CVPR 2026 Papers — Page 21
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
MA-Bench: Towards Fine-grained Micro-Action Understanding
Kun Li (United Arab Emirates University), Dan Guo (Zhejiang University)
RecognitionPose EstimationConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningVision-Language-Action ModelOptical FlowVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes MA-Bench—a benchmark for fine-grained understanding of micro-actions—and MA-Bench-Train, a training corpus, using them to evaluate the performance of 23 multimodal large language models (MLLMs). It generates structured micro-action descriptions and question-answer pairs through a semi-automated process, forming a three-tier evaluation framework (Perception-Comprehension-Reasoning) with 1,000 videos and 12,000 QA pairs.
Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens
Zeyuan Yang (University of Massachusetts Amherst), Chuang Gan (University of Massachusetts Amherst)
Data SynthesisRepresentation LearningLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose the Mirage framework, allowing VLM to generate interleaved reasoning trajectories of latent visual tokens and text without generating pixel-level images.
Machine Unlearning via Adaptive Gradient Reweighting and Multi-stage Objective Optimization
Juxin Lu (Inner Mongolia University), Huaiwen Zhang (Inner Mongolia University)
ClassificationRecognitionSafty and PrivacyImageBiomedical Data
🎯 What it does: This paper proposes a machine unlearning method based on adaptive gradient weighting and multi-stage objective optimization, which can efficiently remove the influence of specified samples or categories from pre-trained models without complete retraining, while maintaining the integrity of the remaining knowledge.
MacTok: Robust Continuous Tokenization for Image Generation
Hengyu Zeng (Fudan University), Jian Pu (Fudan University)
GenerationTransformerAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Propose MacTok, a continuous tokenization method based on ViT, which addresses the posterior collapse problem of KL-VAE under strong compression through random masking, DINO-guided semantic masking, and global/local representation alignment.
MAD: Modality-Adaptive Decoding for Mitigating Cross-Modal Hallucinations in Multimodal Large Language Models
Sangyun Chung (KAIST), Yong Man Ro (KAIST)
TransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose a training-free Modality-Adaptive Decoding (MAD) that eliminates audio-visual cross-modal hallucinations by self-assessing task-relevant modalities and dynamically weighted contrastive decoding.
MAD: Motion Appearance Decoupling for efficient Driving World Models
Ahmad Rahimi (Ecole Polythechnique Federal de Lausanne), Alexandre Alahi (Ecole Polythechnique Federal de Lausanne)
Pose EstimationAutonomous DrivingTransformerVision Language ModelDiffusion modelAuto EncoderVideoText
🎯 What it does: Propose the MAD framework, decomposing a general video diffusion model into motion feedforward (generating pose videos) and appearance synthesizer (rendering realistic RGB videos) to achieve controllable driving world models.
MagicFuse: Single Image Fusion for Visual and Semantic Reinforcement
Hao Zhang (Wuhan University), Jiayi Ma (Wuhan University)
GenerationDiffusion modelImage
🎯 What it does: Propose MagicFuse, which utilizes a single degraded visible image through diffusion models to achieve visual and semantic cross-spectral scene reconstruction.
MAGICIAN: Efficient Long-Term Planning with Imagined Gaussians for Active Mapping
Shiyao Li (cole Nationale des Ponts et Chaussées), Vincent Lepetit (cole Nationale des Ponts et Chaussées)
Robotic IntelligenceGaussian SplattingSimultaneous Localization and MappingPoint Cloud
🎯 What it does: Propose the MAGICIAN framework, achieving efficient long-term planning active mapping through pre-trained occupancy networks and Imagined Gaussians.
MagicQuill V2: Precise and Interactive Image Editing with Layered Visual Cues
Zichen Liu (Hong Kong University Of Science And Technology), Qifeng Chen (Hong Kong University Of Science And Technology)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImageMultimodality
🎯 What it does: Propose the MagicQuill V2 system, which employs hierarchical visual prompts (content layer, spatial layer, structural layer, color layer) to achieve fine-grained image generation and editing.
MajutsuCity: Language-driven Aesthetic-adaptive City Generation with Controllable 3D Assets and Layouts
Zilong Huang (Sun Yat-sen University), Ting Han (Sun Yat-sen University)
GenerationData SynthesisTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityPoint CloudMeshRetrieval-Augmented Generation
🎯 What it does: Propose MajutsuCity — a natural language-based, aesthetically adaptive 3D city generation framework, featuring a four-stage pipeline (scene design, layout generation, asset and material generation, scene synthesis), and implementing an interactive editable MajutsuAgent along with a large-scale multimodal dataset MajutsuDataset.
Make it SING: Analyzing Semantic Invariants in Classifiers
Harel Yadid (Technion Israel Institute of Technology), Guy Gilboa (Technion Israel Institute of Technology)
Explainability and InterpretabilityConvolutional Neural NetworkVision Language ModelContrastive LearningImage
🎯 What it does: Studied mapping the zero space (i.e., directions that do not affect the output) of the final linear layer of a classifier to a multimodal space, and generating equivalent images to explain the semantic information carried by these invariances.
