ECCV 2026 论文列表

Lensless imaging enables compact and versatile computa-tional cameras by replacing bulky optics with thin coded elements. How-ever, reconstruction from the resulting measurements is challenging: large-footprint point-spread functions (PSFs) produce highly multiplexed ob-servations, making inversion severely ill-conditioned and sensitive to cal-ibration errors and model mismatch. While deep learning approaches,including hybrid models that incorporate physics priors, have shownpromise, explicitly maintaining data fidelity throughout the network hier-archy remains difficult. Here, we propose the Integrated Forward–InverseNetwork (IFIN), a physics-guided architecture that interleaves differen-tiable forward projections with learnable inverse updates at every scale,enabling complementary cues to be exploited jointly in the measurementand image domains. This bidirectional coupling supports progressive,physics-consistent refinement and permits system-constrained PSF ker-nel adaptation under model uncertainty. On challenging lensless bench-marks, including a newly introduced dataset, IFIN achieves state-of-the-art reconstruction quality. We further observe competitive performanceon Gaussian deblurring and simulated inline holography reconstruction,suggesting that the same interleaving principle can extend beyond lens-less cameras.

Diffusion Transformers (DiTs) have emerged as a dominantarchitecture for high-quality text-to-image generation, yet their scaleposes challenges for efficient deployment. While truncated singular valuedecomposition (SVD) is a principled tool for parameter reduction, evi-dence from large language models (LLMs) suggests that naive low-rankapproximation can cause catastrophic failure. In contrast, we find thattruncated SVD in DiTs produces smooth degradation even under sub-stantial global compression, with redundancy distributed across projec-tion matrices throughout the whole network rather than concentratedin a few transformer blocks. Building on these insights, we introduceSVDtrunc, a two-step block-level compression scheme, first allocatingranks across blocks and compressing the least important ones via trun-cated SVD under a global parameter budget, and then fine-tuning allblocks with modular knowledge distillation and a rectified-flow objec-tive. We apply SVDtrunc to FLUX.dev across compression levels rangingfrom 40–90 % of the original parameter count. Across three benchmarks,GenEval, HPSv2, and DPG, we outperform all competing approaches.Notably, and in contrast to prior work, we retain near-full performanceat 68 % and remain competitive even at 57 % of the original parameterbudget. Furthermore, we show that SVDtrunc complements step distil-lation and achieves strong results even without fine-tuning, positioningit as a practical continuation of efficiency improvements beyond diffusionstep reduction for large-scale generative models.

Automated layout generation powers a wide range of designs, such as posters, banners, and user interfaces. However, current single-shot generators overlook the iterative nature of human design and often produce aesthetically subpar compositions. We present i-Design, a unified framework that models layout generation as a progressive process, placing k elements at a time, conditioned on previously rendered partial canvases. Our approach combines Progressive Imitation Learning (PIL) on partial canvases, followed by Progressive Aesthetic Policy Optimization (PAPO), a reinforcement learning algorithm that aligns generation with aesthetic preferences. In PAPO, a frozen vision judge evaluates rendered layout pairs, constructs a directed comparison graph, and computes a global aesthetic consensus to derive trajectory-level rewards reflective of visual appeal rather than spatial accuracy. During training, multiple rollouts are rendered through a non-differentiable Skia engine and scored in real time by the judge to enable adaptive aesthetic learning. Experiments on Crello and WebUI benchmarks demonstrate that i-Design surpasses SoTA methods, achieving a 13% improvement in win rate and a 7% gain in mIoU, producing layouts with superior balance, alignment, and visual harmony. Our findings demonstrate that iterative generation and progressive aesthetic optimization offer a powerful new paradigm for human-aligned graphic layout generation.

Chain-of-Thought (CoT) reasoning significantly elevates thecomplex problem-solving capabilities of multimodal large language mod-els (MLLMs). However, adapting CoT to vision typically discretizes sig-nals to fit LLM inputs, causing early semantic collapse and discard-ing fine-grained details. While external tools can mitigate this, theyintroduce a rigid bottleneck, confining reasoning to predefined oper-ations. Although recent latent reasoning paradigms internalize visualstates to overcome these limitations, optimizing the resulting hybriddiscrete-continuous action space remains challenging. In this work, wepropose HyLaR (Hybrid Latent Reasoning), a framework that seamlesslyinterleaves discrete text generation with continuous visual latent rep-resentations. Specifically, following an initial cold-start supervised fine-tuning (SFT), we introduce DePO (Decoupled Policy Optimization) toenable effective reinforcement learning within this hybrid space. DePOdecomposes the policy gradient objective, applying independent trust-region constraints to the textual and latent components, alongside anexact closed-form von Mises-Fisher (vMF) KL regularizer. Extensive ex-periments demonstrate that HyLaR outperforms standard MLLMs andstate-of-the-art latent reasoning approaches across fine-grained percep-tion and general multimodal understanding benchmarks. Code is avail-able at https://github.com/EthenCheng/HyLaR.

