CVPR 2026 Papers — Page 23
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
MoRGS: Efficient Per-Gaussian Motion Reasoning for Streamable Dynamic 3D Scenes
Wonjoon Lee (Yonsei University), Sangyoun Lee (Yonsei University)
OptimizationComputational EfficiencyRepresentation LearningGaussian SplattingOptical FlowVideo
🎯 What it does: Propose an online dynamic scene reconstruction framework called MoRGS, which explicitly models the motion of each Gaussian and utilizes sparse optical flow and motion confidence for constraints, enhancing 4D reconstruction quality and temporal consistency.
MorphAny3D: Unleashing the Power of Structured Latent in 3D Morphing
Xiaokun Sun (Nanjing University), Zhenyu Zhang (Nanjing University)
GenerationTransformerMesh
🎯 What it does: Propose MorphAny3D, an untrained 3D morphing framework based on Trellis Structured Latent (SLAT) representation, capable of generating smooth, semantically coherent 3D morphing sequences across different categories, and supporting advanced applications such as decoupling, dual-target, and style transfer.
MorphSeek: Fine-grained Latent Representation-Level Policy Optimization for Deformable Image Registration
Runxun Zhang (Institute of Automation Chinese Academy of Sciences), Jingwei Wei (Institute of Automation Chinese Academy of Sciences)
OptimizationConvolutional Neural NetworkTransformerReinforcement LearningBiomedical DataMagnetic Resonance ImagingComputed TomographyBenchmark
🎯 What it does: Proposed a deformable image registration framework called MorphSeek, redefining the registration problem as a fine-grained strategy optimization process in the latent feature space.
MOS: Mitigating Optical-SAR Modality Gap for Cross-Modal Ship Re-Identification
Yujian Zhao (Beihang University), Guanglin Niu (Beihang University)
RetrievalDomain AdaptationDiffusion modelContrastive LearningMultimodality
🎯 What it does: Propose the MOS framework to address the modality gap problem in cross-modal ship ReID between optical and synthetic aperture radar (SAR) modalities
MOSAIC-GS: Monocular Scene Reconstruction via Advanced Initialization for Complex Dynamic Environments
Svitlana Morkva (ETH Zürich), Marco Hutter (ETH Zürich)
GenerationDepth EstimationOptimizationComputational EfficiencyGaussian SplattingOptical FlowVideo
🎯 What it does: Propose MOSAIC-GS, a dynamic scene reconstruction framework for monocular video, which achieves real-time rendering and high-fidelity reconstruction using 3D Gaussian Splatting.
Mostly Text, Smart Visuals: Asymmetric Text-Visual Pruning for Large Vision-Language Models
Sijie Li (University of Sheffield), Jungong Han (Tsinghua University)
Computational EfficiencyVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a weight quantization method for large vision-language models (LVLMs) called ATV-Pruning, aiming to achieve high sparsity ratios while maintaining performance.
Motion 3-to-4: 3D Motion Reconstruction for 4D Synthesis
Hongyuan Chen (Zhejiang University), Anpei Chen (Zhejiang University)
GenerationData SynthesisTransformerSupervised Fine-TuningVideoPoint CloudMeshBenchmark
🎯 What it does: Propose the Motion 3-to-4 method, which generates static 3D shapes first based on monocular video and an optional reference mesh, then reconstructs motion to infer complete 4D dynamic objects in a single inference step.
Motion-Aware Animatable Gaussian Avatars Deblurring
Muyao Niu (University of Tokyo), Yinqiang Zheng (University of Tokyo)
RestorationGaussian SplattingVideo
🎯 What it does: Proposes an end-to-end method for directly reconstructing sharp and animatable 3D Gaussian avatars from motion-blurred videos.
MotionCrafter: Dense Geometry and Motion Reconstruction with a 4D VAE
Ruijie Zhu (NTU), Chuanxia Zheng (NTU)
GenerationDiffusion modelAuto EncoderVideoPoint Cloud
🎯 What it does: MotionCrafter is a framework based on video diffusion models that can predict dense 3D point clouds and corresponding 3D scene flow from monocular videos in a single step, achieving joint reconstruction of 4D scene geometry and motion.
MotionEdit: Benchmarking and Learning Motion-Centric Image Editing
Yixin Wan (Tencent AI), Dong Yu (Tencent AI)
Image TranslationReinforcement LearningDiffusion modelFlow-based ModelOptical FlowImageVideoBenchmark
🎯 What it does: Propose the MotionEdit dataset and benchmark, focusing on modifying actions and interactions in images according to text instructions while preserving identity and scene consistency.
MotionEnhancer: Leveraging Video Diffusion for Motion-Enhanced Vision-Language Models
Yifan Xu (Beihang University), Zhipeng Chen (Beijing Digital Native Digital City Research Center)
TransformerSupervised Fine-TuningVision Language ModelDiffusion modelVideo
🎯 What it does: Propose MotionEnhancer, which leverages the motion prior of video diffusion models to align the attention of vision-language models, enhancing their ability to understand fine-grained motion.
