CVPR 2026 Papers — Page 20
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Learning What Matters: Prioritized Concept Learning via Relative Error-driven Sample Selection
Qian Yang (Mila Quebec Ai Institute), Leonid Sigal (University Of British Columbia)
Computational EfficiencyRepresentation LearningSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Propose a data and computation-efficient VLM instruction-tuning framework called PROGRESS, which adaptively selects the most valuable samples by tracking the model's learning progress.
Learning What to Trust: Bayesian Prior-Guided Optimization for Visual Generation
Ruiying Liu (Chinese University of Hong Kong Shenzhen), Chi Zhang (China Telecom)
GenerationOptimizationReinforcement LearningVision Language ModelContrastive LearningImageVideoMultimodality
🎯 What it does: Propose Bayesian Prior-Guided Optimization (BPGO) under the GRPO framework, modeling uncertainty in text-visual correspondence through prior rewards, and adaptively weighting and renormalizing rewards at both group-level and sample-level to enhance semantic alignment and visual quality in visual generation.
Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency
Hakan Emre Gedik (University of Texas at Austin), Alan Bovik (University of Colorado Boulder)
RestorationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Propose a multi-scale patch encoding framework named ReLIQS, which leverages the CLIP visual encoder, saliency estimator, and latent quality axis module to achieve no-reference image quality assessment for images of arbitrary resolution, achieving breakthroughs in preserving original resolution details, maintaining a controllable computational budget, and enabling cross-dataset learning.
LEMON: A Large Endoscopic MONocular Dataset and Foundation Model for Perception in Surgical Settings
Chengan Che (King's College London), Luis C. Garcia-Peraza-Herrera (King's College London)
ClassificationRecognitionObject DetectionSegmentationKnowledge DistillationRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningVideoBiomedical DataBenchmark
🎯 What it does: Proposed a large-scale publicly available surgical video dataset named LEMON (over 4K segments, 938 hours, 85M frames) and trained a new surgical foundation model, LemonFM, based on this dataset.
Lens Component Deletion based on Differentiable Ray Tracing
Wenguan Zhang (Zhejiang University), Qi Li (Zhejiang University)
OptimizationComputational EfficiencyConvolutional Neural NetworkImagePhysics Related
🎯 What it does: Proposed a lens component removal process based on differentiable ray tracing, which automatically identifies and removes lens components with the least impact on imaging, while achieving distortion correction through joint optimization of the optical system and post-processing network.
Lenses: Toward Polysemous Vision-Language Understanding
Hani Alomari, Chris Thomas
RetrievalRepresentation LearningLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose a multi-perspective image-text retrieval framework named Lenses, construct the Lenses dataset, and train a multi-prompt embedding model to capture the multiple meanings of images and their corresponding relationships with text.
LensWalk: Agentic Video Understanding by Planning How You See in Videos
Keliang Li (Institute of Computing Technology, Chinese Academy of Sciences), Shiguang Shan (Institute of Computing Technology, Chinese Academy of Sciences)
TransformerLarge Language ModelAgentic AIVision Language ModelVideoMultimodality
🎯 What it does: Propose LensWalk, an active observation framework based on large language models (LLM) and vision-language models (VLM), which allows the LLM to dynamically plan the temporal range, sampling density, and tool calls for video observations during reasoning, thereby achieving a 'reasoning-planning-observation' loop.
LESA: Learnable Stage-Aware Predictors for Diffusion Model Acceleration
Peiliang Cai (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerMixture of ExpertsDiffusion modelImageVideo
🎯 What it does: Designed a learnable stage-aware prediction framework (LESA) based on Kolmogorov-Arnold Networks (KAN) to accelerate feature cache inference in diffusion transformers (DiT), significantly reducing the number of inference steps and computational load.
Less is More: Data-Efficient Adaptation for Controllable Text-to-Video Generation
Shihan Cheng (Vanderbilt University), Dmitriy Smirnov (Netflix)
GenerationData SynthesisTransformerDiffusion modelVideoText
🎯 What it does: Proposed a data-efficient fine-tuning framework that utilizes sparse low-quality synthetic data to add controllable camera parameters (shutter speed, aperture, color temperature) to a pre-trained text-to-video diffusion model.
Let it Snow! Animating 3D Gaussian Scenes with Dynamic Weather Effects via Physics-Guided Score Distillation
Gal Fiebelman (Hebrew University of Jerusalem), Sagie Benaim (Hebrew University of Jerusalem)
GenerationOptimizationRecurrent Neural NetworkLarge Language ModelDiffusion modelScore-based ModelGaussian SplattingVideoPoint CloudPhysics Related
🎯 What it does: Propose a Physics-Guided Score Distillation framework that combines motion priors from physical simulations with video diffusion models to achieve dynamic weather editing (snow, rain, fog, sandstorms) on static 3D Gaussian fields.
Let VLMs Grade Their Own Thoughts: A Self-Quantification Approach to Reasoning-Aware Reward Modeling
Xing Xi (South China University Of Technology), peilin tong (Ant Group)
Reinforcement Learning from Human FeedbackReinforcement LearningVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: Enable video language models (VLM) to self-assess their reasoning paths and use their own answer confidence as a reward signal, optimizing reasoning ability through reinforcement learning (GRPO).
Let Your Image Move with Your Motion! -- Implicit Multi-Object Multi-Motion Transfer
Yuze Li (Tianjin University), Xinyu Zhang (University of Auckland)
GenerationTransformerDiffusion modelOptical FlowImageVideo
🎯 What it does: Propose FlexiMMT, achieving for the first time implicit image-to-video (I2V) multi-object, multi-motion transfer. Given a multi-object static image and several reference videos, FlexiMMT can independently extract each motion and precisely map it to the corresponding object in the target image, enabling arbitrary motion-object combinations and rearrangements.