Handwritten mathematical expression recognition (HMER)requires reasoning over diverse symbols and structures, yet autoregressivemodels struggle with exposure bias and syntax inconsistency. We presentGryphOne, a discrete diffusion framework which reformulates HMER asiterative symbolic refinement instead of sequential generation. GryphOneprogressively refines symbols and relations, removing autoregression andimproving consistency. Symbol-aware tokenization and random-maskingmutual learning further enhance robustness to handwriting diversity. Onthe MathWriting benchmark, GryphOne achieves 5.51% CER and 59.9%EM (ExpRate), outperforming all reimplemented models in the matchedsetting as well as the commercial HMER system. Held-out evaluation onCROHME 2014–2023 further shows strong cross-dataset generalization.

Medical image visualization requires volumetric rendering al-gorithms that preserve anatomical fidelity while maintaining high ren-dering speeds. To address the high computational cost of large volumet-ric datasets, we propose a Gaussian-based volumetric representation forefficient visualization of dense medical volumes without compromisingstructural and radiometric details. We optimize the proposed represen-tation using Monte Carlo volumetric estimation, which enables trainingon a highly sparse subset of voxels while maintaining consistency withthe dense volumetric objective. In addition, we introduce a curriculumlearning strategy that progressively incorporates structured slice-basedsampling during training. Sparse voxel samples provide an early globalcoverage of the volume, while slice samples capture spatially correlatedregions that aid geometric structure and texture continuity. This combi-nation enables the Gaussian representation to learn anatomical detailsof various structures and corresponding textures from sparse supervisionwhile significantly reducing the computational cost associated with densevoxel processing. The learned representation supports slice-based ren-dering methods such as shear–warp volume rendering, enabling efficientvisualization of multimodal medical datasets including MRI and Cryosec-tion volumes while preserving anatomical structures. Using sparse super-vision, our method achieves up to 43.86 FPS rendering with a compres-sion ratio of 11.31:1.

Small object detection (SOD) remains a challenging task inreal-world applications. Despite recent advances, existing detectors re-main limited by rigid processing that entangle spatial aggregation withimplicit frequency aliasing and truncation, leading to inadequate preser-vation of high-frequency components for SOD. To tackle these limita-tions, we propose a Frequency-Spatial Domain Collaborative DetectionTransformer (FSDC-DETR), a novel collaborative framework that ex-plicitly models complementary spatial and frequency representations.Specifically, we first introduce Dual-Branch Frequency-Spatial AdaptiveFusion (DBFSAF) to enhance frequency diversity and adaptively cap-ture frequency-spatial domain discriminative representations. Buildingon these representations, a frequency-spatial interaction scheme is fur-ther explored within the hybrid encoder to enable progressive featurepropagation to the decoder. In particular, structure-aware frequency-spatial aggregation is achieved through Shunt Frequency-Spatial Fea-ture Fusion (SFS-FF), establishing bidirectional interaction and pro-gressive cross-scale propagation between frequency and spatial repre-sentations for coherent discriminative modeling. Meanwhile, informativehigh-frequency responses are preserved during scale transitions throughFrequency-Spatial Dynamic Downsampling (FSD-Down), thereby mini-mizing frequency degradation throughout multi-scale fusion for the pre-cise SOD. Experimental results demonstrate that FSDC-DETR achievesstate-of-the-art performance, improving AP by 6.4 on VisDrone-DET2019and 6.6 on AITODv2, with gains of 6.8 and 6.9 AP for small objects. Thecode is available at https://github.com/nevereverinsomnia/FSDC-DETR.