MakeAnything: Harnessing Diffusion Transformers for Multi-Domain Procedural Sequence Generation
Yiren Song (National University of Singapore), Mike Zheng Shou (National University of Singapore)
GenerationTransformerSupervised Fine-TuningDiffusion modelFlow-based ModelMultimodalitySequential
🎯 What it does: In this work, we propose the MakeAnything framework, which can generate multi-step creation tutorials (such as painting, crafts, cooking, etc.) from text or a single image, and support decomposing completed works into visual step sequences;
Making the Classification Explanation Faithful to the Confidence Score
Jian-Xun Mi (Chongqing University of Posts and Telecommunications), Weisheng Li (Chongqing University of Posts and Telecommunications)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a black-box explainable method called MHE based on Metropolis-Hastings sampling, which generates explanation maps highly consistent with model confidence;
Making Training-Free Diffusion Segmentors Scale with the Generative Power
Benyuan Meng (Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
SegmentationDiffusion modelImage
🎯 What it does: Proposes a method for automatically aggregating cross-attention maps and performing per-pixel reweighting, enabling training-agnostic diffusion segmenters to significantly improve segmentation performance as the generative model's capabilities increase.
Mamba Learns in Context: Structure-Aware Domain Generalization for Multi-Task Point Cloud Understanding
Jincen Jiang (Bournemouth University), Xuequan Lu (University of Western Australia)
RestorationDomain AdaptationPoint CloudBenchmark
🎯 What it does: Propose a context learning framework based on Mamba (SADG) for multi-task point cloud domain generalization (reconstruction, denoising, registration), maintaining cross-domain structural consistency through structure-aware serialization and hierarchical domain-aware modeling;
MambaLiteUNet: Cross-Gated Adaptive Feature Fusion for Robust Skin Lesion Segmentation
Md Maklachur Rahman (Texas A&M University), Tracy Hammond (Texas A&M University)
SegmentationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: Proposed a lightweight MambaLiteUNet for skin lesion segmentation
MambaSIC: Mamba-based Stereo Image Compression with Bi-directional Multi-reference Entropy Model
Shiyu Qin (Tsinghua University), Jun Zhang (Hong Kong University of Science and Technology)
CompressionImage
🎯 What it does: Designed a stereo image compression framework based on the Mamba visual state space model, named MambaSIC, and proposed a bidirectional multi-reference entropy model to achieve efficient compression
MAMMA: Markerless Accurate Multi-person Motion Acquisition
Hanz Cuevas Velasquez, Michael J. Black (Carnegie Mellon University)
SegmentationPose EstimationOptimizationConvolutional Neural NetworkTransformerVideo
🎯 What it does: Developed an unmarked multi-view motion capture system called MAMMA, which can accurately reconstruct SMPL-X human poses and shapes from multi-camera videos.
MangoBench: A Benchmark for Multi-Agent Goal-Conditioned Offline Reinforcement Learning
Yi Wang (Sun Yat-sen University), Yulan Guo (Sun Yat-sen University)
Reinforcement LearningContrastive LearningBenchmark
🎯 What it does: Proposed the first offline goal-conditioned multi-agent reinforcement learning framework and created the MangoBench benchmark.
ManifoldGD: Training-Free Hierarchical Manifold Guidance for Diffusion-Based Dataset Distillation
Ayush Roy (University at Buffalo, SUNY), Vishnu Suresh Lokhande (University at Buffalo, SUNY)
Data SynthesisKnowledge DistillationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Proposes a completely untrained diffusion model-based dataset distillation framework called ManifoldGD, which generates synthetic samples that preserve both class semantics and data manifold geometry by applying manifold geometric consistency correction guided by patterns at each denoising step.
ManifoldNeuS: Manifold-aware View Optimizability for Pose-Free Neural Surface Reconstruction
Xinxin Liu (Northwestern Polytechnical University), Qing Wang (Northwestern Polytechnical University)
OptimizationNeural Radiance FieldImage
🎯 What it does: Propose the ManifoldNeuS method, which can jointly optimize camera poses and neural implicit surface geometry without requiring known camera poses, achieving high-quality 3D reconstruction.
MANSION: Multi-floor lANguage-to-3D Scene generatIOn for loNg-horizon tasks
Lirong Che (Tsinghua University), Jian Su (AgiBot)
GenerationData SynthesisTransformerLarge Language ModelAgentic AITextMesh
🎯 What it does: This paper proposes the MANSION framework, which leverages multimodal large language models and geometric solvers to convert natural language descriptions into multi-level, navigable 3D building scenes, and releases the MansionWorld dataset containing over 1,000 multi-story buildings;
Mantis: A Versatile Vision-Language-Action Model with Disentangled Visual Foresight
Yi Yang (Shanghai Jiao Tong University), Zhijie Deng (Shanghai Jiao Tong University)
Robotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Propose the Mantis Visual-Language-Action model, which improves the efficiency and robustness of robotic task execution by decoupling visual foresight from action prediction.
MAPo: Motion-Aware Partitioning of Deformable 3D Gaussian Splatting for High-Fidelity Dynamic Scene Reconstruction
Han Jiao (Zhejiang University), Huaizhong Lin (Zhejiang University)
GenerationGaussian SplattingVideo
🎯 What it does: Propose the MAPo framework, which enhances the quality of dynamic scene reconstruction in variable 3D Gaussian splatting through motion-aware partitioning methods.
Mapping Networks
Lord Sen (National Institute of Technology Rourkela), Shyamapada Mukherjee (National Institute of Technology Rourkela)
ClassificationSegmentationImageTime Series
🎯 What it does: Proposed and validated a mapping network framework that generates target network parameters through low-dimensional latent vectors and fixed mapping weights, thereby avoiding direct training of the target network.
MapReduce LoRA: Advancing the Pareto Front in Multi-Preference Optimization for Generative Models
Chieh-Yun Chen (Georgia Tech), Humphrey Shi (Georgia Tech)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerFlow-based ModelImageText
🎯 What it does: Jointly optimize generative models with multiple reward objectives to enhance their alignment with multi-dimensional human preferences.