MotionHiFlow: Text-to-Motion via Hierarchical Flow Matching
Heng Li (Sun Yat-sen University), Jian-Fang Hu (Sun Yat-sen University)
GenerationGraph Neural NetworkTransformerDiffusion modelFlow-based ModelAuto EncoderTextMultimodality
🎯 What it does: Propose MotionHiFlow, a hierarchical flow matching framework that generates text-driven 3D human motions layer by layer from coarse to fine temporal scales;
MotionMaster: Generalizable Text-Driven Motion Generation and Editing
Nan Jiang (Peking University), Yixin Zhu (Peking University)
GenerationConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalitySequential
🎯 What it does: Propose the MotionMaster framework, which fine-tunes a pre-trained multi-modal large language model (MLLM) on large-scale motion data to achieve end-to-end text-driven human motion generation and editing.
MotionScale: Reconstructing Appearance, Geometry, and Motion of Dynamic Scenes with Scalable 4D Gaussian Splatting
Haoran Zhou (National University of Singapore), Gim Hee Lee (National University of Singapore)
GenerationNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: Reconstruct 4D scenes from monocular videos to achieve high-fidelity appearance, accurate geometry, and coherent motion representations.
MotionV2V: Editing Motion in a Video
Ryan Burgert (Google), Nataniel Ruiz (Google)
Object TrackingGenerationDiffusion modelGaussian SplattingVideo
🎯 What it does: Propose MotionV2V, which enables editing of objects, camera motion, and timeline by leveraging sparse trajectories dragged by users in videos.
Motus: A Unified Latent Action World Model
Hongzhe Bi (Tsinghua University), Jun Zhu (Horizon Robotics)
Robotic IntelligenceTransformerMixture of ExpertsVision-Language-Action ModelDiffusion modelAuto EncoderWorld ModelOptical FlowVideoTextMultimodality
🎯 What it does: Propose Motus, a unified latent action world model that integrates five paradigms: world models, vision-language-action models, inverse dynamics, video generation, and joint prediction. It leverages Mixture-of-Transformers to fuse pre-trained experts, combines optical flow-based latent actions for cross-modal learning, and achieves transfer from large-scale unlabeled videos to target robot data through three-stage training and a six-layer data pyramid.
MoVie: Broaden Your Views with Human Motion for Action Detection
Di Yang (University of Science and Technology of China), François Brémond (Inria Center at Universit´ Cˆte d'Azur)
RecognitionPose EstimationTransformerVideoMultimodality
🎯 What it does: Proposes a unified action video framework called MoVie, which first decomposes structured human motion into motion primitives via a motion dictionary, and then enhances the spatiotemporal representation for action detection by guiding visual features through orthogonal projection;
MovieRecapsQA: A Multimodal Open-Ended Video Question-Answering Benchmark
Shaden Shaar (Cornell University), Bharath Hariharan (Cornell University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Designed and released MovieRecaps QA: a multimodal open-ended video QA benchmark, generating approximately 8.2K QA pairs using movie recap videos, subtitles, and summaries, and proposing a fact-based evaluation method without reference answers.
MoVieS: Motion-Aware 4D Dynamic View Synthesis in One Second
Chenguo Lin (Peking University), Yadong Mu (ByteDance)
GenerationData SynthesisDepth EstimationComputational EfficiencyTransformerGaussian SplattingVideoPoint Cloud
🎯 What it does: Proposed the MoVieS model, which can perform 4D dynamic scene instant reconstruction, view synthesis, and 3D point tracking from monocular video within one second.
Moving Border Ownership for Event-based Motion Segmentation
Zhiyuan Hua (University of Maryland, College Park), Yiannis Aloimonos (University of Maryland, College Park)
SegmentationDomain AdaptationConvolutional Neural NetworkRecurrent Neural NetworkSequential
🎯 What it does: Propose a framework for moving boundary ownership detection and motion segmentation based on event cameras, achieving zero-shot transfer to real data through training with synthetic data.
MPL: Match-guided Prototype Learning for Few-shot Action Recognition
Feng Yang (Chongqing University of Posts and Telecommunications), Junwei Han (Chongqing University of Posts and Telecommunications)
RecognitionMeta LearningConvolutional Neural NetworkTransformerVision-Language-Action ModelVideo
🎯 What it does: Propose a Match-guided Prototype Learning (MPL) framework that uses matching information to drive prototype learning, explicitly enhancing class prototype representations for few-shot action recognition and achieving finer-grained alignment through hierarchical matching.
MR-RAG: Multimodal Relevance-Aware Retrieval-Augmented Generation for Medical Visual Question Answering
Xuze Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
RetrievalVision Language ModelContrastive LearningImageTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: This paper proposes a two-stage retrieval-augmented generation framework called MR-RAG, which introduces multimodal relevance discrimination in both the retrieval and generation stages for medical visual question answering tasks.
MR. Illuminate: Zero-Shot Low-Light Image Enhancement with Diffusion Prior
Joshua Cho (University of Illinois Urbana Champaign), David Forsyth (University of Illinois Urbana Champaign)
RestorationDiffusion modelImage
🎯 What it does: This paper proposes MR. Illuminate, a zero-shot low-light image enhancement method that achieves high-quality image restoration using a frozen diffusion model without requiring additional training, optimization, or assumptions.