Leveraging Class Distributions in CLIP for Weakly Supervised Semantic Segmentation
Ziqian Yang (Xi'an Jiaotong-Liverpool University), Jimin Xiao (Xi'an Jiaotong-Liverpool University)
SegmentationTransformerContrastive LearningImage
🎯 What it does: Proposed the CD-CLIP framework, leveraging CLIP's full category distribution and DINO's superclass boundary information to refine and constrain the Class Activation Map (CAM) for weakly supervised semantic segmentation.
Leveraging Multispectral Sensors for Color Correction in Mobile Cameras
Luca Cogo (University of Milano Bicocca), Raimondo Schettini (University of Milano Bicocca)
RestorationConvolutional Neural NetworkImageMultimodality
🎯 What it does: This work proposes an end-to-end color correction framework that integrates high-resolution RGB and low-resolution multispectral (MS) sensor images, achieving unified processing of illumination estimation, illumination suppression, and color space conversion;
Leveraging Verifier-Based Reinforcement Learning in Image Editing
Hanzhong Guo (University of Hong Kong), Weilin Huang (ByteDance Seed)
Image TranslationSupervised Fine-TuningReinforcement LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes a Reasoning Reward Model (RRM) based on a Verifier for image editing using RLHF, constructs a two-stage training pipeline (SFT+GCPO), and applies RRM to GRPO training for editing models.
LF-BVN: Blind-View Network for Self-Supervised Light Field Denoising
Longzhao Guo (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
RestorationDepth EstimationImage
🎯 What it does: This paper proposes a self-supervised light field image denoising framework named LF-BVN (Blind-View Network), which achieves training without clean images by leveraging multi-view consistency in light fields.
LIBERO-Plus: A Progressive Robustness Benchmark for Visual-Language-Action Models
Senyu Fei (Tongji University), Xipeng Qiu (Fudan University)
Supervised Fine-TuningVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Proposes LIBERO-Plus, an automated, scalable, and fine-grained vision-language-action (VLA) robustness benchmark based on LIBERO, covering seven disturbance dimensions (object layout, camera perspective, robot initial state, language instruction, illumination, background texture, sensor noise), with five levels of difficulty set for each dimension.
LiDAR Prompted Spatio-Temporal Multi-View Stereo for Autonomous Driving
Qihao Sun (Alibaba Group), Ruifeng Li (Harbin Institute of Technology)
Depth EstimationAutonomous DrivingConvolutional Neural NetworkTransformerPrompt EngineeringImagePoint Cloud
🎯 What it does: Integrate sparse LiDAR hints with multi-view stereo reconstruction to propose DriveMVS, achieving multi-view, temporally consistent, and cross-domain metric depth estimation.
LiDAR-to-4DRadar Diffusion Bridge via Cross-Modal Alignment and Translation in Latent Space
Dazhong Shen (Nanjing University of Aeronautics and Astronautics), Ying Sun (National University of Defense Technology)
Image TranslationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelAuto EncoderContrastive LearningMultimodalityPoint Cloud
🎯 What it does: Propose a 4D millimeter-wave radar tensor generation method based on LiDAR input, achieving LiDAR-to-4D-Radar translation;
LiDAS: Lighting-driven Dynamic Active Sensing for Nighttime Perception
Simon de Moreau (Valeo), Fabien Moutarde (Mines Paris PSL)
Object DetectionSegmentationAutonomous DrivingOptimizationConvolutional Neural NetworkImage
🎯 What it does: Designed and implemented a closed-loop active lighting system called LiDAS based on high-resolution headlights, capable of real-time adjustment of illumination distribution to enhance the detection and segmentation performance of night-time cameras.
LiDeRe: A Lightweight Readout for Fast and Data-Efficient Dense Prediction
Timo Lüddecke (University of Göttingen), Alexander Ecker (MPI for Dynamics and Self-Organization)
Object DetectionSegmentationPose EstimationComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Design a lightweight dense prediction reader, LiDeRe, on large frozen visual backbones (e.g., DINOv3), leveraging learnable interpolation and content-guided attention to achieve high-resolution, fine-grained pixel-level predictions.
Life-IQA: Boosting Blind Image Quality Assessment through GCN-enhanced Layer Interaction and MoE-based Feature Decoupling
Long Tang (Shanghai Jiao Tong University), Liang Yuan (Beijing University of Chemical Technology)
Graph Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: This paper proposes a BIQA framework called Life-IQA, which utilizes a Transformer decoder to achieve multi-layer feature interaction and decoding, thereby improving the accuracy of blind image quality assessment.
LifeEval: A Multimodal Benchmark for Assistive AI in Egocentric Daily Life Tasks
Hengjian Gao (Shanghai Artificial Intelligence Laboratory), Guangtao Zhai (Shanghai Jiao Tong University)
TransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Propose the LifeEval benchmark to evaluate the performance of multimodal large language models in first-person real-time task assistance scenarios;
Lifelong Imitation Learning with Multimodal Latent Replay and Incremental Adjustment
Fanqi Yu (Istituto Italiano di Tecnologia), Vittorio Murino (Istituto Italiano di Tecnologia)
Representation LearningRobotic IntelligenceTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose a lifelong imitation learning framework that achieves continuous task learning and prevents catastrophic forgetting through multimodal latent replay and incremental feature adaptation.