MapRoute:Precise-Concept Erasing Mappers via Semantic Routing
Sihao Li (Harbin Institute of Technology), Yunyun Yang (Harbin Institute of Technology)
Representation LearningTransformerDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Insert a lightweight linear Mapper module after the text encoder to perform conditional identity mapping for target concepts, enabling precise concept elimination without modifying the main model parameters.
MAPS: Preserving Vision-Language Representations via Module-Wise Proximity Scheduling for Better Vision-Language-Action Generalization
Chengyue Huang (Georgia Institute of Technology), Zsolt Kira (Georgia Institute of Technology)
Robotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelMultimodality
🎯 What it does: Propose the MAPS (Module-Wise Proximity Scheduling) framework, which improves the fine-tuning process of vision-language-action (VLA) models by applying adjustable proximity constraints to each module of pre-trained vision-language models (VLM), preserving pre-trained knowledge while enhancing generalization capabilities.
MARCO: Navigating the Unseen Space of Semantic Correspondence
Claudia Cuttano (Politecnico di Torino), Stefan Roth (TU Darmstadt)
TransformerImageBenchmark
🎯 What it does: Under the single-encoder framework based on DINOv2, the MARCO model is proposed, adopting a coarse-to-fine hierarchical learning and self-distillation training framework to expand sparse keypoint supervision into dense semantic correspondence across the entire image.
Mario: Multimodal Graph Reasoning with Large Language Models
Yuanfu Sun (New York University Shanghai), Qiaoyu Tan (New York University Shanghai)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityGraph
🎯 What it does: Proposed a two-stage framework named Mario, which uses a graph-structured Vision-Language Model (VLM) to align text and image features, followed by multimodal adaptive reasoning at the node level through a Large Language Model (LLM) and a lightweight router.
MARIS: Marine Open-Vocabulary Instance Segmentation
Bingyu Li (University of Science and Technology of China), Xuelong Li (Institute of Artificial Intelligence TeleAI China Telecom)
SegmentationDepth EstimationTransformerPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: Created the MARIS fine-grained ocean open-vocabulary instance segmentation benchmark, and proposed a unified framework combining the Geometric Prior Enhancement Module (GPEM) and Semantic Alignment Injection Mechanism (SAIM) to address underwater visual degradation and semantic ambiguity.
Mark4D: Temporally-Consistent Watermarking for 4D Gaussian Splatting
Jaejin Lee (Pohang University of Science and Technology), Won Hwa Kim (Pohang University of Science and Technology)
GenerationData SynthesisVision Language ModelNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: Achieve hidden and temporally consistent watermark embedding in the dynamic 4D Gaussian splatting (4DGS) model;
Markovian Scale Prediction: A New Era of Visual Autoregressive Generation
Yu Zhang (Tongji University), Longbing Cao (Macquarie University)
GenerationTransformerAuto EncoderImage
🎯 What it does: Propose a visual autoregressive generative model named Markov-VAR, which modifies the scale prediction into a Markov process and introduces a sliding window historical compensation mechanism.
MarkushGrapher-2: End-to-end Multimodal Recognition of Chemical Structures
Tim Strohmeyer (IBM Research), Peter Staar (IBM Research)
RecognitionDrug DiscoveryTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose MarkushGrapher-2, achieving end-to-end identification of chemical structures and multimodal Markush structures by integrating image, text, and layout information;
MARSS: Radar Semantic Segmentation via Modular Attention and State Space Models
Fengyu Chen (Tsinghua University), Qingmin Liao (Tsinghua University)
SegmentationAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: Designed a modular attention radar semantic segmentation framework called MARSS to address the anisotropy, sparsity, and noise issues in radar spectra
Mask to Align, Weight to Disambiguate: Reliable Unsupervised Cross-Modal Hashing with Masked-Weight Contrast
Fan Yang (Nanjing University of Finance and Economics), Haikun Xu (Nanjing University of Finance and Economics)
RetrievalRepresentation LearningTransformerContrastive LearningMultimodality
🎯 What it does: Propose an unsupervised cross-modal hashing framework called UWMCH, which achieves robust representations against incomplete observations and semantic ambiguity through pre-fusion token masking and weighted contrastive learning.
MaskAdapt: Learning Flexible Motion Adaptation via Mask-Invariant Prior for Physics-Based Characters
Soomin Park (KAIST), Sung-Hee Lee (KAIST)
Robotic IntelligenceReinforcement LearningDiffusion modelGenerative Adversarial NetworkTextSequentialPhysics Related
🎯 What it does: Propose the MaskAdapt framework, which first trains a mask-invariant base controller and then learns residual strategies on top of it to achieve flexible local action adaptation for physical controllers, supporting motion composition and text-based local tracking.
MaskDexGrasp: Generative Masked Modeling for Part-Aware Dexterous Grasp Synthesis
Binghui Zuo (Southeast University), Yangang Wang (Southeast University)
GenerationData SynthesisConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderTextMultimodalityPoint Cloud
🎯 What it does: Propose a component-oriented palm and finger joint discretization tokenizer and bidirectional masked Transformer, capable of generating editable and controllable multi-finger grasping poses based on object geometry information and text task descriptions.
MaskDiME: Adaptive Masked Diffusion for Precise and Efficient Visual Counterfactual Explanations
Changlu Guo (Technical University of Denmark), Morten Rieger Hannemose (Technical University of Denmark)
Autonomous DrivingExplainability and InterpretabilityDiffusion modelImage
🎯 What it does: Propose MaskDiME, a training-free diffusion model framework that achieves efficient and accurate visual counterfactual explanations by adaptively applying dual masks to locally control updates during the reverse diffusion process.