MRD: Multi-resolution Retrieval-Detection Fusion for High-Resolution Image Understanding
Fan Yang (HITSZ), Kaihao Zhang (HITSZ)
Object DetectionRetrievalVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposed an unsupervised multi-resolution retrieval-detection fusion framework MRD to enhance the understanding ability of MLLM on high-resolution images.
MRI Contrast Enhancement Kinetics World Model
Jindi Kong (Case Western Reserve University), Shuo Li (Case Western Reserve University)
GenerationData SynthesisDiffusion modelWorld ModelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a world model-based MRI contrast enhancement dynamics simulation framework, MRI CEKWorld, to generate continuous time-point contrast-enhanced images from single non-contrast MRI scans, avoiding the need for contrast agent injection.
MS-Temba: Multi-Scale Temporal Mamba for Understanding Long Untrimmed Videos
Arkaprava Sinha (UNC Charlotte), Srijan Das (UNC Charlotte)
Object DetectionVideo
🎯 What it does: This paper proposes a multi-scale temporal Mamba architecture, MS-Temba, for efficient temporal action detection in long-duration, untrimmed videos;
MS^2Gait: A Multi-Scale Spatio-Temporal Fusion Network for LiDAR-based Gait Recognition
Shenyin Xu (Wuhan University), Xin Tian (Wuhan University)
RecognitionTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposed the MS²Gait multi-scale spatiotemporal framework for directly processing LiDAR raw point cloud gait recognition;
MSCD-GS: Motion-Separated Cooperative Deblurring Dynamic Reconstruction via Gaussian Splatting
Yongjian Liao (Huazhong University of Science and Technology), Luxin Yan (Huazhong University of Science and Technology)
RestorationGenerationGaussian SplattingImageVideo
🎯 What it does: Propose a collaborative defocusing 4D Gaussian Splatting method based on motion separation, named MSCD-GS, which can reconstruct high-quality dynamic scenes and perform novel view synthesis from motion-blurred videos captured by monocular cameras.
MSGNav: Unleashing the Power of Multi-modal 3D Scene Graph for Zero-Shot Embodied Navigation
Xun Huang (Xiamen University), Chenglu Wen (Xiamen University)
Robotic IntelligenceGraph Neural NetworkVision Language ModelVision-Language-Action ModelMultimodalityGraph
🎯 What it does: Propose a zero-shot navigation system, MSGNav, based on the multimodal 3D scene graph (M3DSG), which uses visual image edges instead of text relationships to support open vocabulary while retaining visual information.
MSJoE: Jointly Evolving MLLM and Sampler for Efficient Long-Form Video Understanding
Wenhui Tan (Renmin University of China), Jian Luan (Xiaomi Inc.)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the MSJoE framework, which jointly evolves multimodal large language models (MLLM) with a lightweight keyframe sampler to efficiently select the most informative keyframes in long video question answering;
MSPT: Efficient Large-Scale Physical Modeling via Parallelized Multi-Scale Attention
Pedro M. P. Curvo (University of Amsterdam), Maksim Zhdanov (University of Amsterdam)
Computational EfficiencyTransformerMeshBenchmarkPhysics Related
🎯 What it does: Proposed a multi-scale patch Transformer (MSPT) for large-scale physical modeling, achieving parallel processing of local self-attention and global pooling through spherical tree partitioning.
MSRL: Scaling Generative Multimodal Reward Modeling via Multi-Stage Reinforcement Learning
Chenglong Wang (Northeastern University), Tong Xiao (Northeastern University)
Knowledge DistillationReinforcement Learning from Human FeedbackReinforcement LearningVideoTextMultimodalityChain-of-Thought
🎯 What it does: Proposes a multi-stage reinforcement learning framework, MSRL, which first trains a reward model on large-scale text preference data using RL, and then transfers it to multimodal tasks through caption-based RL and cross-modal knowledge distillation, significantly enhancing the performance of multimodal reward models.
MTA: Multimodal Task Alignment for BEV Perception and Captioning
Yunsheng Ma (Bosch Research North America and Bosch Center for Artificial Intelligence), Liu Ren (Bosch Research North America and Bosch Center for Artificial Intelligence)
Autonomous DrivingTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes a multimodal task alignment framework called MTA, which enhances two tasks: 3D perception in BEV views and natural language description (captioning) through joint training.
MU-GeNeRF: Multi-view Uncertainty-guided Generalizable Neural Radiance Fields for Distractor-aware Scene
Wenjie Mu (Tongji University), Guang Chen (Tongji University)
RestorationTransformerNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Propose a multi-view uncertainty-guided generalizable neural radiance field (MU-GeNeRF) that achieves high-quality reconstruction in real-world scenes with temporary occlusions.