LIFT and PLACE: A Simple, Stable, and Effective Knowledge Distillation Framework for Lightweight Diffusion Models
Hyunsoo Han (Ulsan National Institute of Science and Technology), Jaejun Yoo (Ulsan National Institute of Science and Technology)
CompressionKnowledge DistillationDiffusion modelImage
🎯 What it does: Proposed LIFT (Linear Fit Distillation) and PLACE (Partitioned Local Adaptive Coefficient Estimation) methods to address the instability in knowledge distillation caused by capacity gaps during diffusion model compression.
Lifting Unlabeled Internet-level Data for 3D Scene Understanding
Yixin Chen (State Key Laboratory of General Artificial Intelligence, BIGAI), Siyuan Huang (State Key Laboratory of General Artificial Intelligence, BIGAI)
Object DetectionSegmentationData SynthesisData-Centric LearningLarge Language ModelVision Language ModelVideoTextPoint Cloud
🎯 What it does: Built an automated data engine for 3D scene understanding from unlabeled internet videos, generating the SceneVerse++ dataset and evaluating it on tasks such as 3D detection, VQA, and VLN.
Lighting in Motion: Spatiotemporal HDR Lighting Estimation
Christophe Bolduc (Universit' e Laval), Jean-François Lalonde
GenerationDiffusion modelOptical FlowImageVideo
🎯 What it does: This work proposes LIMO, which utilizes diffusion models to estimate spatiotemporal high dynamic range (HDR) illumination from single-frame images or videos. By 'filling' specular and diffuse spheres at specified 3D positions, it ultimately generates HDRI maps applicable for virtual object insertion.
Lighting-grounded Video Generation with Renderer-based Agent Reasoning
Ziqi Cai (Peking University), Boxin Shi (Peking University)
GenerationData SynthesisDiffusion modelAuto EncoderVideoTextMultimodality
🎯 What it does: This paper proposes LiVER, a lighting-driven video generation framework based on diffusion models, which utilizes a renderer proxy to convert text instructions into a controllable 3D scene proxy (including geometric layout, camera trajectory, and HDR environment light), and injects these physical information into the video diffusion model, achieving precise control over lighting, layout, and camera.
LightMover: Generative Light Movement with Color and Intensity Controls
Gengze Zhou (Adelaide University), Scott Cohen (Adobe Research)
GenerationTransformerDiffusion modelAuto EncoderImageVideo
🎯 What it does: Design the LightMover framework, leveraging video diffusion models to achieve controllable movement of light position, color, and intensity on a single image, generating physically consistent shadows, reflections, and lighting attenuation changes while avoiding re-rendering;
LightRR: A Lightweight Network for Single Image Reflection Removal
Wenbin Yin (East China Normal University), Guixu Zhang (East China Normal University)
RestorationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: Propose LightRR, a lightweight single-image reflection removal network that achieves efficient reflection removal using waveform frequency splitting and Mamba state space modeling.
LightSplat: Fast and Memory-Efficient Open-Vocabulary 3D Scene Understanding in Five Seconds
Jaehun Bang (Ulsan National Institute of Science and Technology), Kyungdon Joo (Ulsan National Institute of Science and Technology)
SegmentationRetrievalComputational EfficiencyVision Language ModelGaussian SplattingPoint Cloud
🎯 What it does: Propose LightSplat, an open-source training-free, fast, and low-memory vocabulary-based 3D scene understanding framework, achieving one-time semantic injection and single-step clustering by injecting 2D semantics with 2-byte indices into 3D Gaussian Splatting, supporting fast retrieval.
Linear Fundamental Matrix Estimation from 7 or 5 Points
Taci Ata Kucukpinar (University of Missouri), Kannappan Palaniappan (University of Missouri)
Pose EstimationImage
🎯 What it does: This paper proposes a linear minimal solver for the V-Umlaut configuration, which estimates the fundamental matrix in two-view cameras using five real corresponding points and two virtual midpoints.
Linear Image Generation by Synthesizing Exposure Brackets
Yuekun Dai (Nanyang Technological University), Nanxuan Zhao (Adobe Research)
GenerationTransformerSupervised Fine-TuningDiffusion modelFlow-based ModelImageText
🎯 What it does: Propose a text-to-linear image generation framework that can directly synthesize linear (scene-referenced) images with full dynamic range and supports post-exposure editing, white balance adjustment, and other edits.
Linguistic Priors for Visual Decoupling: Towards Symmetric Vision-Brain Alignment
Dongjun Liu (Hangzhou Dianzi University), Wanzeng Kong (Hangzhou Dianzi University)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: Proposes a vision disentanglement method guided by language priors, using text priors to guide the separation of foreground objects and background in visual images, thereby achieving better alignment between visual and brain signals.
Linking Modality Isolation in Heterogeneous Collaborative Perception
Changxing Liu (Nanjing University of Science and Technology), Siheng Chen (Nanjing University of Science and Technology)
Autonomous DrivingMultimodality
🎯 What it does: Propose the CodeAlign framework to address the modality isolation problem in heterogeneous collaborative perception caused by the lack of co-occurrence data across different modalities, by constructing modality-specific discrete feature spaces through codebooks and achieving cross-modal alignment via feature-code-feature (FCF) translation.
Linking Perception, Confidence and Accuracy in MLLMs
Yuetian Du (Zhejiang University), Qiang Zhu (Zhejiang University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMixture of ExpertsVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose Confidence-Driven Reinforcement Learning (CDRL) to calibrate model confidence by incorporating original-noise image pairs and confidence rewards into multi-modal large language models. Based on this confidence, develop the Confidence-Aware Test-Time Scaling (CA-TTS) framework, integrating three modules: Self-Consistency, Self-Reflection, and Self-Check, ultimately significantly enhancing visual reasoning performance.
