ECCV 2026 论文列表

Multimodal Diffusion Large Language Models (MDLLMs)achieve high-concurrency generation through parallel masked decoding,yet the architectures remain prone to multimodal hallucinations. Thisstructural vulnerability stems from an algorithmic flaw: the decoderranks candidate tokens based on textual likelihood without verifyinglocalized visual support. We establish that this language-only rankinginduces an objective mismatch, where language probability mass acts asa misspecified proxy for the intended multimodal task. Consequently, wereinterpret hallucination as a localized optimization error, a phenomenonwhere the decoder exploits language shortcuts to maximize a proxy scoreat the expense of visual grounding. To address this objective mismatch,we introduce VISAGE, a training-free decoding framework that cali-brates the objective at inference time. VISAGE estimates the proxydiscrepancy by quantifying the spatial entropy of cross-attention distri-butions. By enforcing a localization consensus across attention heads, themethod penalizes spatially uniform distributions and re-ranks token com-mitments to favor visually grounded outcomes. We provide an analyti-cal stability guarantee establishing that VISAGE maintains a boundedobjective loss under estimation error. Evaluations across hallucination-sensitive and general-purpose benchmarks demonstrate the robustnessof the framework, yielding relative gains of 8.59% on MMMU-val and7.75% on HallusionBench.

Scene graph generation (SGG) approaches can be broadlyclassified into detector-based and query-based methods according to theirunderlying reasoning mechanisms. However, the discrepancy in their pre-dictive behaviors, induced by these distinct mechanisms, has not beensystematically analyzed. In this work, we design a controlled experimen-tal setup to examine prediction discrepancies from the perspective ofdetector-conditioned reachability. The results suggest clear complemen-tary clues. Motivated by this observation, we introduce a Dual-SGGmethod that consolidates both reasoning mechanisms via a dual-querydesign, thereby leveraging the complementary predictive behaviors ofboth detector-based and query-based methods. Extensive experimentson the Visual Genome, Open Images v6, and GQA-200 datasets demon-strate the effectiveness of the proposed method. Code is available at:Dual-SGG.

Novel view synthesis (NVS) through non-planar refractivesurfaces presents fundamental challenges due to severe, spatially varyingoptical distortions. While recent representations like NeRF and 3D Gaus-sian Splatting (3DGS) excel at NVS, their assumption of straight-lineray propagation fails under these conditions, leading to significant arti-facts. To overcome this limitation, we introduce RefracGS, a frameworkthat jointly reconstructs the refractive water surface and the scene be-neath the interface. Our key insight is to explicitly decouple the refractiveboundary from the target objects: the refractive surface is modeled via aneural height field, capturing wave geometry, while the underlying sceneis represented as a 3D Gaussian field. We formulate a refraction-awareGaussian ray tracing approach that accurately computes non-linear raytrajectories using Snell’s law and efficiently renders the underlying Gaus-sian field while backpropagating the loss gradients to the parameterizedrefractive surface. Through end-to-end joint optimization of both rep-resentations, our method ensures high-fidelity NVS and view-consistentsurface recovery. Experiments on both synthetic and real-world sceneswith complex waves demonstrate that RefracGS outperforms prior re-fractive methods in visual quality, while achieving ∼15× faster trainingand real-time rendering at 200 FPS.

Cross-view object geo-localization (CVOGL) aims to deter-mine the location of a specific object in high-resolution satellite imagerygiven a query image with a point prompt. Existing approaches treatCVOGL as a one-shot detection process, directly regressing object loca-tions from cross-view information aggregation, but they are vulnerableto feature noise and lack mechanisms for error correction. In this paper,we propose ReCOT, a Recurrent Cross-view Object geo-localizationTransformer, which models CVOGL as a recurrent localization process.ReCOT introduces a set of learnable tokens that encode task-specificintent from the query image and prompt embeddings, and iteratively at-tend to the reference features to refine the predicted location. To enhancethis recurrent process, we incorporate two complementary modules: (1) aSAM-based knowledge distillation strategy that transfers segmentationpriors from the Segment Anything Model (SAM) to provide clearer se-mantic guidance without additional inference cost, and (2) a ReferenceFeature Enhancement Module (RFEM) that introduces hierarchical at-tention to emphasize object-relevant regions in the reference features.Extensive experiments on CVOGL benchmarks demonstrate that Re-COT achieves state-of-the-art (SOTA) performance while significantlyreducing parameters compared to previous SOTA approaches. Our codeis available at https://github.com/Temperature-ai/ReCOT.git.