Masked Auto-Regressive Variational Acceleration: Fast Inference Makes Practical Reinforcement Learning
Yuxuan Gu (Peking University), He Sun (Peking University)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelScore-based ModelImageText
🎯 What it does: This paper proposes the MARVAL framework, which compresses the Masked Auto-Regressive Diffusion model into a single-step generator and further enhances the generation quality through reinforcement learning.
Masked Region Transformer for Layered Image Generation and Editing at Scale
Zhicong Tang, Yuhui Yuan (Canva Research)
GenerationKnowledge DistillationTransformerDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Propose the Masked Region Transformer (MRT) to generate and edit multi-layer transparent images;
Masked Representation Modeling for Domain-Adaptive Segmentation
Wenlve Zhou (South China University of Technology), Biyun MA (South China University of Technology)
SegmentationDomain AdaptationTransformerAuto EncoderImageBenchmark
🎯 What it does: Propose Masked Representation Modeling (MRM), performing random masking and reconstruction in the latent representation space of semantic segmentation as an auxiliary objective for unsupervised domain adaptation (UDA) tasks;
Masked-Diffusion Autoencoders for 3D Medical Vision Representation Learning
Jiachen Tu (University of Illinois Urbana Champaign), Hoifung Poon (Microsoft)
Representation LearningConvolutional Neural NetworkDiffusion modelScore-based ModelAuto EncoderBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose Masked-Diffusion Autoencoders (MDAE), a self-supervised framework for learning unified representations in 3D medical imaging, which learns joint structural and textural representations by simultaneously applying spatial masking and diffusion noise to volumes.
MaskFocus: Focusing Policy Optimization on Critical Steps for Masked Image Generation
Guohui Zhang (MoE Key Lab of BIPC, USTC), Feng Zhao (MoE Key Lab of BIPC, USTC)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageText
🎯 What it does: This paper proposes the MaskFocus framework, which optimizes the sampling process of mask generation models through reinforcement learning, focusing on key sampling steps to enhance image generation quality.
Masking Matters: Unlocking the Spatial Reasoning Capabilities of LLMs for 3D Scene-Language Understanding
Yerim Jeon (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
TransformerLarge Language ModelVision Language ModelPoint Cloud
🎯 What it does: Proposes 3D-SLIM, a decoder masking strategy that can directly replace traditional causal masks, significantly enhancing LLM's reasoning capabilities in 3D scene-language understanding.
Masking Teacher and Reinforcing Student for Distilling Vision-Language Models
Byung-Kwan Lee, Ryo Hachiuma
Knowledge DistillationTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Proposes the Masters framework, which employs teacher weight amplitude masking and progressive recovery (mask-progressive) along with knowledge distillation using offline reinforcement learning (offline RL) to compress large vision-language models.
MASQuant: Modality-Aware Smoothing Quantization for Multimodal Large Language Models
Lulu Hu (Alibaba Cloud Computing, Alibaba Group), Yongliang Tao (Alibaba Cloud Computing, Alibaba Group)
Computational EfficiencyTransformerLarge Language ModelMultimodality
🎯 What it does: Perform post-training quantization on multi-modal large language models, proposing the MASQuant framework to address the mismatch issues of channel-level smoothing in multi-modal scenarios.
MatAnyone 2: Scaling Video Matting via a Learned Quality Evaluator
Peiqing Yang (Nanyang Technological University), Qingyi Tao (SenseTime Research)
SegmentationConvolutional Neural NetworkTransformerVideoBenchmark
🎯 What it does: Proposed the MatAnyone 2 video matting framework, with the core innovation being the learning-based matting quality evaluator (MQE), used for online supervision and offline data filtering, and constructed the VMReal real-time video matting dataset with a scale of 28K videos and 2.4M frames.
Match-and-Fuse: Consistent Generation from Unstructured Image Sets
Kate Feingold (Weizmann Institute of Science), Tali Dekel (Weizmann Institute of Science)
GenerationGraph Neural NetworkDiffusion modelImageText
🎯 What it does: Proposes a zero-shot, no-training-required image collection generation method called Match-and-Fuse, which can generate new collections maintaining cross-image consistency given a source image collection and text prompts.
MatchED: Crisp Edge Detection Using End-to-End, Matching-based Supervision
Bedrettin Cetinkaya (Middle East Technical University), Emre Akbas (Middle East Technical University)
SegmentationConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: Propose MATCHED, a lightweight plug-and-play one-to-one matching supervision module that can seamlessly connect with any edge detection network to achieve end-to-end generation of sharp, single-pixel-wide edges.
Matching Every Pair to Track Every Point: PairFormer for All-Pairs Tracking and Video Trajectory Fields
Guangyang Wu (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
Object TrackingData SynthesisTransformerVideo
🎯 What it does: Investigate full-sequence tracking, proposing PairFormer Transformer to generate full-sequence pixel trajectory fields with a single forward inference;
MatchMask: Mask-Centric Generative Data Augmentation for Label-Scarce Semantic Segmentation
Yuqi Lin, Kaipeng Zhang (Zhejiang University)
SegmentationData-Centric LearningDiffusion modelImage
🎯 What it does: For semantic segmentation tasks with scarce annotations, MatchMask, a mask-centric generative data augmentation method, is proposed. It can synthesize high-quality, alignment-accurate image-mask pairs using a small number of annotated samples, significantly improving segmentation model performance.