LinVideo: A Post-Training Framework towards O(n) Attention in Efficient Video Generation
Yushi Huang (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelVideoBenchmark
🎯 What it does: A data-agnostic post-training framework named LINVIDEO is proposed for trained video diffusion models, achieving efficient video generation by layer-by-layer automatic selective replacement of high-complexity softmax attention with low-complexity linear attention.
Lipschitz Optimization for Formal Verification of Homographies
Jean-Guillaume Durand (Joby Aviation), Alessio Lomuscio (Safe Intelligence)
OptimizationImageBenchmark
🎯 What it does: Proposed a formal robustness verification framework for homography induced by camera motion, closed-form derivation of homography caused by 6-DOF pose perturbations, and achieved upper and lower bounds estimation of pixel values through piecewise linear approximation and Lipschitz optimization;
LiREC-Net: A Target-Free and Learning-Based Network for LiDAR, RGB, and Event Calibration
Aditya Ranjan Dash (RPTU University Kaiserslautern Landau), Didier Stricker (RPTU University Kaiserslautern Landau)
Pose EstimationAutonomous DrivingTransformerOptical FlowMultimodalityPoint Cloud
🎯 What it does: Proposes a modality-agnostic learning-based tri-modal (LiDAR-RGB-event) extrinsic calibration network, LiREC-Net, which can directly learn calibration from three sensor data in natural scenes.
Lite Any Stereo: Efficient Zero-Shot Stereo Matching
Junpeng Jing (Imperial College London), Krystian Mikolajczyk (Imperial College London)
Depth EstimationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Designed a lightweight stereo matching network called Lite Any Stereo, which can achieve high-precision disparity estimation on real-world images without requiring domain fine-tuning.
LitePT: Lighter Yet Stronger Point Transformer
Yuanwen Yue (ETH Zurich), Konrad Schindler (ETH Zurich)
Object DetectionSegmentationConvolutional Neural NetworkTransformerPoint Cloud
🎯 What it does: Designed and implemented a lightweight point cloud Transformer called LitePT, integrating convolution and attention modules to provide an efficient and scalable backbone network for point cloud analysis.
LiteSense: Lifting Lightweight ToF with RGB for High-Resolution Metric Depth Estimation
Yusheng Li (Tongji University), Jianmei Wang
Depth EstimationConvolutional Neural NetworkImageMultimodality
🎯 What it does: Propose a lightweight RGB-ToF fusion framework named LiteSense, which integrates low-resolution ToF depth and Compact Normalized Histogram (CNH) with RGB images to generate high-resolution depth maps with absolute scale.
LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging
Zhijian Shu (Nanjing University of Posts and Telecommunications), Xiao-Xiao Long (Nanjing University)
GenerationPose EstimationComputational EfficiencyTransformerSupervised Fine-TuningImagePoint Cloud
🎯 What it does: Propose LiteVGGT, an improvement over the Visual Geometry Grounded Transformer (VGGT), significantly enhancing speed and memory usage in long-sequence 3D reconstruction tasks through a geometry-aware cached token merging technique, while maintaining or even improving reconstruction and camera pose estimation accuracy.
Live Interactive Training for Video Segmentation
Xinyu Yang (Cornell University), Jennifer J. Sun (Cornell University)
ClassificationSegmentationTransformerSupervised Fine-TuningPrompt EngineeringImageVideo
🎯 What it does: Propose a Live Interactive Training (LIT) framework that enables video segmentation models to learn user corrections in real-time during inference and adaptively improve.
LiveGesture: Streamable Co-Speech Gesture Generation Model
Muhammad Usama Saleem (University of North Carolina), Pu Wang (University of Central Florida)
GenerationPose EstimationTransformerMeshAudio
🎯 What it does: Propose LiveGesture, a zero-latency, streaming, real-time full-body speech-driven gesture generation framework capable of synchronously generating diverse, beat-aligned SMPL-X motions under arbitrary-length speech inputs.
LLaDA-MedV: Exploring Large Language Diffusion Models for Biomedical Image Understanding
Xuanzhao Dong (Arizona State University), Yalin Wang (Arizona State University)
RecognitionTransformerSupervised Fine-TuningVision Language ModelDiffusion modelBiomedical Data
🎯 What it does: By migrating the large-scale language diffusion model LLaDA to the medical imaging field and combining it with visual instruction tuning, the LLaDA-MedV model was constructed to achieve visual-language understanding in medical imaging.
LLaDA-V: Large Language Diffusion Models with Visual Instruction Tuning
Zebin You (Renmin University of China), Chongxuan Li (Renmin University of China)
GenerationLarge Language ModelVision Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Proposed a fully diffusion-based multimodal large language model, LLaDA-V, integrating visual instruction tuning with masked diffusion models.
LLaMo: Scaling Pretrained Language Models for Unified Motion Understanding and Generation with Continuous Autoregressive Tokens
Zekun Li (Brown University), Abhay Mittal (Meta)
GenerationTransformerLarge Language ModelFlow-based ModelAuto EncoderMultimodality
🎯 What it does: Extend pre-trained large language models (LLMs) to achieve unified motion-language understanding and generation, supporting text-to-motion and motion-to-text tasks.
LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Models
Guolei Huang (Southeast University), Yongjun Shen (Southeast University)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: This paper addresses the safety evaluation of multi-modal multi-turn dialogues by constructing the MMDS safety dataset and proposing a specialized review model, LLaVAShield.