MuCo: Multi-turn Contrastive Learning for Multimodal Embedding Model
Geonmo Gu (NAVER AI Lab), Dongyoon Han (NAVER AI Lab)
Computational EfficiencyRepresentation LearningData-Centric LearningSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: This paper proposes a multi-round contrastive learning framework, MuCo, which processes multiple query-target pairs from the same image in one go using a conversational structure, enhancing the training efficiency and effectiveness of multi-modal embeddings.
MUFASA: A Multi-Layer Framework for Slot Attention
Sebastian Bock (TU Darmstadt), Stefan Roth (TU Darmstadt)
SegmentationRepresentation LearningTransformerContrastive LearningImageVideo
🎯 What it does: In unsupervised object learning, multi-layer features from visual Transformers are utilized for multi-layer slot attention, fusing slots from different layers to generate a unified object representation.
MuKV: Multi-Grained KV Cache Compression for Long Streaming Video Question-Answering
Junbin Xiao (University of Science and Technology of China), Angela Yao (National University of Singapore)
RetrievalCompressionTransformerLarge Language ModelVideo
🎯 What it does: Designed and implemented a multi-granularity KV cache compression framework called MuKV to improve the answer accuracy and online inference efficiency for long video question answering.
Multi-Crit: Benchmarking Multimodal Judges on Pluralistic Criteria-Following
Tianyi Xiong (University of Maryland), Heng Huang (University of Maryland)
Large Language ModelPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Propose the Multi-Crit benchmark, which evaluates multi-modal classifiers using multi-dimensional criteria
Multi-Hierarchical Contrastive Spectral Fusion for Multi-View Clustering
Bing Cai (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
Representation LearningAuto EncoderContrastive LearningMultimodality
🎯 What it does: Proposes a multi-view clustering framework named MCSF, which combines multi-level contrastive learning with deep spectral embedding. It learns view-specific encoders, spectral embeddings, and consensus representations, achieving structural preservation, view consistency, and consensus refinement through multi-level contrastive loss.
Multi-level Causal LLM-based Text-to-Motion Generation with Human Alignment
Xiaodong Chen (University of Science and Technology of China), Wu Liu (University of Science and Technology of China)
GenerationReinforcement Learning from Human FeedbackConvolutional Neural NetworkTransformerLarge Language ModelReinforcement LearningAuto EncoderVideoText
🎯 What it does: Built MoTiGA, a multi-layer causal LLM-integrated text-to-action generation framework, aligned with human preferences.
Multi-Metric Representation Learning Strategy Based on Clustering for Fine-Grained Multimodal Sentiment Analysis
Yidan Wang (Hebei University), Xiaoxiao Liu (Hebei University)
ClassificationRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerMultimodality
🎯 What it does: Propose the MMRest model, addressing the overlapping sentiment centers problem in multimodal sentiment analysis through clustering and multi-metric learning.
Multi-modal Frequency Decomposition Network for Semantic Scene Completion
Die Zuo (Tianjin University), Di Lin (Tianjin University)
SegmentationConvolutional Neural NetworkImageMultimodality
🎯 What it does: Proposed a lightweight multi-modal frequency domain decomposition network called MFDNet, achieving semantic scene completion using MAFF and FDC;
Multi-Modal Image Fusion via Intervention-Stable Feature Learning
Xue Wang (Yunnan University), Runzhuo Ma (Hong Kong Polytechnic University)
RestorationConvolutional Neural NetworkMultimodalityBiomedical DataMagnetic Resonance ImagingPositron Emission Tomography
🎯 What it does: Train a multi-modal image fusion network using three masking strategies based on causal intervention, learning features robust to interventions, and performing feature fusion during the fusion stage using Causal Feature Integrator (CFI).
Multi-Modal Representation Learning via Semi-Supervised Rate Reduction for Generalized Category Discovery
Wei He (Beijing University of Posts and Telecommunications), Chun-Guang Li (Beijing University of Posts and Telecommunications)
ClassificationRepresentation LearningVision Language ModelContrastive LearningMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a multi-modal representation learning framework named SSR-2-GCD based on semi-supervised coding rate reduction, which can simultaneously identify both known and unknown classes in scenarios where known and unknown classes are mixed;
Multi-modal Test-time Adaptation via Adaptive Probabilistic Gaussian Calibration
Jinglin Xu (Institute of Software Chinese Academy of Sciences), Fanjiang Xu (Institute of Software Chinese Academy of Sciences)
Domain AdaptationContrastive LearningMultimodality
🎯 What it does: Propose a multi-modal test-time adaptation method based on adaptive probability Gaussian calibration (AdaPGC), which explicitly models class-conditional distributions and continuously updates statistics during inference, supplemented by adaptive contrastive imbalance correction to address distribution asymmetry caused by single-modal distortions.
Multi-Paradigm Collaborative Adversarial Attack Against Multi-Modal Large Language Models
Yuanbo Li (Jiangnan University), Josef Kittler (University of Surrey)
Adversarial AttackLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a multi-paradigm collaborative adversarial attack framework named MPCAttack, which can generate adversarial examples with high transferability against multi-modal large language models (MLLMs).