MatE: Material Extraction from Single-Image via Geometric Prior
Zeyu Zhang (University of Science and Technology of China), Yang Cao (University of Science and Technology of China)
GenerationDepth EstimationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Provides PBR material extraction from a single real-world image, first using geometric priors for view distortion correction, then employing a dual-branch diffusion model to generate high-fidelity albedo, normal, roughness, and height maps;
Material Magic Wand: Material-Aware Grouping of 3D Parts in Untextured Meshes
Umangi Jain (University of Toronto), Zhiqin Chen (Adobe Research)
SegmentationTransformerContrastive LearningMeshBenchmark
🎯 What it does: Proposed the Material Magic Wand tool, which automatically retrieves and groups other parts with the same material by clicking on a part in textureless, pre-segmented 3D meshes, achieving efficient interactive material assignment;
MatLat: Material Latent Space for PBR Texture Generation
Kyeongmin Yeo (Korea Advanced Institute Of Science And Technology), Minhyuk Sung (Korea Advanced Institute Of Science And Technology)
GenerationData SynthesisDiffusion modelAuto EncoderTextMesh
🎯 What it does: Proposes the MATLAT framework, which generates high-quality PBR textures on a given 3D mesh through a two-stage process utilizing a pre-trained diffusion model.
MatMart: Material Reconstruction of 3D Objects via Diffusion
Xiuchao Wu, Chengfei Lyu
GenerationDiffusion modelImage
🎯 What it does: Proposes the MatMart framework, which achieves two-stage recovery of 3D object materials using diffusion models: first performing per-view material estimation and baking it into UV space, then generating prior-guided materials based on the estimated material and geometric priors, ultimately obtaining high-fidelity PBR materials.
MatPedia: A Universal Generative Foundation for High-Fidelity Material Synthesis
Di Luo (Nankai University), Chunchao Guo (Tencent Hunyuan)
GenerationTransformerVision Language ModelDiffusion modelRectified FlowAuto EncoderImageTextMultimodalityPhysics Related
🎯 What it does: Proposes MatPedia, a unified physically-based rendering (PBR) material generation framework that adopts a joint RGB-PBR representation, enabling three tasks—text-to-material, image-to-material, and intrinsic decomposition—within a single network.
MatSpray: Fusing 2D Material World Knowledge on 3D Geometry
Philipp Langsteiner (University of Tübingen), Hendrik Lensch (University of Tübingen)
GenerationData SynthesisDiffusion modelGaussian SplattingImage
🎯 What it does: Achieving relightable 3D asset reconstruction under arbitrary lighting conditions by projecting 2D PBR material prediction based on diffusion models into a 3D Gaussian sparse representation.
MaxMark: High-Capacity Diffusion-Native Watermarking via Robust and Invertible Latent Embedding
Xuanhang Chang (Harbin Institute of Technology), Yu Li (Zhejiang University)
GenerationSafty and PrivacyDiffusion modelFlow-based ModelImageText
🎯 What it does: Designed and implemented a high-capacity, robust diffusion-native watermarking framework called MaxMark, which can embed a large amount of information in the latent space of LDMs while preserving the latent distribution through invertible neural networks.
MCHDoc: A Comprehensive Benchmark for Reading Multi-Carrier Chinese Historical Documents
Yijun Sheng (Southeast University), Hui Xue (Southeast University)
RecognitionRestorationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs the MCHDoc benchmark for Chinese historical document identification and post-correction, spanning six carriers and three thousand years of history, and systematically evaluates the performance of various large multimodal language models on this benchmark.
MD2E: Modeling Depth-to-Edge Cues for Monocular Metric Depth Estimation
Chao Ning (University of Tokyo), Naoto Yokoya (University of Tokyo)
Depth EstimationTransformerImagePoint Cloud
🎯 What it does: Propose the MD2E method, which converts dense depth labels into edge labels and utilizes the spectral quantile estimator (SQE) to perform metric depth estimation on monocular images without requiring camera intrinsic parameters.
MDCS-MoAME: Multi-directional Composite Scanning with Mixture of Attention and Mamba Experts for Cancer Survival Prediction
Linjie Qu (Xiamen University), Leyi Wei (Macao Polytechnic University)
ClassificationMixture of ExpertsMultimodalityBiomedical Data
🎯 What it does: Propose the MDCS-MoAME model, which uses multi-directional composite scanning combined with Mamba and attention-based mixture of experts for multi-modal survival prediction on WSIs (whole slide images) and genomic data.
MDS-VQA: Model-Informed Data Selection for Video Quality Assessment
Jian Zou (City University of Hong Kong), Kede Ma (Google Inc)
Data-Centric LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningVideo
🎯 What it does: Proposes MDS-VQA, an active data selection framework combining model difficulty prediction with content diversity, to select challenging and diverse video samples for fine-tuning VQA models under limited annotation budgets.
MeanFlow Transformers with Representation Autoencoders
Zheyuan Hu (Sony AI), Stefano Ermon (Stanford University)
GenerationTransformerFlow-based ModelAuto EncoderImageBenchmark
🎯 What it does: Train the MeanFlow (MF) model in the high-dimensional semantic latent space of the Representation Autoencoder (RAE) to achieve single-step or few-step generation while significantly reducing training and inference costs.
MeanFuser: Fast One-Step Multi-Modal Trajectory Generation and Adaptive Reconstruction via MeanFlow for End-to-End Autonomous Driving
Junli Wang (Chinese Academy of Sciences), Qichao Zhang (University of Chinese Academy of Sciences)
Autonomous DrivingTransformerFlow-based ModelTime SeriesBenchmark
🎯 What it does: Propose MeanFuser, an end-to-end multi-modal trajectory generation and adaptive reconstruction framework based on one-shot sampling, for autonomous driving planning.