LLM-Guided Probabilistic Fusion for Label-Efficient Document Layout Analysis
Ibne Farabi Shihab (Iowa State University), Anuj Sharma (Iowa State University)
Object DetectionSegmentationTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: Propose a semi-supervised document layout analysis framework that combines visual detection results with structural priors inferred by a large language model (LLM) through probabilistic fusion.
LLMind: Bio-inspired Training-free Adaptive Visual Representations for Vision-Language Models
Soumyaratna Debnath (Nanyang Technological University), Lin Wang (Nanyang Technological University)
Representation LearningVision Language ModelMultimodality
🎯 What it does: Proposes the LLMind framework, which uses a Bio-Inspired Adaptive Sampling Strategy (BASS) based on Möbius transformation and closed-loop semantic feedback (CSF) to adaptively allocate pixel budgets to vision-language models (VLM) without training the model.
LNEM: Lunar Neural Elevation Model
Suwan Lee (KENTECH), Seokju Lee (KENTECH)
Depth EstimationNeural Radiance FieldImageBenchmarkPhysics Related
🎯 What it does: Propose LNEM, a lunar digital elevation model method leveraging a strict pushbroom camera model and differentiable voxel rendering, and release the Lunar Studio multi-track dataset
Local Motion Matters: A Deconstruct-Recompose Paradigm for Reinforcement Learning Pre-training from Videos
Jinwen Wang (Beijing Jiaotong University), Kai Lv (Beijing Jiaotong University)
Reinforcement LearningAuto EncoderWorld ModelOptical FlowVideo
🎯 What it does: Proposes the Deconstruct-Recompose Paradigm (DRP) for pre-training in reinforcement learning by decomposing global actions into local atomic actions and learning their spatiotemporal relationships.
Local Precise Refinement: A Dual-Gated Mixture-of-Experts for Enhancing Foundation Model Generalization against Spectral Shifts
Xi Chen (National University of Defense Technology), Shen Yan (National University of Defense Technology)
SegmentationDepth EstimationDomain AdaptationTransformerMixture of ExpertsImageBenchmark
🎯 What it does: Propose SpectralMoE, a dual-gate Mixture-of-Experts fine-tuning framework addressing spectral shift and spatial heterogeneity in remote sensing images, achieving fine-grained, spatially adaptive feature refinement.
Localizing, Structuring, and Rendering: Bridging 3D and 2D Vision-Language-Action Models for Robotic Manipulation
Yunlong Zhao, Xiu Su
Robotic IntelligenceTransformerVision-Language-Action ModelMultimodalityPoint Cloud
🎯 What it does: Propose the DiffRender-VLA framework, which utilizes differentiable rendering to convert 3D spatial information into 2D images directly usable by vision-language action (VLA) models, achieving end-to-end gradient flow in robotic manipulation;
Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images
Yikun Ji (Shanghai Jiao Tong University), Jianfu Zhang (Shanghai Jiao Tong University)
Anomaly DetectionExplainability and InterpretabilitySupervised Fine-TuningReinforcement LearningVision Language ModelImageTextChain-of-Thought
🎯 What it does: Propose a two-stage 'Locate-Then-Examine' framework based on a vision-language model to locate suspicious regions and re-examine them to determine whether an image is AI-generated;
Locate-then-Sparsify: Attribution Guided Sparse Strategy for Visual Hallucination Mitigation
Tiantian Dang (Chinese Academy of Sciences), Shuhui Wang (Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyVision Language ModelMultimodality
🎯 What it does: Proposes LTS-FS, a hierarchical feature guidance-based visual hallucination mitigation framework that achieves precise feature steering in large vision-language models through hierarchical attribution and sparse modulation.
LocateAnything3D: Vision-Language 3D Detection with Chain-of-Sight
Yunze Man (University of Illinois Urbana Champaign), Zhiding Yu (NVIDIA)
Object DetectionTransformerImageTextMultimodality
🎯 What it does: Propose a monocular 3D detection method LocateAnything3D based on Vision-Language Models (VLM), utilizing Chain-of-Sight to link 2D detection and 3D estimation into a self-regressive token sequence, supporting open-world categories, free text, and visual prompts;
LoD-Loc v3: Generalized Aerial Localization in Dense Cities using Instance Silhouette Alignment
Shuaibang Peng (National University of Defense Technology), Shen Yan (National University of Defense Technology)
Pose EstimationAutonomous DrivingConvolutional Neural NetworkPrompt EngineeringSimultaneous Localization and MappingImage
🎯 What it does: Propose the LoD-Loc v3 method, achieving aerial positioning of low-detail LoD models through instance silhouette alignment, addressing cross-scenario generalization and localization ambiguity in dense urban scenes.
LoFA: Learning to Predict Personalized Prior for Fast Adaptation of Visual Generative Models
Yiming Hao (Chinese University of Hong Kong Shenzhen), Xiaoguang Han (Chinese University of Hong Kong Shenzhen)
GenerationTransformerPrompt EngineeringDiffusion modelFlow-based ModelImageVideoText
🎯 What it does: Propose the LoFA framework, which directly generates complete LoRA weights from user prompts through a two-stage hypernetwork, enabling second-level fast adaptation for visual generation models.
LoG3D: Ultra-High-Resolution 3D Shape Modeling via Local-to-Global Partitioning
Xinran Yang (Nanjing University), Yanwen Guo (Nanjing University)
GenerationConvolutional Neural NetworkTransformerAuto EncoderMesh
🎯 What it does: This paper proposes the LoG3D framework, which performs variational autoencoder reconstruction of 3D shapes at ultra-high resolution (up to 2048³) using unsigned distance fields (UDF).