Vision-Language Models (VLMs) create a severe visual fea-ture bottleneck by using a crude, asymmetric connection that links onlythe output of the vision encoder to the input of the large language model(LLM). This static architecture fundamentally limits the ability of LLMsto achieve comprehensive alignment with hierarchical visual knowledge,compromising their capacity to accurately integrate local details withglobal semantics into coherent reasoning. 5 To resolve this, we introduceCross-Layer Injection (CLI), a novel framework that forges a dy-namic “many-to-many” bridge between the two modalities. CLI con-sists of two synergistic, parameter-efficient components: an AdaptiveMulti-Projection (AMP) module that harmonizes features from di-verse vision layers, and an Adaptive Gating Fusion (AGF) mech-anism that empowers the LLM to selectively inject the most relevantvisual information based on its real-time decoding context. We validatethe effectiveness and versatility of CLI by integrating it into LLaVA-OneVision and LLaVA-1.5. Extensive experiments on 28 diverse bench-marks demonstrate significant performance improvements, establishingCLI as a scalable paradigm that unlocks deeper multimodal understand-ing by granting LLMs on-demand access to the full visual hierarchy. Codeis available at https://github.com/codefuse-ai/CLI.

Large view synthesis models synthesize novel views throughcross-view attention without explicit 3D representations, and recent stud-ies have shown that they learn accurate spatial correspondence fromRGB supervision alone. We observe that this correspondence generalizesbeyond appearance. When non-photorealistic signals such as binary en-coded panoptic labels are passed through the model, they are propagatedto novel views with consistent spatial structure. These results indicatethat the correspondence learned for RGB view synthesis can also propa-gate view-independent per-pixel labels. From this observation, we studyhow large view synthesis models can be extended beyond appearancerendering to 3D scene understanding. We propose a panoptic segmen-tation pipeline that reuses a frozen view synthesis model to propagatepanoptic labels from input views to novel views, without 3D reconstruc-tion or any segmentation-specific training of the view synthesis model.Given panoptic labels on the input views, we encode them into binarychannel representations and pass them through the same model to rendertarget-view segmentation. On ScanNet, our method achieves segmenta-tion quality on par with Gaussian based approaches requiring explicit3D reconstruction, while outperforming them in novel view synthesis bymore than 7 dB. The label propagation also transfers across datasets,surpassing these approaches on Replica without any fine-tuning.

Open-Vocabulary object Detection (OVD) typically relies onVision–Language Models (VLMs) to associate visual regions with arbi-trary textual concepts. To compensate for the limited localization capa-bility of VLMs, existing approaches often incorporate large pre-trainedvisual features that inherently encode object-centric spatial cues. How-ever, these features are commonly introduced through direct fusion orshallow adaptation, which fails to explicitly align semantic representa-tions with object-level spatial information, resulting in imprecise local-ization. In this paper, we propose a novel framework for explicit seman-tic–spatial alignment in open-vocabulary object detection. Rather thandirectly merging pre-trained object-centric visual features, our methodprogressively aligns semantic representations with object-level spatialcues through a staged alignment process. Specifically, we introduce alightweight spatial adapter that spatially recalibrates auxiliary object-centric visual features to suppress background responses and emphasizesalient object regions. Building upon this, we design a frequency-awarefusion mechanism that decomposes the adapted object-centric visual fea-tures into spectral components and adaptively injects object-sensitivespatial cues into semantic features at multiple levels. This targeted fu-sion strategy enables effective semantic alignment while preserving theoriginal semantic embedding space. Extensive experiments on the OV-COCO and OV-LVIS benchmarks demonstrate the effectiveness of ourapproach against state-of-the-art methods.

As neuromorphic sensors, event cameras asynchronously recordchanges in brightness as streams of sparse events with the advantages ofhigh temporal resolution and high dynamic range. Reconstructing inten-sity images from events is a highly ill-posed task due to the inherent am-biguity of absolute brightness. Early methods generally follow an end-to-end regression paradigm, directly mapping events to intensity frames ina deterministic manner. While effective to some extent, these approachesoften yield perceptually inferior results and struggle to scale up in modelcapacity and training data. In this work, we propose EvDiff, an event-based diffusion model that follows a surrogate training framework to pro-duce high-quality videos. To reduce the high computational cost of high-frame-rate video generation, we design an event-based diffusion modelthat performs only a single forward diffusion step, equipped with a tem-porally consistent EvEncoder. Furthermore, our novel Surrogate TrainingFramework eliminates the dependence on paired event–image datasets,allowing the model to leverage large-scale image datasets for higher ca-pacity. The proposed EvDiff is capable of generating high-quality col-orful videos solely from monochromatic event streams. Experiments onreal-world datasets demonstrate that our method strikes a sweet spotbetween fidelity and realism, outperforming existing approaches on bothpixel-level and perceptual metrics. The code will be released publicly.