Aerial geo-localization is increasingly essential for large-scale spatial retrieval. To overcome the limitations of single-retrieval systems caused by modality-specific uncertainty in open environments, this paper introduces the unified Aerial Multi-view Geo-localization (AMGL) task. However, when applying universal multimodal retrieval paradigms to this task, the semantic bias introduced solely by textual instructions fails to resolve and reconstruct the differentiated distribution structures necessary for cross-view alignment. Consequently, this paper proposes the Rectified Embedding Flow Learning (REFL) framework, formulating crossdomain alignment as a directed conditional distribution transport process. Initially, REFL employs velocity-prior flow learning to fit continuous ordinary differential equation trajectories, deriving transformation priors that map single-view distributions to a latent shared manifold. Subsequently, a trajectory-guided embedding rectification mechanism continuously transports query features to the target view distribution, explicitly compensating for distribution shifts. Extensive evaluations on the Aerial MVGL benchmark demonstrate that REFL achieves state-of-theart performance, yielding an average R@1 of 44.83% and R@10 of 65.85%. The code and benchmark are available at https://github.com/rhaohur/REFL.

Video diffusion transformers (DiTs) generate high-fidelityand temporally coherent videos, yet motion control remains implicit,primarily relying on text prompts. As a result, achieving desired motionoften requires extensive prompt engineering and repeated resampling.While fine-tuning models with additional spatial prompts (e.g., bound-ing boxes or point trajectories) enables explicit control, it demands sub-stantial data curation and computation, and may compromise the gener-ative capabilities of pretrained models. Consequently, training-free mo-tion control using such spatial prompts has been explored in U-Net–basedvideo diffusion models, but remains largely unexplored for DiTs. We in-troduce Qwerty, a training-free framework that enables flexible motioncontrol in pretrained image-to-video DiTs via user-defined object warpingand optical flow. We carefully manipulate the 3D full attention of DiTsby warping the frame-invariant semantic subspace of queries. We findthat the noise predicted by the query-warped DiT naturally guides thediffusion trajectory toward the desired motion, and further show thatleveraging this noise as self-guidance for latent optimization improvescontrol stability and visual quality. Experiments show that Qwertyachieves the most effective motion control among existing training-freeapproaches on a recent image-to-video DiT, with performance compara-ble to fine-tuning–based methods.

Hallucinations in medical vision language models arise whenlanguage priors override subtle visual evidence during decoding. Existingtraining-free contrastive decoding methods suppress these priors by con-trasting logits against a globally perturbed image, but global perturba-tions destroy the very diagnostic cues they are meant to protect —stan-dard VCD blur flips nearly one in five CheXpert labels on held-out chestradiographs. We propose Proximity-Constrained CounterfactualDecoding (PCCD), which admits a counterfactual view only whenit satisfies both a global and an object-masked feature-space similaritybound, constraining diagnostic drift while increasing the contrastive mar-gin for visually grounded tokens to first order under explicitly stated as-sumptions. Two complementary branches address heterogeneous halluci-nation drivers: Object-Aware VCD preserves attribution-indicated lesionregions while attenuating contextual co-occurrence priors, and Latent-Disentangled VCD applies structure-preserving style edits to suppressattribute biases. When neither branch passes the proximity gate, PCCDfalls back to greedy decoding, guaranteeing it never degrades below theunmitigated baseline. Across radiology VQA, chest X-ray report gener-ation, and a new CXR existence probe (Med-POPE), PCCD improvesGREEN AUC, QAAS, and RadGraph F1 by up to →+2.5 pp over SOTAwith one auxiliary forward pass per token and no weight updates.