LogCD: Local-to-global Consistency Distillation for Few-step Image Generation
Qingsong Xie (OPPO AI Center), Haonan Lu (OPPO AI Center)
GenerationKnowledge DistillationDiffusion modelFlow-based ModelGenerative Adversarial NetworkImageOrdinary Differential Equation
🎯 What it does: Propose a two-stage Local-to-Global Consistency Distillation (LogCD) method, compressing large-scale latent diffusion models (LDM) and correction flow models (RFM) to generate high-quality images in only 2~4 steps.
Logit-Margin Repulsion for Backdoor Defense
Zhiguo Yang (University of Science and Technology of China), Wenjie Ruan (University of Science and Technology of China)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposes a general backdoor defense method called Logit-Margin Repulsion (LMR), which significantly suppresses the logits of the backdoor class by using selective cross-entropy and logit margin constraints on clean samples, and further eliminates backdoor channels during the pruning stage, ultimately achieving efficient removal of traditional backdoors and conditional backdoors (quantization/trimming-induced).
LoL: Longer than Longer, Scaling Video Generation to Hour
Justin Cui (UCLA), Cho-Jui Hsieh (UCLA)
GenerationTransformerAuto EncoderVideo
🎯 What it does: Propose a lightweight, training-agnostic multi-head RoPE jittering method to address the attention sink-collapse problem, enabling real-time streaming of infinitely long video generation.
Long-RVOS: A Comprehensive Benchmark for Long-term Referring Video Object Segmentation
Tianming Liang (Sun Yat-sen University), Jian-Fang Hu (Sun Yat-sen University)
SegmentationTransformerVision Language ModelOptical FlowVideoMultimodalityBenchmark
🎯 What it does: Proposed the Long-RVOS long-duration referential video object segmentation benchmark and provided the reference baseline method ReferMo
Long-SCOPE: Fully Sparse Long-Range Cooperative 3D Perception
Jiahao Wang (Tsinghua University), Jianqiang Wang (Hong Kong Polytechnic University)
Object DetectionDepth EstimationAutonomous DrivingTransformer
🎯 What it does: Propose a fully sparse long-term collaborative 3D perception framework named Long-SCOPE
Long-Tail Internet Photo Reconstruction
Yuan Li (Cornell University), Ruojin Cai (Harvard University)
Pose EstimationDepth EstimationTransformerSupervised Fine-TuningImagePoint CloudBenchmark
🎯 What it does: To address the long-tail distribution of internet photos, we propose the MegaDepth-X dataset and a sparse-aware sampling strategy, fine-tuning 3D foundation models to improve reconstruction performance in sparse, noisy, and unevenly distributed long-tail scenarios.
LongStream: Long-Sequence Streaming Autoregressive Visual Geometry
Chong Cheng (HKUST (GZ)), Hao Wang (HKUST (GZ))
Pose EstimationDepth EstimationAutonomous DrivingTransformerSimultaneous Localization and MappingVideoPoint CloudSequential
🎯 What it does: Propose LongStream, a model for long-sequence streaming autoregressive visual geometry, capable of achieving stable kilometer-scale 3D reconstruction in online settings without future information;
LongVideo-R1: Smart Navigation for Low-cost Long Video Understanding
Jihao Qiu (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose LongVideo-R1, design an active reasoning agent based on a multimodal large language model (LLM) that can achieve long video question answering under low computational budget.
LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling
Zuhao Yang (MiroMind AI), Lidong Bing (MiroMind AI)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes LongVT, an end-to-end agent-based framework that enables large-scale multimodal models to achieve self-reflection and fine-grained temporal localization in long videos through alternating visual tool calls and reasoning (iMCoTT), thereby better answering questions with sparse evidence in long videos.
Look Before You Fuse: 2D-Guided Cross-Modal Alignment for Robust 3D Detection
Xiang Li (Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences), Bin Kong (Institute of Intelligent Machines, Hefei Institutes of Physical Science, Chinese Academy of Sciences)
Object DetectionDepth EstimationAutonomous DrivingConvolutional Neural NetworkTransformerImagePoint Cloud
🎯 What it does: This paper proposes a three-stage method based on 2D detection priors (PGDC, DAGF, SGDM), which improves geometric alignment issues in multi-modal BEV fusion by correcting projection errors between point clouds and images.
LookasideVLN: Direction-Aware Aerial Vision-and-Language Navigation
Yuwei Ning (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
Robotic IntelligenceGraph Neural NetworkTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a new aerial vision-and-language navigation (Aerial VLN) method called LookasideVLN, which utilizes directional cues in instructions to achieve more precise spatial reasoning and efficient path planning.
Looking Beyond the Window: Global-Local Aligned CLIP for Training-free Open-Vocabulary Semantic Segmentation
ByeongCheol Lee (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
SegmentationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Proposed a global-local aligned CLIP framework (GLA-CLIP) that does not require fine-tuning during training, addressing issues of window boundary inconsistency and grid artifacts by introducing cross-window global context during sliding window inference.
LoPrune: Efficient Data Pruning for LoRA-Based Fine-Tuning of Vision Transformer
Qiang He (Huazhong University of Science and Technology), Yun Yang (Swinburne University of Technology)
ClassificationObject DetectionTransformerSupervised Fine-TuningImage
🎯 What it does: To address the training redundancy problem in LoRA fine-tuning of visual Transformers on IoT devices, the LoPrune data pruning method is proposed, which can quickly evaluate sample importance and remove redundant samples by using only a single epoch, thus significantly reducing training overhead.