Multi-Patch Global-to-Local Transformer Architecture For Efficient Flow Matching and Diffusion Model
Quan Dao (Rutgers University), Dimitris Metaxas (Rutgers University)
GenerationComputational EfficiencyTransformerDiffusion modelFlow-based ModelImage
🎯 What it does: Propose a multi-scale Patch Transformer (MPDiT), which uses large Patches in the early network stages to capture global information and then upsamples to small Patches in later stages to refine local details, thereby significantly reducing computational costs while maintaining generation quality.
Multi-Prototype Compactness and Boundary-Aware Synthesis for Unsupervised Anomaly Detection
Kailun Liao (Wuhan University), Jinsheng Xiao (Wuhan University)
Data SynthesisAnomaly DetectionContrastive LearningImage
🎯 What it does: Propose the PGBL framework, combining multi-prototype compact constraints, boundary-aware anomaly synthesis, and discriminative boundary refinement to achieve unsupervised anomaly detection and localization.
Multi-Scale Gaussian-Language Map for Zero-shot Embodied Navigation and Reasoning
Sixian Zhang (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Shuqiang Jiang (University of Chinese Academy of Sciences)
SegmentationRepresentation LearningRobotic IntelligenceVision-Language-Action ModelGaussian SplattingImageTextMultimodalityPoint Cloud
🎯 What it does: This paper proposes a multi-scale Gaussian Language Map (GLMap) for zero-shot embodied navigation and reasoning.
Multi-Scale Gradient-Guided Unrolling Architecture with Adaptive Mamba for Compressive Sensing
Le Yang (Northwestern Polytechnical University), Hongping Gan (Northwestern Polytechnical University)
RestorationImageMagnetic Resonance Imaging
🎯 What it does: Proposed a multi-scale gradient-guided Mamba architecture for deep unfolded networks (DUNs) to achieve compressed sensing (CS) image reconstruction.
Multi-Scale Local Speculative Decoding for Image Generation
Elia Peruzzo (Qualcomm AI Research), Amirhossein Habibian (Qualcomm AI Research)
GenerationTransformerDiffusion modelImage
🎯 What it does: Proposes a multi-scale local speculative decoding (MULO-SD) framework, which generates candidate images using a low-resolution draft model and an upsampler, then parallel verifies and locally resamples them with a high-resolution target model to significantly accelerate autoregressive image generation.
Multi-SpatialMLLM: Multi-Frame Spatial Understanding with Multi-Modal Large Language Models
Runsen Xu (FAIR, Meta), Kevin J. Liang (FAIR, Meta)
Pose EstimationDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageVideoTextMultimodalityPoint CloudBenchmark
🎯 What it does: Proposed and trained a multi-modal large language model called Multi-SpatialMLLM, specifically designed for cross-frame spatial understanding and reasoning;
Multi-speaker Attention Alignment for Multimodal Social Interaction
Liangyang Ouyang (University of Tokyo), Yoichi Sato (University of Tokyo)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose a multi-speaker attention alignment method to enhance the performance of multi-modal large language models in multi-speaker social interaction tasks.
Multi-view Consistent 3D Gaussian Head Avatars 'without' Multi-view Generation
Aviral Chharia (Carnegie Mellon University), Fernando De la Torre (Carnegie Mellon University)
GenerationGenerative Adversarial NetworkGaussian SplattingImagePoint Cloud
🎯 What it does: Achieving high-fidelity, multi-view consistent 3D Gaussian head avatar generation under extremely limited resource conditions.
Multi-view Crowd Tracking Transformer with View-Ground Interactions Under Large Real-World Scenes
Qi Zhang (Shenzhen University), Hui Huang (Shenzhen University)
Object TrackingTransformerVideoBenchmark
🎯 What it does: Propose a multi-perspective crowd tracking model based on Transformer, named MVTrackTrans, and collect and annotate two large-scale real-world multi-perspective tracking datasets, MVCrowdTrack and CityTrack.
Multi-View Hierarchical Alignment Learning for Spatial Transcriptomics
Zhengzhong Zhu (Sichuan University), Jiangping Zhu (Sichuan University)
Representation LearningGraph Neural NetworkTransformerContrastive LearningGraphBiomedical Data
🎯 What it does: Propose a multi-view hierarchical alignment learning framework that integrates spatial coordinates and gene expression information to achieve more robust spatial domain clustering.
Multi-view Pyramid Transformer: Look Coarser to See Broader
Gyeongjin Kang (Sungkyunkwan University), Eunbyung Park (Yonsei University)
GenerationComputational EfficiencyRepresentation LearningTransformerGaussian SplattingImageBenchmark
🎯 What it does: Propose Multi-View Pyramid Transformer (MVP), a scalable Transformer that can process tens to hundreds of images in a single forward pass and rapidly reconstruct large-scale 3D scenes.
MultiAnimate: Pose-Guided Image Animation Made Extensible
Yingcheng Hu (Chinese Academy of Sciences), Songhua Liu (Shanghai Jiao Tong University)
GenerationConvolutional Neural NetworkTransformerDiffusion modelVideo
🎯 What it does: This paper proposes a scalable multi-character pose-driven image animation framework called MultiAnimate, which can generate interactive videos between multiple characters while maintaining the identity consistency of each character.