We introduce PRISM-VO, a novel pure optimization-basedsparse photometric visual odometry framework for focused plenopticcameras. The core of PRISM-VO is a novel photometric plenoptic bun-dle adjustment which jointly optimizes camera poses and inverse depthvalues of points in a sliding window. By combining geometric depth froma single plenoptic image with temporal multi-view constraints, PRISM-VO achieves accurate and drift-resilient motion estimation. Through ex-plicit modeling of the plenoptic projection, PRISM-VO provides reliablemetric-scale reconstructions, overcoming the scale ambiguity of monocu-lar SLAM algorithms. Importantly, our approach relies solely on a singleplenoptic sensor and avoids complex initialization, as depth priors arecomputed directly from plenoptic imaging.Experiments show that PRISM-VO outperforms the current state-of-the-art plenoptic visual odometry method on indoor and outdoor scenes.The proposed approach rivals other optimization- and learning-basedmethods while accurately and reliably recovering a metric scale of thescene.Project page: https://prism-vo.github.io/.

Multi-View Pedestrian Detection (MVPD) aims to detectpedestrians in the form of a bird’s eye view map from multi-view images.Recent MVPD methods adopt a unified framework that projects 2D im-age features into a 3D world space and aggregates them into a singlefeature. Although they are effective, they struggle to generalize to un-seen camera configurations during training due to two main issues. First,they are difficult to capture accurate visual geometry across views in un-seen camera configurations. Second, they make detection models highlydependent on distortion patterns during training arising from their im-age feature projection. To address these, we leverage a visual geometricfoundation model and propose MV2GF. This foundation model has ex-hibited strong generalization in capturing visual geometry across viewsand predicting accurate 3D attributes in diverse camera configurations.MV2GF fuses task-specific features with general-purpose geometric fea-tures extracted by the foundation model to effectively capture the visualgeometry even in unseen camera configurations. Furthermore, MV2GFprojects each pixel in the image features to an appropriate 3D locationusing 3D pointmaps predicted by the foundation model, preventing thedetection model from depending on distortion patterns during training.Our experiments demonstrate the effectiveness of leveraging a visual geo-metric foundation model for MVPD and that MV2GF generalizes betterthan existing methods.

Test-time adaptation (TTA) aims to mitigate distributionshifts by adapting models with unlabeled target data at inference time.While TTA with vision-language models (VLMs) has shown promisingresults in classification, extending it to medical image segmentation re-mains challenging. In this setting, the adaptation gains from optimizingon VLM-generated predictions are often outweighed by the degradationto the VLM’s strong pretrained features caused by noisy, update-drivenlearning, resulting in limited and unstable improvements. We thereforepropose Memory-Supported Synergistic Adaptation (MSSA), anovel training-free TTA framework for medical image segmentation.Without updating model parameters, MSSA dynamically selects reli-able image–text predictions to construct an online memory, uses themas text-guided semantic priors, and couples them with cross-image struc-tural alignment for robust adaptation. Specifically, MSSA consists of (i) anoise-aware memory construction module that filters and stabilizes cross-modal predictions, and (ii) a relevance-driven prototype alignment mod-ule that aligns the target sample with structurally consistent memorysamples and their reliable predictions to improve adaptation. Extensiveexperiments on multiple medical segmentation benchmarks demonstratethat MSSA consistently improves VLM-based segmentation models andoutperforms existing fine-tuning-based TTA methods by a clear mar-gin, with gains of up to 12.2% DSC and 11.7% mIoU. Project Page:https://lingrayy.github.io/MSSA/.

Image restoration aims to reconstruct high-quality images from degraded low-quality inputs. As the computational demands of image restoration models continue to rise, there is growing interest in lightweight architectures optimized for fast and efficient inference. Logic gate networks (LGNs), which operate using fundamental logic operations such as NAND and XOR, have recently emerged as a promising direction for achieving highly efficient computation. However, their potential remains largely untapped in the domain of image restoration. In this work, we introduce LogicIR, the first LGN specifically designed for image restoration tasks. LogicIR incorporates a UNet-inspired architecture composed entirely of logic gates. In addition, we propose a differentiable bit decoding layer and an index shuffling mechanism that improves information propagation across logic gates. Experimental results across multiple image restoration benchmarks demonstrate that LogicIR achieves strong performance with significantly reduced computational cost, establishing LogicIR as a viable and efficient alternative for image restoration. The source code is available at our project page https://github.com/jimmy9704/LogicIR.