Measure The Feature Universe: Topology-based Pseudo Labeling and Gravity Consistency for Source-Free Domain Adaptation
Jae Yun Lee (Sogang University), Sung In Cho (Sogang University)
Data SynthesisDomain AdaptationImage
🎯 What it does: This paper proposes a source-agnostic domain adaptation method. It first constructs a feature universe (via virtual feature sampling and k-NN graph) to achieve topology-aware pseudo-label propagation, then combines gravity consistency to adaptively adjust consistency regularization, ultimately achieving high-quality self-training on the target domain.
Measuring the (Un)Faithfulness of Concept-Based Explanations
Shubham Kumar (University of Illinois Urbana-Champaign), Narendra Ahuja (University of Illinois Urbana-Champaign)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: This paper proposes the Surrogate Faithfulness (SURF) evaluation framework to measure the credibility of unsupervised concept-based explanation methods (U-CBEM) for the final outputs of deep visual models, and conducts a comprehensive benchmark evaluation of existing U-CBEM based on this framework.
Mechanisms of Object Localization in Vision-Language Models
Timothy Schaumlöffel (Goethe University Frankfurt), Gemma Roig (Goethe University Frankfurt)
Object DetectionExplainability and InterpretabilityTransformerVision Language ModelImage
🎯 What it does: This paper conducts mechanism explainability experiments on two visual-language models, LLaVA-1.5 and InternVL-3.5, to investigate how models achieve object localization, revealing that localization depends on 'containerization' mechanisms, the complementarity of global and local perspectives, and the role of a small number of task-critical attention heads.
Med-CMR: A Fine-Grained Benchmark Integrating Visual Evidence and Clinical Logic for Medical Complex Multimodal Reasoning
Haozhen Gong (National University Of Singapore), Hongwei Bran Li (National University Of Singapore)
TransformerLarge Language ModelVision Language ModelMultimodalityBiomedical DataBenchmark
🎯 What it does: Constructed and released the Med-CMR benchmark, providing 20,653 visual question-answer pairs covering 11 organ systems, 12 imaging modalities, and decomposing medical reasoning into 7 visual dimensions and 4 reasoning dimensions;
MedCLIPSeg: Probabilistic Vision-Language Adaptation for Data-Efficient and Generalizable Medical Image Segmentation
Taha Koleilat (Concordia University), Hassan Rivaz (Concordia University)
SegmentationTransformerVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: Developed MedCLIPSeg, a text-driven medical image segmentation framework that utilizes probabilistic cross-modal attention.
MedFG-VQA: Low-Frequency Memory and Graph Attention for Lightweight Medical VQA
Haowen Gu (Nanjing University of Science and Technology), Fumin Shen (University of Electronic Science and Technology of China)
Graph Neural NetworkTransformerLarge Language ModelVision Language ModelMultimodalityBiomedical DataBenchmark
🎯 What it does: Propose a lightweight medical vision question answering framework MedFG-VQA, leveraging frequency domain memory fusion and graph attention to enhance global and local feature representations.
MedGRPO: Multi-Task Reinforcement Learning for Heterogeneous Medical Video Understanding
Yuhao Su (Northeastern University), Ziyan Wu (United Imaging Intelligence)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoBiomedical DataBenchmark
🎯 What it does: Constructed a large-scale medical video instruction dataset, MedVidBench (531k video-instruction pairs), and proposed the MedGRPO framework. The framework first performs supervised fine-tuning on Qwen2.5-VL-7B, then enhances multi-scale medical video understanding performance through cross-dataset reward normalization and multi-task reinforcement learning guided by a medical LLM judge.
Medic-AD: Towards Medical Vision-Language Model's Clinical Intelligence
Woohyeon Park (AIDAS Laboratory, Seoul National University), Jaeyoung Do (AIDAS Laboratory, Seoul National University)
Anomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposed MEDIC-AD, a multi-stage medical vision-language model, aimed at enhancing anomaly detection, symptom tracking, and visual interpretability capabilities.
MedKCO: Medical Vision-Language Pretraining via Knowledge-Driven Cognitive Orchestration
Chenran Zhang (Southeast University), Yi Zhou (Nanjing University of Science and Technology)
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningTextMultimodalityBiomedical Data
🎯 What it does: Propose a medical vision-language pre-training framework called MedKCO, which integrates two-level knowledge-driven curriculum learning (label hierarchy and description hierarchy) along with an adaptive asymmetric contrastive loss to enhance the model's alignment and representation learning of medical images and text.
MedLIME: A Distribution-Aligned and Evidence-Supported Framework for Medical Saliency Explanations
Raghav Magazine (Carnegie Mellon University), Min Xu (Carnegie Mellon University)
Anomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkTransformerSupervised Fine-TuningAuto EncoderBiomedical DataUltrasound
🎯 What it does: Propose the MedLIME framework, which generates explainable saliency maps for medical image anomaly localization by leveraging generative masking, supervised test-time adaptation, and evidence-supported regularization.
MedLoc-R1: Performance-Aware Curriculum Reward Scheduling for GRPO-Based Medical Visual Grounding
Guangjing Yang (Beijing University of Posts and Telecommunications), Qicheng Lao (Beijing University of Posts and Telecommunications)
Object DetectionReinforcement LearningVision Language ModelBiomedical Data
🎯 What it does: Propose MedLoc-R1, a performance-aware reward scheduling framework designed to enhance GRPO training in medical vision localization, addressing the issue of sparse rewards.