MultiBanana: A Challenging Benchmark for Multi-Reference Text-to-Image Generation
Yuta Oshima (University of Tokyo), Hiroki Furuta (Google DeepMind)
GenerationVision Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: Proposed and implemented the MultiBanana benchmark to evaluate the performance of multi-reference text-to-image generation models under varying numbers of references, cross-domain scenarios, scale differences, rare concepts, and multilingual instructions.
MultiCrafter: High-Fidelity Multi-Subject Generation via Disentangled Attention and Identity-Aware Preference Alignment
Tao Wu (Zhejiang University), Xi Li (Zhejiang University)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningMixture of ExpertsDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Propose a two-stage separated training multi-agent image generation framework called MultiCrafter. First, improve subject fidelity through explicit position supervision and MoE-LoRA, then optimize human aesthetics and text consistency using online reinforcement learning (IPPO).
Multigrain-aware Semantic Prototype Scanning and Tri-Token Prompt Learning Embraced High-Order RWKV for Pan-Sharpening
Junfeng Li (Shenzhen Campus of Sun Yat-sen University), Wenqi Ren (Shenzhen Campus of Sun Yat-sen University)
Super ResolutionTransformerPrompt EngineeringImage
🎯 What it does: Propose a high-order RWKV model combining multigrain-aware semantic prototype scanning and tri-token prompting for remote sensing image fusion and super-resolution (pan-sharpening) tasks.
Multimodal Causality-Driven Representation Learning for Generalizable Medical Image Segmentation
Xusheng Liang (City University of Hong Kong), Jiebo Luo (CAIR, HKISI, Chinese Academy of Sciences)
SegmentationTransformerVision Language ModelContrastive LearningBiomedical Data
🎯 What it does: Proposed a medical image segmentation framework called MCDRL based on multimodal causal intervention, which utilizes CLIP's visual-textual capabilities to localize lesion candidate regions and constructs a text prompt library for causal intervention.
Multimodal Continual Instruction Tuning with Dynamic Gradient Guidance
Songze Li (Harbin Institute of Technology Institute of Automation Chinese Academy of Sciences), Zhongjie Wang (Harbin Institute of Technology Institute of Automation Chinese Academy of Sciences)
OptimizationRepresentation LearningSupervised Fine-TuningMultimodality
🎯 What it does: Propose a dynamic gradient guidance method to approximate old task gradients in multi-modal continual instruction tuning, reducing catastrophic forgetting.
Multimodal Distribution Matching for Vision-Language Dataset Distillation
Jongoh Jeong (KAIST), Kuk-Jin Yoon (KAIST)
RetrievalComputational EfficiencyKnowledge DistillationVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a multi-modal data distillation method (MDM) that matches multi-modal distributions in a joint embedding space
Multimodal Learning on Low-Quality Data with Conformal Predictive Self-Calibration
Xun Jiang (Tongji University), Xing Xu (Tongji University)
Data-Centric LearningMultimodalityBenchmark
🎯 What it does: Proposed a unified multimodal low-quality data learning framework named CPSC, leveraging conformal prediction to achieve self-calibration mechanisms during training;
Multimodal Protein Language Models for Enzyme Kinetic Parameters: From Substrate Recognition to Conformational Adaptation
Fei Wang (Hefei University of Technology), Jingwen Yang (Hefei University of Technology)
Drug DiscoveryTransformerLarge Language ModelMixture of ExpertsMultimodalityBiomedical Data
🎯 What it does: To predict enzyme kinetic parameters (k_cat, K_m, K_i), this paper proposes a stage-wise multi-modal fusion method called ERBA, which integrates enzyme sequences, substrate chemical structures, and 3D geometric information of active sites into a pre-trained protein language model (PLM) for refined learning.
Multimodal RewardBench 2: Evaluating Omni Reward Models for Interleaved Text and Image
Yushi Hu (FAIR at Meta Superintelligence Labs), Marjan Ghazvininejad (FAIR at Meta Superintelligence Labs)
Reinforcement Learning from Human FeedbackLarge Language ModelMultimodalityBenchmark
🎯 What it does: Designed and released Multimodal RewardBench 2 (MMRB2) - a reward model evaluation benchmark containing text-image hybrid tasks.
Multimodal Semantic Bias Mitigation for Diverse Text-To-3D Generation
Yukuan Min (Xidian University), Cheng Deng (Xidian University)
GenerationLarge Language ModelSupervised Fine-TuningPrompt EngineeringDiffusion modelMultimodalityBenchmark
🎯 What it does: Proposes a framework that first generates 3D assets using a text-to-3D large model, leverages an evaluation model to identify word-level deviations, generates diverse text-3D pairs based on these deviations, and finally fine-tunes the original model with these pairs to enhance diversity and semantic consistency.
