods. Moreover, it achievessuperior efficiency, being up to 15.3× faster than prior self-supervisedmethods. https://github.com/palmdong/RobSelf

Reinforcement-learning alignment for autoregressive imagegeneration is bottlenecked by expensive rollouts: in GRPO-style groupoptimization, sampling G rollouts per prompt improves exploration andwithin-group credit estimation but multiplies sampling cost. This chal-lenge is amplified by vector-quantized (VQ) tokenization, where nearbycodebook IDs in embedding space can be locally substitutable; however,GRPO-style updates provide explicit positive credit only to sampled IDs,leading to insufficient coverage of success-supported alternatives undersmall-G training. We propose GR-GRPO (Graph-Regularized GRPO),which densifies token-level supervision without additional rollouts by dif-fusing positive evidence from positive rollouts over a precomputed code-book K-NN graph to construct graph-diffused soft targets, and regulariz-ing the policy toward these targets via an auxiliary cross-entropy term. Aconfidence-adaptive gate modulates diffusion strength based on the pol-icy’s logit margin, and a difficulty-aware budget allocation scheme fur-ther improves rollout utilization across prompts. Extensive experimentson compositional image generation (GenEval) and preference-style align-ment (DrawBench) validate the effectiveness of GR-GRPO. In particular,GR-GRPO achieves a substantially better compute–performance trade-off, surpassing large-group training (G=64) with ∼ 4.6× less rollout timeon Janus-Pro-1B. Our code and model weights are available athttps://github.com/vivoCameraResearch/GR-GRPO.

Unified anomaly detection requires modeling highly heteroge-neous normal data without access to anomalous samples. While foundationmodels like DINOv2 provide rich token representations, leveraging thesespaces for explicit density estimation remains challenging. Energy-BasedModels (EBMs) offer a principled formulation, but their training inhigh-dimensional token spaces is unstable due to anisotropy and strongcross-dimensional correlations, which degrades finite-step Markov ChainMonte Carlo (MCMC) sampling. We identify this instability as fundamen-tally geometric and introduce ReFP-AD (Rectified Flow Preconditioningfor Anomaly Detection), which learns a geometric reparameterization thatmaps high-dimensional embeddings into a well-conditioned latent spacevia an optimal transport (OT)-coupled rectified flow. This precondition-ing enables stable persistent contrastive divergence with preconditionedStochastic Gradient Langevin Dynamics (SGLD) in full-dimensional tokenspaces. Anomaly scores are then derived from the learned energy landscapeusing gradient norms. Under a strict unified protocol on the MVTec-ADand VisA datasets, ReFP-AD achieves 98.6%/97.9% Image/Pixel AUROCon MVTec-AD and 97.3%/99.0% on VisA, outperforming prior unifiedEBM baselines by up to +10.8% in Image AUROC. Ablation experimentsdemonstrate that geometric reparameterization is critical for finite-stepMCMC and accurate anomaly localization in high-dimensional tokenspaces. Code is available at https://github.com/CLendering/ReFP-AD

CLIP and its variants are widely adopted visual backbonesin multimodal systems, but their pretraining remains dominated by de-scriptive image-text alignment. As downstream applications increasinglydemand visually grounded commonsense inference and compositionalreasoning, it remains unclear whether CLIP-style encoders can supportsuch reasoning without architectural changes. To address this, we presentReasonCLIP-58M, a continual pretraining framework that integrates large-scale reasoning supervision into CLIP-style models through our two-stage strategy, which progressively integrates reasoning signals while pre-serving descriptive alignment, followed by category-structured reason-ing supervision. To support this framework, we construct two comple-mentary datasets and a benchmark: ReasonLite-42M, with open-form,visually verifiable reasoning captions; ReasonPro-16M, with category-specific reasoning supervision; and RCLIP-Bench for diagnostic evalua-tion of visually grounded reasoning. We train a family of ReasonCLIPthat improves visually grounded commonsense and compositional reason-ing while also enhancing zero-shot retrieval performance. As a drop-invisual encoder for multimodal large language models such as LLaVA-NeXT, ReasonCLIP delivers consistent gains without additional infer-ence cost, demonstrating that structured reasoning supervision enhancesthe expressive capacity of CLIP-style visual representations. All datasets,models, and training code are available at https://github.com/RISys-Lab/ReasonCLIP.