Rolling-shutter cameras introduce characteristic distortionswhen imaging fast moving objects, and these effects are typically treatedas artifacts to be corrected. In this work, we instead leverage rolling-shutter distortions as a valuable source of temporal information to es-timate the 3D translational and angular velocities of rapidly movingspherical objects from a single rolling-shutter frame. We design a robustand easily detectable spherical pattern and propose a correspondence-free formulation that recovers motion by enforcing geometric consis-tency in a back-projection framework. By exploiting the geometry ofthe sphere, translational and rotational motions are decoupled and esti-mated through a two-stage optimization process, enabling reliable veloc-ity recovery even for textureless objects. Extensive experiments on bothsynthetic and real datasets demonstrate accurate and robust estimationof motion parameters under challenging high-speed conditions.

Egocentric video understanding requires procedural reasoning under partial observability and continuously shifting viewpoints. Current multimodal large language models (MLLMs) struggle with this setting, often generating plausible but visually inconsistent or weakly grounded responses. We introduce EgoVITA, a framework that decomposes egocentric video reasoning into a structured plan-then-verify process. The model first generates an egocentric plan: a causal sequence of anticipated actions from a first-person perspective. This plan is then evaluated by an exocentric verification stage that uses third-person reasoning over the same video to verify its spatiotemporal and logical consistency, without exocentric video input. This decomposition enables cross-perspective feedback without requiring paired ego-exo supervision. To train this reasoning process, we adopt Group Relative Policy Optimization (GRPO) with two dense reward signals: one that grounds anticipated actions in subsequent visual observations and another that reinforces consistent third-person verification. EgoVITA achieves state-of-the-art performance on egocentric reasoning benchmarks, outperforming Qwen2.5-VL-7B by +7.7 on EgoBlind and +4.4 on EgoOrient, while maintaining strong generalization on exocentric video tasks with only 52k training samples.

Accurately forecasting human trajectories from an egocen-tric perspective plays a central role in applications such as humanoidrobotics, wearable sensing systems, and assistive navigation. However,progress in this direction remains limited due to the scarcity of egocen-tric trajectory datasets collected in real-world environments. Address-ing this need, we introduce EgoTraj, an egocentric multimodal opendataset recorded using Meta Quest Pro (MQPro). EgoTraj contains 75sequences of human navigation collected from multiple MQPro wearersin real-world urban environments. Each recording provides synchronizedRGB video along with ground-truth data, including 6-degree-of-freedomhead poses, per-frame 3D gaze vectors, and scene annotations. To thebest of our knowledge, EgoTraj differs from typical egocentric trajectorydatasets by capturing long-horizon, self-chosen pedestrian navigation onconsumer AR headset across urban routes with broad participant diver-sity. We benchmark several state-of-the-art trajectory prediction modelsand ablate the contributions of gaze, scene, and motion cues. The resultshighlight the utility of EgoTraj for AR-based perception, navigation, andassistive systems. The EgoTraj dataset, code, and EgoViz Dashboard arepublicly available at https://github.com/yehiahmad/EgoTraj.

Object detection for Unmanned Aerial Vehicles (UAVs) work-ing in open and dynamic environments is a highly challenging task.While Vision-Language Models (VLMs) have offered a powerful solutionfor universal object detection, adapting them to UAV scenarios remainsnon-trivial due to a substantial domain gap between VLM pre-trainingdata and aerial imagery. The prevailing Parameter-Efficient Fine-Tuning(PEFT) methods prove ineffective in bridging this gap, as VLMs’ “natural-scene, foreground-dominant” visual priors misalign with the “bird’s-eye-view, background-dominant, small-object” characteristics of UAV data.To address this issue, we propose DroneFINE, a novel PEFT paradigmcomprising two domain-aware complementary modules tailored for VLM-based drone image detectors. Specifically, a data-dependent, foreground-aware, and multi-path adaptation mechanism named HyperAdapter isdesigned, which overcomes the static structural constraints of PEFT. Inaddition, a background suppression algorithm named SemanticGate isdeveloped, which is a text-conditioned guidance strategy that employsbackground vocabulary to actively guide the model in suppressing re-sponses from irrelevant regions. Extensive experiments on VisDrone andUAVDT demonstrate that DroneFINE significantly outperforms existingPEFT methods and achieves performance comparable to full fine-tuningwhile substantially reducing the fine-tuned parameters.