MultiModalPFN: Extending Prior-Data Fitted Networks for Multimodal Tabular Learning
Wall Kim (Samsung Electronics), Hanul Kim (Seoul National University of Science and Technology)
ClassificationRepresentation LearningTransformerVision Language ModelGenerative Adversarial NetworkImageTextMultimodalityTabularElectronic Health RecordsFinance Related
🎯 What it does: Proposed a multimodal TabPFN (MMPFN) framework capable of uniformly processing three different modalities of data: tables, images, and text.
Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex
Alexandru Brateanu (University of Manchester), Cosmin Ancuti (University Politehnica Timisoara)
RestorationConvolutional Neural NetworkImage
🎯 What it does: Proposed a lightweight low-light image enhancement framework called Multinex, which utilizes the Retinex residual formula and multi-perspective prior representations to achieve the separation and fusion of exposure and color.
MultiShotMaster: A Controllable Multi-Shot Video Generation Framework
Qinghe Wang (Dalian University of Technology), Xu Jia (Dalian University of Technology)
GenerationData SynthesisTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelVideoText
🎯 What it does: This paper proposes the MultiShotMaster framework, achieving controllable multi-shot video generation. It supports text-driven cross-shot consistency, subject customization with motion control, background customization, and flexible configuration of the number of shots and duration.
MuM: Multi-View Masked Image Modeling for 3D Vision
David Nordström (Chalmers University of Technology), Georg Bökman (University of Amsterdam)
ClassificationSegmentationPose EstimationDepth EstimationTransformerAuto EncoderContrastive LearningImagePoint Cloud
🎯 What it does: Proposed a multi-view masked image modeling (MuM) framework that extends MAE to any number of images from the same scene, using a ViT-L encoder-decoder for 3D visual pre-training.
MUSE: Harnessing Precise and Diverse Semantics for Few-Shot Whole Slide Image Classification
Jiahao Xu (Chongqing University), Nankun Mu (Chongqing University)
ClassificationTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose a framework using Random Multi-Perspective Semantic Enhancement (MUSE) for few-shot whole-slide image classification, combining sample-level fine-grained semantic refinement with multi-perspective text retrieval for model optimization.
Muses: Designing, Composing, Generating Nonexistent Fantasy 3D Creatures without Training
Hexiao Lu (Nanjing University), Zhenyu Zhang (Nanjing University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelTextMultimodalityMesh
🎯 What it does: Propose the Muses framework, which utilizes a 3D skeleton-driven design-combination-generation process to generate fantasy 3D creatures without existing counterparts, without requiring training;
MusicInfuser: Making Video Diffusion Listen and Dance
Susung Hong (University of Washington), Steven M. Seitz (University of Washington)
GenerationDiffusion modelVideoMultimodalityAudio
🎯 What it does: This paper proposes the MusicInfuser method, which adapts music input to generate high-quality dance videos by integrating a zero-initialized audio cross-attention module and LoRA into a pre-trained text-to-video diffusion model.
MUST: Modality-Specific Representation-Aware Transformer for Diffusion-Enhanced Survival Prediction with Missing Modality
Kyungwon Kim (Yonsei University), Dosik Hwang (Yonsei University)
ClassificationTransformerDiffusion modelMultimodalityBiomedical Data
🎯 What it does: Propose the MUST framework to model the missing modality problem in multimodal survival prediction.
MuViT: Multi-Resolution Vision Transformers for Learning Across Scales in Microscopy
Albert Dominguez Mantes, Martin Weigert (ScaDS.AI and TU Dresden)
SegmentationTransformerAuto EncoderBiomedical Data
🎯 What it does: Propose the MUVIT architecture, which maps multi-resolution microscopy images to a unified world coordinate system and implements cross-scale attention within a single Transformer encoder.
MV-Fashion: Towards Enabling Virtual Try-On and Size Estimation with Multi-View Paired Data
Hunor Laczkó (Universitat Autonoma De Barcelona), Meysam Madadi (Computer Vision Center)
Image TranslationGenerationPose EstimationDepth EstimationTransformerNeural Radiance FieldVideoMultimodalityBenchmark
🎯 What it does: This paper constructs and releases the MV-Fashion multi-view synchronized video dataset, and implements baseline experiments on virtual try-on, size estimation, and novel view synthesis on it.
MV-RoMa: From Pairwise Matching into Multi-View Track Reconstruction
Jongmin Lee (Korea Advanced Institute of Science and Technology), Sungjoo Yoo (Seoul National University)
Pose EstimationDepth EstimationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Propose the MV-RoMa model, which can perform dense matching on multi-view images in one go, directly generating consistent correspondences for structural reconstruction.
MV-TAP: Tracking Any Point in Multi-View Videos
Jahyeok Koo (KAIST AI), Seungryong Kim (KAIST AI)
Object TrackingTransformerVideo
🎯 What it does: Proposes a multi-view video-based point tracking framework, MV-TAP, which can robustly predict 2D pixel trajectories and visibility of arbitrary points in synchronized multi-view videos.