Unified multimodal models have recently shown remarkablegains in both capability and versatility, yet most leading systems are stilltrained from scratch and require substantial computational resources. Inthis paper, we show that competitive performance can be obtained farmore efficiently by strategically fusing publicly available models special-ized for either generation or understanding. Our key design is to re-tain the original blocks while additionally interleaving multimodal self-attention blocks throughout the networks. This double fusion mechanism(1) effectively enables rich multi-modal fusion while largely preservingthe original strengths of the base models, and (2) catalyzes synergisticfusion of high-level semantic representations from the understanding en-coder with low-level spatial signals from the generation encoder. By train-ing with only ∼35B tokens, this approach achieves strong results acrossmultiple benchmarks: 0.91 on GenEval for compositional text-to-imagegeneration, 82.16 on DPG-Bench for complex text-to-image generation,6.06 on GEditBench, and 3.77 on ImgEdit-Bench for image editing. Wewill fully release the entire suite of code, model weights, and datasets tosupport future research on unified multimodal modeling.

Neural rendering techniques allow for accurate reconstruc-tion of the geometry and color appearance of 3D scenes. Some methodshave extended their use to additional imaging modalities, such as multi-spectral, infrared, or polarimetric data. However, all of these approachesrequire expensive sensors and calibrated setups to capture new multi-modal frames for each new scene. We propose Spectral and PolarimetricImplicit Learned Representation (SPoILeR), a novel method to obtainmulti-view consistent renderings of unconventional modalities for sceneswhere either only RGB frames or very few of the additional modali-ties are available. Thanks to a multimodal pre-training phase, the modellearns the mutual correlation between different modalities. This step al-lows predicting accurate renderings of unconventional modalities duringa fine-tuning phase supervised only by RGB images. Experimental resultsshow that the approach can accurately render infrared, polarimetric, andmultispectral frames for scenes where no input sample captured by thesetypes of sensors is provided.

Generative vision-language models (VLMs) such as Qwen-VLand LLaVA achieve strong zero-shot performance on tasks overlappingwith their pretraining distribution, yet fail on specialized domains wherethe required discriminative features were never learned, a regime weterm distant out-of-distribution (OOD). Standard adaptation methodscannot overcome this representational absence because they operatewithin the encoder’s existing feature space. However, VLMs retain arobust descriptive capacity even when discrimination collapses: a modelthat cannot classify a medical scan can still articulate its visual patterns.Exploiting this asymmetry, we introduce Inductive Visual Logic (IVL), atraining-free framework that constructs classification knowledge from themodel’s surviving descriptive ability. IVL extracts visual traits from few-shot support images through dual-mode prompting, combining semanticdescriptions with primitive visual observations, and organizes them intoper-class trait dictionaries. At inference, hierarchical filtering identifiesspatially grounded trait evidence for classification. Across multiple distant-OOD benchmarks, IVL achieves the highest aggregate accuracy under twoVLM backbones while producing interpretable, trait-traceable predictions.

Self-collision remains a persistent challenge in SMPL-basedhuman pose estimation and motion generation. Under extreme articula-tions or stochastic motion synthesis, generated meshes frequently exhibitself-penetrations, leading to physically implausible results. We proposePoseShield, a neural collision constraint defined directly in SMPL posespace. We formulate collision correction as a constrained optimizationproblem and connect the learned constraint with the Eikonal equation.Enforcing Eikonal regularization ensures non-vanishing gradients nearthe collision boundary, improving numerical stability and robustness ofthe optimization process. Unlike prior methods that operate in the meshspace or rely on heuristic penalties, our approach operates directly in thelow-dimensional space of human poses and is theoretically grounded. Thesame learned constraint extends to human motion sequences, providinga generator-agnostic post-hoc collision corrector without retraining theunderlying motion model. Experiments on a newly constructed SMPLpose benchmark show that our method achieves a 95.8% success rateand outperforms state-of-the-art baselines.

Generating clean line art from conti, rough storyboard sketchesused in webtoon production, remains a labor-intensive process requiringsubstantial artistic expertise. Conti sketches provide only coarse geom-etry and sparse semantic cues, making it difficult to recover detailedline art while preserving consistent character identity across sequentialpanels. Existing sketch-to-image or diffusion-based translation modelsprocess images independently and therefore struggle to maintain cross-panel identity coherence. We present PanelDiff, a diffusion transformerframework for conti-to-line-art generation that explicitly models cross-panel character consistency. PanelDiff integrates three components. Amulti-reference character conditioning module encodes character labelsand multiple line-art exemplars into compact identity tokens for robustidentity guidance under ambiguous conti inputs. A memory-augmentedrepresentation bank accumulates character-aware features from previ-ously generated panels, while a similarity interpreter dynamically re-trieves relevant entries for the current panel. A panel-aware diffusiontransformer then jointly attends to conti structure, textual descriptions,identity tokens, and retrieved memory features to produce coherent line-art sequences. To support this task, we construct a conti–line-art datasetwith character identity annotations and sequential panel structures. Ex-periments show that PanelDiff improves line-art fidelity, identity preser-vation, and cross-panel consistency over strong baselines, and ablationstudies verify the contribution of each component. The code and datasetare available online (huggingface) ⋆ .