LOREAL: Mitigating Low-Resolution Challenges in Vision-Language Models with Attribute-driven Prompt Self-Distillation
Xucong Wang, Yang Wang (University Of Science And Technology Of China)
ClassificationDomain AdaptationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityChain-of-Thought
🎯 What it does: Enhancing the robustness of vision-language models at low resolution, the LOREAL framework is proposed, which leverages resolution-robust attributes generated by LLMs and learns through cross-modal Meta-Net and dual student self-distillation;
LoST: Level of Semantics Tokenization for 3D Shapes
Niladri Shekhar Dutt (University College London), Xuelin Chen
GenerationRepresentation LearningTransformerDiffusion modelContrastive LearningTextPoint CloudMesh
🎯 What it does: Propose a novel 3D shape segmentation method called LoST, which sorts segmentation by semantic significance, enabling any prefix to decode complete and semantically reasonable shapes; simultaneously, a new semantic alignment loss called RIDA is trained, and an efficient autoregressive generative model LoST-GPT is built using this segmentation.
LottieGPT: Tokenizing Vector Animation for Autoregressive Generation
Junhao Chen (Shenzhen International Graduate School, Tsinghua University), Ruqi Huang (Shenzhen International Graduate School, Tsinghua University)
GenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Proposed the LottieGPT framework to generate editable, resolution-independent vector animations from text, images, or keyframes.
Love Me, Love My Label: Rethinking the Role of Labels in Prompt Retrieval for Visual In-Context Learning
Tianci Luo (Tsinghua University), Chun Yuan (Tsinghua University)
Object DetectionSegmentationGenerationRetrievalTransformerPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningImage
🎯 What it does: Designed the LaPR framework, incorporating label information into prompt retrieval in Vision-and-Language Context Learning (VICL). A mixture-of-experts network and a query-adaptive router are employed to generate label-aware prompt and query embeddings, improving the quality of prompt selection.
Low-Rank Residual Diffusion Models
Junfu Tan, Jiang Yuan
RestorationDiffusion modelImage
🎯 What it does: Propose a low-rank residual diffusion model (LRDM), which projects residuals into a low-rank subspace during near-field image restoration tasks to achieve more efficient and structured recovery.
Low-Rank Test-Time Training for Pre-Trained Point Cloud Models
Ouyangzi Ye (Zhejiang University), Jun Xiao (Nanyang Technological University)
ClassificationDomain AdaptationTransformerPoint Cloud
🎯 What it does: Designed and implemented a lightweight test-time training framework called LoTT-PC based on LoRA to enhance the robustness of 3D point cloud classification models in out-of-distribution (OOD) scenarios.
Low-Resolution Editing is All You Need for High-Resolution Editing
Junsung Lee (Seoul National University), Bohyung Han (University of Wisconsin-Madison)
GenerationSuper ResolutionDiffusion modelImage
🎯 What it does: Proposed a high-resolution image editing task and designed the ScaleEdit framework, achieving precise high-resolution image editing through a block-based approach that leverages low-resolution editing results and a detail transfer function.
LRDUN: A Low-Rank Deep Unfolding Network for Efficient Spectral Compressive Imaging
He Huang (Wuhan University), Wei He (Wuhan University)
RestorationCompressionConvolutional Neural NetworkImagePhysics Related
🎯 What it does: Propose the low-rank deep unfolding network LRDUN to reconstruct compressed hyperspectral images and significantly improve reconstruction quality.
LRHDR: Learning Representation-enhanced HDR Video Reconstruction
Chenzhuo Liao (Nanjing University), Xuemei Hu (Nanjing University)
RestorationRepresentation LearningConvolutional Neural NetworkVideo
🎯 What it does: Propose the LRHDR framework to recover HDR videos from LDR videos with alternating exposure, avoiding explicit cross-exposure alignment and improving reconstruction quality in scenarios with large motion and exposure inconsistency.
LS-ViT: Least-Squares Hessian Based Block Reconstruction for Low-Bit Post-Training Quantization of Vision Transformers
Hyunha Hwang (Seoul National University), Hyuk-Jae Lee (Seoul National University)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: Proposes a block reconstruction framework called LS-ViT based on least squares Hessian estimation for low-bit post-training quantization of Vision Transformers;
LumiMotion: Improving Gaussian Relighting with Scene Dynamics
Joanna Kaleta (Warsaw University of Technology), Marek Kowalski (Microsoft)
RestorationNeural Radiance FieldGaussian SplattingImageVideoBenchmark
🎯 What it does: Inverse rendering using a 2D Gaussian Splatting model in dynamic scenes first learns scene geometry and dynamics through a deformation network. Subsequently, after freezing the geometry, material properties (diffuse reflection, roughness) and environment lighting are jointly optimized, achieving separation and relighting of materials and lighting.
LUMINA: A Multi-Vendor Mammography Benchmark with Energy Harmonization Protocol
Hongyi Pan (Northwestern University), Ulas Bagci (Istanbul University)
ClassificationImage HarmonizationConvolutional Neural NetworkTransformerBiomedical DataBenchmark
🎯 What it does: Created and publicly released a multi-vendor, high-energy and low-energy annotated full-scenario digital mammography image dataset named LUMINA, and proposed a foreground-only CDF histogram matching energy homogenization preprocessing method on this dataset; subsequently conducted systematic baseline evaluations on three tasks: diagnosis, BI-RAD classification, and breast density prediction.