MedMO: Grounding and Understanding Multimodal Large Language Model for Medical Images
Ankan Deria (Mohamed bin Zayed University of Artificial Intelligence), Imran Razzak (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyBenchmark
🎯 What it does: Trained and released MedMO—a medical multimodal large language model based on Qwen3-VL—that achieves unified processing of tasks such as precise alignment of medical images and text, report generation, VQA, and localization through a four-stage post-training process (SFT, instruction tuning, high-resolution image supervision, reinforcement learning).
MedTVT-R1: A Multimodal LLM Empowering Medical Reasoning and Diagnosis
Yuting Zhang (Hong Kong University of Science & Technology), Kaishun Wu (Hong Kong University of Science & Technology)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageTextMultimodalityTabularBiomedical DataElectrocardiogramChain-of-Thought
🎯 What it does: Propose MedTVT-R1, a multimodal large language model capable of simultaneously processing electrocardiograms (ECG), chest X-rays, and laboratory tables, supporting long-text medical reasoning and diagnosis for multiple diseases.
MemFlow: A Lightweight Forward Memorizing Framework for Quick Domain Adaptive Feature Mapping
Jianming Lv (South China University of Technology), Xueqi Cheng (Institute of Computing Technology Chinese Academy of Sciences)
Domain AdaptationSpiking Neural NetworkImageBenchmark
🎯 What it does: Propose a lightweight forward memory framework called MemFlow, which uses a frozen deep backbone network for feature extraction. It achieves feature-to-label memory encoding, distributed storage, and retrieval through randomly connected neurons and spiking propagation. It also supports memory reinforcement on unlabeled data to enable fast domain adaptation.
MEMO: Human-like Crisp Edge Detection Using Masked Edge Prediction
Jiaxin Cheng (Capital One), Yicong Zhou (Capital One)
SegmentationTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes the MASKED EDGE PREDICTION (MEMO) framework, achieving human-like clear edge detection through masked training and confidence ranking inference.
Memory Matters: Boosting Training-Free Zero-Shot Temporal Action Localization with a Learnable Lookup Table
Han Jiang (Xi'an Jiaotong University), Jihua Zhu (Shandong University)
Object DetectionTransformerVision Language ModelVideoText
🎯 What it does: Designed a training-free zero-shot temporal action localization method that dynamically accumulates historical video knowledge through a learnable lookup table during testing, and achieves more precise action localization by adapting lookup table entries and text prototypes via residual learning.
Memory-Augmented Scene Understanding and Exploration for Open-World Aerial Object-Goal Navigation
Jiacong Zhou (Hangzhou Dianzi University), Jun Yu (Harbin Institute of Technology)
TransformerLarge Language ModelVision Language ModelVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Proposes OctMem-Agent to address the problem of target navigation for drones in large-scale outdoor environments using visual and linguistic descriptions.
Memory-Efficient Fine-Tuning Diffusion Transformers via Dynamic Patch Sampling and Block Skipping
Sunghyun Park (Qualcomm AI Research), Seokeon Choi (Qualcomm AI Research)
GenerationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: To address the significant memory consumption of large Diffusion Transformers (DiT) during personalized fine-tuning, this paper proposes the DiTBlockSkip scheme, combining dynamic patch sampling with block skipping and precomputed residual features, significantly reducing training memory usage while maintaining generation quality.
Memory-Efficient Transfer Learning with Fading Side Networks via Masked Dual Path Distillation
Yutong Zhang (Beihang University), Yunhong Wang (Beihang University)
Computational EfficiencyKnowledge DistillationMultimodality
🎯 What it does: Proposes Masking Dual Path Distillation (MDPD), achieving memory-efficient transfer learning by using a lightweight side network and main network to distill knowledge mutually during training; after training, only the main network is retained for inference to avoid inference overhead from the side network.
MER-Tracker: Towards High-Speed 3D Point Tracking via Multi-View Event-RGB Hybrid Cameras
Yiqian Chang (Harbin Institute of Technology), Peixi Peng (Peng Cheng Laboratory)
Object TrackingTransformerImageMultimodalityPoint Cloud
🎯 What it does: Propose MER-Tracker, which utilizes four 30fps RGB cameras and two event cameras synchronized for data collection, achieving high-frame-rate (150fps) 3D point tracking.
MERG3R: A Divide-and-Conquer Approach to Large-Scale Neural Visual Geometry
Leo Kaixuan Cheng (University Of Toronto), Nandita Vijaykumar (University Of Toronto)
Pose EstimationDepth EstimationComputational EfficiencyRepresentation LearningContrastive LearningSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Propose the MERG3R framework, which utilizes training-agnostic chunking and sorting strategies to split large-scale unordered image sets into subsets. Each subset is reconstructed using geometric foundation models (e.g., VGGT, π3, etc.) for local reconstruction. Subsequently, all subsets are concatenated into a consistent global 3D model through global alignment and confidence-weighted bundle adjustment.
Merge3D: Efficient 3D Multimodal LLMs via Joint 2D-3D Token Merging
Tianbo Pan (National University of Singapore), Xinchao Wang (National University of Singapore)
Computational EfficiencyRepresentation LearningLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Propose Merge3D, a token merging framework based on 2D semantics and 3D geometric information, to significantly reduce the number of visual tokens in multi-view videos, thereby enhancing the inference speed of 3D multimodal LLMs.