How can we teach large multimodal models (LMMs) newskills without erasing prior abilities? We study sequential fine-tuningon five target skills while monitoring general ability on eight held-outbenchmarks across three model families. Surprisingly, we find that per-formance lost on held-out tasks after fine-tuning on one skill can partlyrecover when the model is subsequently tuned on a different skill. Wetrace this behavior to a measurable shift in the output token distribu-tion, manifested through a simple counting-bias probe that shows theshift co-varies with forgetting. Guided by this insight, we identify twosimple, robust tuning recipes that learn strongly while limiting drift: (i)updating only the self-attention projection layers (SA Proj., ∆ learning+24.9 / ∆ held-out forgetting -0.6), and (ii) updating only the MLPGate&Up while freezing the Down projection (+30.5 / -2.1). Both sub-stantially outperform full-LLM tuning (+31.8 / -23.3) in the learning–forgetting trade-off. We also compare against common forgetting mitiga-tion methods—Learning without Forgetting (LwF), LoRA, Mixture-of-Experts, and weight-space interpolation (WiSE-FT)—and find that ourselective tuning recipes match or exceed their learning–stability balancewhile remaining simpler, requiring no replay, auxiliary parameters, orper-stage tuning. These results hold across LLaVA-OneVision, LLaVA-NeXT, and Qwen2.5-VL, confirming that the key to teaching LMMs newskills without forgetting lies in controlling output distribution shift bychoosing which components to tune. Code will be made available.

The advancement of Video Object Segmentation (VOS) isimpeded by its heavy reliance on expensive annotations. This has led toa landscape of compromised alternatives: unsupervised methods provideweak information, few-shot learners are plagued by referential ambiguityfrom distractors, and referring techniques are limited by a lack of visionpriors. We revisit Few-Shot Referring Video Object Segmentation (FSR-VOS) under a streamlined and practical formulation that segments tar-get objects using only query-level textual descriptions and a small set ofannotated support images. This dual conditioning enables coarse-to-finedisambiguation and precise boundary refinement with minimal annota-tion cost. To address the challenge of complex multi-source informationfusion, we propose the QST-SAM framework. Specifically, a SupportInstruction Compressor is introduced to refine and condense the sup-port information into compact representations. The Q-S-T Transformerfurther facilitates comprehensive integration of multi-modal supervisorycues from query, support, and text branches. The Multi-Priors Prompt-ing module provides the SAM2 prompt encoder with enriched guidancefor precise, context-aware segmentation. We further establish a new FS-RVOS benchmark derived from Ref-Youtube-VOS and Ref-DAVIS17, onwhich QST-SAM achieves state-of-the-art performance, validating its ro-bustness and effectiveness in segmentation.

We study positional encodings for multi-view transformersthat process tokens from a set of posed input images, and seek a mech-anism that encodes patches uniquely, allows SE(3)-invariant attentionwith multi-frequency similarity, and can adapt to the geometry of theunderlying 3D scene. We find that prior (absolute or relative) encod-ing schemes for multi-view attention do not meet these desiderata, andpresent RayRoPE to address this gap. RayRoPE represents patch po-sitions based on associated rays and computes query-frame projectivecoordinates to ensure SE(3) invariance. To adapt to scene geometry,RayRoPE predicts (without direct supervision) a per-token depth to ob-tain its position along the corresponding ray, while also modeling uncer-tainty and analytically computing the expected positional encoding. Wevalidate our method on the tasks of novel-view synthesis, stereo depthestimation, and feed-forward 3DGS reconstruction. While remaining ef-ficient, RayRoPE consistently improves over alternate position encodingschemes (e.g. 24% relative improvement on LPIPS in RE10K).