LumiX: Structured and Coherent Text-to-Intrinsic Generation
Xu Han (Huazhong University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: Propose the LumiX framework to achieve multi-attribute (color, albedo, irradiance, depth, normal) consistent generation based on text;
Lumosaic: Hyperspectral Video via Active Illumination and Coded-Exposure Pixels
Dhruv Verma (University of Toronto), Alex Mariakakis (University of Toronto)
Optical FlowVideo
🎯 What it does: Designed and implemented Lumosaic, capable of capturing hyperspectral videos with 31 bands at 30fps and VGA resolution using programmable narrowband LEDs and pixel-level coded exposure.
LuxRemix: Lighting Decomposition and Remixing for Indoor Scenes
Ruofan Liang (Meta Reality Labs), Christian Richardt (Meta Reality Labs)
Image HarmonizationDiffusion modelGaussian SplattingImage
🎯 What it does: Proposes a complete system for separating, harmonizing, and real-time remixing of lighting in multi-view indoor scenes, enabling independent control of each light source.
LVLM-Aided Alignment of Task-Specific Vision Models
Alexander Koebler (Goethe University Frankfurt), Florian Buettner (German Cancer Consortium)
Explainability and InterpretabilityTransformerVision Language ModelImageBiomedical Data
🎯 What it does: Propose a framework that utilizes a large vision-language model to assist in aligning a small task-specific visual model, automatically identifying and correcting the model's dependence on pseudo-relevant features.
Lyapunov Probes for Hallucination Detection in Large Foundation Models
Bozhi Luan (Beihang University), Zhaoxin Fan (Beihang University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelTextMultimodality
🎯 What it does: Proposed a lightweight probe based on Lyapunov stability theory for detecting hallucinations in large-scale language models and multimodal models.
Lynx: Towards High-Fidelity Personalized Video Generation
Shen Sang (ByteDance Inc), Linjie Luo (ByteDance Inc)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageVideoText
🎯 What it does: Developed a model called Lynx that generates high-fidelity personalized videos from a single reference image.
M^3KG-RAG: Multi-hop Multimodal Knowledge Graph-enhanced Retrieval-Augmented Generation
Hyeongcheol Park (Korea University), Sangpil Kim (Korea University)
GenerationRetrievalGraph Neural NetworkTransformerLarge Language ModelAgentic AIVision Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Propose a multi-hop, multimodal knowledge graph enhanced retrieval-augmented generation framework (M3 KG-RAG), which constructs multi-hop, multimodal knowledge graphs through a lightweight multi-agent pipeline and achieves modality-aware retrieval and fine-grained pruning to improve reasoning and answer credibility in multimodal large language models.
M3DLayout: A Multi-Source Dataset of 3D Indoor Layouts and Structured Descriptions for 3D Generation
Yiheng Zhang (Tsinghua University), Yuwang Wang (Tsinghua University)
GenerationData SynthesisLarge Language ModelDiffusion modelTextMultimodalityPoint CloudMesh
🎯 What it does: Built a large-scale multi-source 3D indoor layout dataset named M3DLayout, providing corresponding structured text descriptions for text-driven 3D scene generation.
M3DocDep: Multi-modal, Multi-page, Multi-document Dependency Chunking with Large Vision-Language Models
Joongmin Shin (Korea University), Heuiseok Lim (Korea University)
RetrievalTransformerVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Developed the M3DOCDEP pipeline, which leverages large vision-language models to first recover block-level dependency trees in multi-page industrial documents, then generates structured blocks based on the tree structure to enhance the effectiveness of retrieval-augmented generation (RAG).
M3Grounder: Mask-Based Multi-Span and Multi-Granular Grounding for Document QA
Venkata Kesav Venna (BharatGen), Ravi Kiran Sarvadevabhatla (BharatGen)
Object DetectionSegmentationRetrievalTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a hybrid vision-language and segmentation model called M3Grounder for pixel-level, multi-level, multi-paragraph evidence localization in document question answering;
M4-RAG: A Massive-Scale Multilingual Multi-Cultural Multimodal RAG
David Anugraha (Stanford University), Genta Indra Winata (Capital One)
RetrievalTransformerVision Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a large-scale multilingual, multicultural, and multimodal retrieval-augmented generation (RAG) evaluation framework, M4-RAG, to assess retrieval effectiveness for visual question answering tasks across different languages and modalities.
M4-SAM: Multi-Modal Mixture-of-Experts with Memory-Augmented SAM for RGB-D Video Salient Object Detection
Jiyuan Liu (Hangzhou Dianzi University), Zhi Liu (Shandong University)
SegmentationTransformerMixture of ExpertsVideoMultimodality
🎯 What it does: A prompt-free end-to-end framework named M-SAM is proposed for the RGB-D video salient object detection task, which performs salient object segmentation based on SAM2;
M4Human: A Large-Scale Multimodal mmWave Radar Benchmark for Human Mesh Reconstruction
Junqiao Fan (Nanyang Technological University), Fangqiang Ding (Massachusetts Institute of Technology)
Pose EstimationTransformerImageMultimodalityPoint CloudMeshBenchmark
🎯 What it does: This paper proposes M4Human—a large-scale multimodal dataset containing millimeter-wave radar, RGB-D camera, and high-precision MoCap annotations—and develops the first RT-Mesh method for human mesh reconstruction using raw radar tensors (RT).
M4V: Multimodal Mamba for Efficient Text-to-Video Generation
Jiancheng Huang (Meituan University of Technology Sydney), Lin Ma (Meituan University of Technology Sydney)
GenerationDiffusion modelVideoTextMultimodality
🎯 What it does: Proposes a multi-modal diffusion framework M4V based on Mamba, achieving efficient fusion and spatiotemporal modeling of text and video through the MM-DiM module.