Robust 3D object detection in adverse weather conditions is challenging due to sensor limitations. Although combining complementary modalities such as LiDAR and 4D RADAR has shown promise, the sparsity of these sensors becomes apparent in adverse weather with reduced reflections, leading to objects with few or no point cloud returns. To address this limitation, camera sensors provide visual cues even when LiDAR and RADAR signals are weakened. However, cameras themselves are also vulnerable to adverse weather, where some regions become unreliable due to snow or rain occluding the camera lens. While some camerafusion methods designed for adverse weather learn to weigh image regions via confidence maps, these maps receive no direct supervision and are learned solely through the detection loss. We introduce Reliability-Aware Fusion (RAF), which explicitly supervises per-pixel reliability estimation and provides a direct learning signal for identifying and suppressing unreliable visual cues. Our framework leverages pretrained LiDAR–RADAR networks, keeping their backbones frozen while only training the added camera branch, BEV fusion encoder, and detection head. Extensive experiments on the K-Radar and VoD datasets demonstrate that integrating RAF consistently improves detection accuracy over LiDAR–RADAR baselines, achieving up to +6.5 AP and +7.4 AP gains. Code is BEV 3D available at https://github.com/parkie0517/RAF.

Real-world visual perception demands joint handling of low-level image degradation and high-level semantic understanding, yet thesetwo objectives inherently conflict: restoration seeks fine-grained texturerecovery while segmentation prioritizes semantic contrast. We proposethe nested unfolding network (NUN), a principled framework that re-solves this conflict by nesting one deep unfolding network (DUN) in-side another. NUN embeds a degradation-resistant unfolding network(DeRUN) within each stage of a segmentation-oriented unfolding network(SODUN), enabling both tasks to optimize in their own subspaces whileinteracting in a controlled manner. DeRUN handles unknown degrada-tion through proximal gradient unfolding with learnable operators thatimplicitly approximate it, while SODUN performs reversible foreground-background estimation. A bi-directional unfolding interaction mechanismuses IQA to select optimal DeRUN outputs, and a cross-stage consis-tency loss ensures robust predictions under varying restoration quality.Theoretically, under local assumptions, we show that NUN reduces di-rect parameter-level gradient conflict through disjoint parameter setsand achieves a degradation-sensitivity bound, where degradation affectssegmentation only through the inner-loop optimization and restorationapproximation errors. We instantiate NUN on concealed object segmen-tation and demonstrate consistent superiority over SOTA alternativesacross 12 benchmarks. The code is available at https://github.com/ChunmingHe/NUN.

Large Vision-Language Models (LVLMs) have shown remark-able potential across a wide array of vision-language tasks, leading totheir adoption in critical domains such as finance and healthcare. How-ever, their growing deployment also introduces significant security andprivacy risks. Malicious actors could potentially exploit these models toextract sensitive information, highlighting a critical vulnerability. Recentstudies show that LVLMs often fail to consistently refuse instructionsdesigned to compromise user privacy. While existing work on privacyprotection has made meaningful progress in preventing the leakage ofsensitive data, they are constrained by limitations in both generalizationand non-destructiveness. They often struggle to robustly handle unseenprivacy-related queries and may inadvertently degrade a model’s per-formance on standard tasks. To address these challenges, we introduceNeural Gate, a novel method for mitigating privacy risks through neuron-level model editing. Our method improves a model’s privacy safeguardsby increasing its rate of refusal for privacy-related questions, cruciallyextending this protective behavior to novel sensitive queries not encoun-tered during the editing process. Neural Gate operates by learning a fea-ture vector to identify neurons associated with privacy-related conceptswithin the model’s representation of a subject. This localization then pre-cisely guides the update of model parameters. Through comprehensiveexperiments on MiniGPT and LLaVA, we demonstrate that our methodsignificantly boosts the model’s privacy protection while preserving itsoriginal utility. The code is available at https://github.com/Xiangkui-Cao/Neural-Gate.

Visual encoding followed by token condensing has becomethe standard architectural paradigm in multi-modal large language mod-els (MLLMs). Many recent MLLMs increasingly favor global native-resolution visual encoding over slice-based methods. To investigate thistrend, we systematically compare their behavior on vision-language un-derstanding and attention patterns, revealing that global encoding en-hances overall capability but at the expense of greater computationaloverhead. To address this issue, we present NaVLM-PVC, an MLLM cen-tered upon our proposed Progressive Visual Compression (PVC) method,which can be seamlessly integrated into standard Vision Transformer(ViT) to enable efficient native-resolution encoding. The PVC approachconsists of two key modules: (i) refined patch embedding, which supportsflexible patch-size scaling for fine-grained visual modeling, (ii) windowedtoken compression, hierarchically deployed across ViT layers to progres-sively aggregate local token representations. Jointly modulated by thesetwo modules, a widely pretrained ViT can be reconfigured into an effi-cient architecture while largely preserving generality. Evaluated acrossextensive benchmarks, the transformed ViT, termed ViT-PVC, demon-strates competitive performance with MoonViT while reducing TTFT(time-to-first-token) by 2.4×, when developed within an identical MLLMarchitecture. Building upon ViT-PVC, NaVLM-PVC also achieves com-petitive performance to Qwen2-VL, while further reducing TTFT by1.9×. We will release all code and checkpoints to support future researchon efficient MLLMs.