MergeVLA: Cross-Skill Model Merging Toward a Generalist Vision-Language-Action Agent
Yuxia Fu (University of Queensland), Yadan Luo (University of Queensland)
Representation LearningRobotic IntelligenceTransformerSupervised Fine-TuningVision-Language-Action ModelImageVideoTextMultimodality
🎯 What it does: This paper proposes MergeVLA, a vision-language-action (VLA) framework designed to inherently support model merging, enabling multiple single-skill expert models to be combined into a single general-purpose robot control model without requiring retraining.
MERIT: Multi-domain Efficient RAW Image Translation
Wenjun Huang (University of California, Irvine), Mohsen Imani (University of California, Irvine)
Image TranslationTransformerGenerative Adversarial NetworkImage
🎯 What it does: Propose MERIT, a unified multi-domain RAW-to-RAW translation framework capable of performing image translation between any camera domains using a single model;
MERLIN: Building Low-SNR Robust Multimodal LLMs for Electromagnetic Signals
Junyu Shen (Tsinghua University), Maosong Sun (Tsinghua University)
RestorationKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMultimodalityBenchmarkPhysics Related
🎯 What it does: Propose the MERLIN framework to build a robust multimodal large language model (EM MLLM) in low signal-to-noise ratio (SNR) environments.
Mesh-Pro: Asynchronous Advantage-guided Ranking Preference Optimization for Artist-style Quadrilateral Mesh Generation
Zhen Zhou (Institute of Automation, Chinese Academy of Sciences), Chunchao Guo (Tencent Hunyuan)
GenerationOptimizationTransformerReinforcement LearningMesh
🎯 What it does: Propose Mesh-Pro, an asynchronous reinforcement learning (RL)-based autoregressive quadrilateral mesh generation framework, incorporating Advantage-Guided Ranking Preference Optimization (ARPO) and a diagonal-aware hybrid triangulation method for quadrilateral mesh labeling.
Mesh4D: 4D Mesh Reconstruction and Tracking from Monocular Video
Zeren Jiang (University of Oxford), Andrea Vedaldi (University of Oxford)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderGaussian SplattingVideoPoint CloudMesh
🎯 What it does: Achieve complete 3D mesh and motion (4D) reconstruction using monocular video through an encoder-decoder network and diffusion model
MeshFlow: Efficient Artistic Mesh Generation via MeshVAE and Flow-based Diffusion Transformer
Weiyu Li (Meta AI), Andrea Vedaldi (Meta AI)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelRectified FlowAuto EncoderContrastive LearningMesh
🎯 What it does: Propose MeshFlow, a method that encodes a continuous latent space through MeshVAE and uses Rectified Flow Transformer to parallelly generate high-quality artistic 3D meshes.
MeshMosaic: Scaling Artist Mesh Generation via Local-to-Global Assembly
Rui Xu (University of Hong Kong), Taku Komura (University of Hong Kong)
GenerationData SynthesisRecurrent Neural NetworkTransformerPoint CloudMesh
🎯 What it does: Propose MeshMosaic, a framework that splits artistic meshes into local patches, autoregressively generates them, and reassembles them into complete meshes, capable of generating high-resolution artistic meshes with over 100,000 triangles;
MeshRipple: Structured Autoregressive Generation of Artist-Meshes
Junkai Lin (Huazhong University of Science and Technology), Wei Yang (Huazhong University of Science and Technology)
GenerationTransformerMesh
🎯 What it does: This paper proposes an autoregressive mesh generation framework called MeshRipple, which leverages frontier-aware BFS tokenization, expansion prediction, and sparse contextual attention to efficiently generate high-poly, structurally complete artistic meshes.
MeshSplatting: Differentiable Rendering with Opaque Meshes
Jan Held (University of Liège), Andrea Tagliasacchi (Simon Fraser University)
GenerationOptimizationComputational EfficiencyGaussian SplattingMesh
🎯 What it does: Propose MeshSplatting, which performs end-to-end optimization on connected, fully opaque triangular meshes using differentiable rendering to achieve high-quality real-time novel view synthesis.
MeshWeaver: Sparse-Voxel-Guided Surface Weaving for Autoregressive Mesh Generation
Jiale Xu (Tencent), Ying Shan (Tencent)
GenerationTransformerLarge Language ModelMesh
🎯 What it does: Propose the MeshWeaver framework, treating mesh generation as a surface weaving process, predicting the next vertex instead of individual coordinates, achieving shorter sequences;
Meta-CoT: Enhancing Granularity and Generalization in Image Editing
Shiyi Zhang (Shenzhen International Graduate School, Tsinghua University), Yansong Tang (Hunyuan, Tencent)
Image TranslationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the Meta-CoT framework, which enhances the fine-grained reasoning and generalization capabilities of a unified multimodal image editing model through triplet decomposition (task, target, required understanding ability), meta-task decomposition, and the CoT-Editing consistency reward, while constructing a benchmark with 21 editing tasks.
Meta-FC: Meta-Learning with Feature Consistency for Robust and Generalizable Watermarking
Yuheng Li (Yangzhou University), Guodong Long (University Of Technology Sydney)
Meta LearningImage
🎯 What it does: Proposed a training strategy called Meta-FC based on meta-learning and feature consistency to enhance the robustness and generalization performance of deep learning watermark models under various distortion conditions.
Meta-Learning In-Context Enables Training-Free Cross Subject Brain Decoding
Mu Nan (University of Hong Kong), Andrew F. Luo (University of Hong Kong)
RetrievalMeta LearningTransformerSupervised Fine-TuningContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Constructed a training-free, cross-subject, and cross-scanner fMRI visual decoding framework called BrainCoDec, which infers encoding models for each voxel using hierarchical context learning, then decodes image embeddings through functional inversion with multi-voxel context.