Autoregressive (AR)-Di_x001B_usion hybrid paradigms combine AR'sstructured semantic modeling with di_x001B_usion's high-_x001C_delity synthesis, yetsu_x001B_er from a dual speed bottleneck: the sequential AR stage and theiterative multi-step denoising of the di_x001B_usion vision decode stage. Exist-ing methods address each in isolation without a uni_x001C_ed principle design.We observe that the per-position prediction entropy of continuous-spaceAR models naturally encodes spatially varying generation uncertainty,which simultaneously governing draft prediction quality in the AR stageand re_x001D_ecting the corrective e_x001B_ort required by vision decoding stage,which is not fully explored before. Since entropy is inherently tied toboth bottlenecks, it serves as a natural unifying signal for joint accel-eration. In this work, we propose Drift-AR, which leverages entropysignal to accelerate both stages: 1) for AR acceleration, we introduceEntropy-Informed Speculative Decoding that align draft_x0015_target entropydistributions via a causal-normalized entropy loss, resolving the entropymismatch that causes excessive draft rejection; 2) for visual decoder ac-celeration, we reinterpret entropy as the physical variance of the initialstate for an anti-symmetric drifting _x001C_eld_x0016_high-entropy positions acti-vate stronger drift toward the data manifold while low-entropy positionsyield vanishing drift_x0016_enabling single-step (1-NFE) decoding without it-erative denoising or distillation. Moreover, both stages share the sameentropy signal, which is computed once with no extra cost. Experimentson MAR, TransDi_x001B_, and NextStep-1 demonstrate 3.8_x0015_5.5× speedup withgenuine 1-NFE decoding, matching or surpassing original quality. Codeis available at https://github.com/aSleepyTree/Drift-AR.

Visual recognition models often fail when deployed in newenvironments. Domain Generalization (DG) addresses this by learningrepresentations that remain invariant to environment-specific variations.Recent approaches increasingly rely on large vision-language models, as-suming that preserving their expressive visual representations improvesrobustness. However, we show that such visual expressiveness can insteadpropagate spurious cues that tie representations to the training environ-ments, hindering invariant learning. We therefore discard visual guidanceand instead treat the language embedding space as the primary sourceof domain invariance, naturally acting as an information bottleneck thatpreserves core semantics while suppressing domain-specific variations.Extensive experiments across diverse backbones exhibit state-of-the-artperformance and further analyze what makes guidance effective for ro-bust generalization. These findings shift the focus of DG from improvingrepresentations to designing supervision that enforces invariance.

Text-based person anomaly search (TPAS) refers to the taskof retrieving people exhibiting normal or anomalous behaviors from nat-ural language descriptions. Existing TPAS models often learn a singlejoint embedding where appearance, action, and background informationare entangled, causing over-reliance on identity cues, poor alignment foraction-centric queries, and limited semantic connection between actionsand places where they occur. To address these issues, we propose Se-mantic Decoupled Alignment (SeDA), a disentangled vision–languageretrieval framework that explicitly factorizes both visual and textualrepresentations into appearance, action, and background components.SeDA introduces Semantic Token Projection, which derives three se-mantic queries from the global [CLS] token, softly aggregates modal-ity tokens relevant to each factor, and recomposes the resulting factortokens into a compact retrieval embedding. To enforce factor-specific se-mantics, we decompose each caption into appearance/action/backgroundsub-captions and supervise the corresponding tokens with a Feature De-coupling Loss, combined with contrastive and image–text matching ob-jectives. On the Person Anomaly Benchmark (1M pairs), SeDA achieves86.45% R@1 (+1.52 over SOTA), improves average multi-weather R@1by +2.43, and gains +3.74 R@1 under out-of-distribution evaluation.https://github.com/ErgastiAlex/SeDA

Recently, diffusion models have been widely adopted in gen-erative modeling and have served as foundational models for many imagegeneration tasks. To control the generation without costly re-training orfine-tuning, many works seek inference-time guidance methods to steerthe latent via a differentiable objective at inference time. However, thesemethods cannot effectively preserve the original Gaussian distributionbecause they introduce distributional drift, thereby degrading the sam-ple quality. To address this gap, we propose DiffRGD, a distribution-aware guidance framework that explicitly preserves the latent Gaussianstructure. DiffRGD formulates each sampling step as a constrained op-timization problem on a spherical manifold induced by the latent Gaus-sian distribution, and solves it efficiently via Riemannian Gradient De-scent (RGD). DiffRGD is a plug-and-play method that can be seamlesslyintegrated into any pre-trained diffusion model. Extensive experimentsdemonstrate that DiffRGD outperforms previous methods in most im-age restoration and conditional generation tasks. Our project page isavailable at https://diffrgd.github.io/.