MV2UV: Generating High-quality UV Texture Maps with Multiview Prompts
Zheng Zhang (Hisilicon Linx Lab, Huawei), Yuan Liu (Hong Kong University of Science and Technology)
GenerationPrompt EngineeringDiffusion modelAuto EncoderImage
🎯 What it does: Proposes a two-stage texture generation framework MV2UV, which first generates multi-view images using a multi-view diffusion model, and then uses these images as semantic prompts for diffusion-based texture generation in the UV space, automatically completing occluded regions and eliminating view inconsistencies.
MV3DIS: Multi-View Mask Matching via 3D Guides for Zero-Shot 3D Instance Segmentation
Yibo Zhao (Nankai University), Jin Xie (Nanjing University)
SegmentationTransformerVision Language ModelPoint CloudMesh
🎯 What it does: Propose MV3DIS, a zero-shot 3D instance segmentation framework based on 3D-guided multi-view mask matching.
MVGGT: Multimodal Visual Geometry Grounded Transformer for Multiview 3D Referring Expression Segmentation
Changli Wu (Xiamen University), Liujuan Cao (Xiamen University)
SegmentationTransformerVision Language ModelImageTextMultimodalityPoint CloudBenchmark
🎯 What it does: Proposed the multi-view 3D referential expression segmentation (MV-3DRES) task, designed an end-to-end multi-modal visual geometry alignment Transformer (MVGGT) model, and created the MVRefer benchmark dataset.
MVInverse: Feed-forward Multiview Inverse Rendering in Seconds
Xiangzuo Wu (Tsinghua University), Yuan Liu (Hong Kong University of Science and Technology)
RestorationDepth EstimationTransformerOptical FlowImageVideo
🎯 What it does: Propose a fast inverse rendering framework called MVInverse for multi-view scenarios, which can simultaneously infer geometric, material, and lighting attributes, and perform multi-view consistent relighting and material editing within seconds.
mVLM: A Vision Language Model for mNPUs
Zijie Chen (Shanghai Jiao Tong University), Haiming Jin (Shanghai Jiao Tong University)
GenerationComputational EfficiencyVision Language ModelImageText
🎯 What it does: Proposed a lightweight vision-language model µVLM for low-power µNPU, combining OverMod encoder and AttSSM decoder to enable real-time image captioning on µNPU.
MVLM: Template-Free Tracking via Vision-Language Margin Confidence and Memory-Gated Tracking
Dae-Hyeon Park (Inha University), Seung-Hwan Bae (Inha University)
Object TrackingTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a template-agnostic object tracking method entirely based on natural language descriptions, achieving dynamic search strategies through visual-language correlation and temporal memory.
MVP: Multiple View Prediction Improves GUI Grounding
Yunzhu Zhang (Zhejiang University), Linchao Zhu (Zhejiang University)
Object DetectionTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a training-agnostic multi-view prediction (MVP) framework that enhances the stability and accuracy of GUI grounding coordinate prediction through attention-guided view cropping and multi-coordinate clustering methods.
NAF: Zero-Shot Feature Upsampling via Neighborhood Attention Filtering
Loïck Chambon (Valeo.ai), Matthieu Cord (Valeo.ai)
SegmentationDepth EstimationSuper ResolutionImageVideo
🎯 What it does: Proposed a zero-shot feature upsampling method based on Neighborhood Attention Filtering (NAF), which can enhance the low-resolution features output by any Vision Foundation Model (VFM) to arbitrary resolutions without requiring retraining specific to a particular VFM.
NAMI: Efficient Image Generation via Bridged Progressive Rectified Flow Transformers
Yuhang Ma (360 AI Research), Yuhui Yin (360 AI Research)
GenerationTransformerRectified FlowImage
🎯 What it does: Propose a bridge-connected multi-stage rectified flow transformer (NAMI) that achieves high-quality image synthesis through a staged generation process, significantly reducing inference latency;
Nano-EmoX: Unifying Multimodal Emotional Intelligence from Perception to Empathy
Jiahao Huang (Fujian Normal University), Zhide Chen (Fujian Normal University)
RecognitionTransformerVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Designed and implemented the Nano-EmoX multimodal language model and the P2E progressive training framework for unified perception, understanding, and interaction across three hierarchical emotional tasks.
NanoSD: Edge Efficient Foundation Model for Real Time Image Restoration
Subhajit Sanyal (Samsung Research Institute), Amit Satish Unde (Samsung Research Institute)
RestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Propose NanoSD, an efficient diffusion model tailored for edge devices, designed for real-time image restoration;
Narrative Weaver: Towards Controllable Long-Range Visual Consistency with Multi-Modal Conditioning
Zhengjian Yao (Peking University), Yanye Lu (Peking University)
GenerationTransformerLarge Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Propose the Narrative Weaver framework to achieve multi-modal controllable, long-term visual consistency generation.
NaTex: Seamless Texture Generation as Latent Color Diffusion
Zeqiang Lai (MMLab, CUHK), Chunchao Guo (Tencent Hunyuan)
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderPoint CloudMesh
🎯 What it does: NaTex proposes a framework for directly generating textures in 3D space, using a latent color diffusion model to replace traditional multi-view image generation and baking processes;
Native and Compact Structured Latents for 3D Generation
Jianfeng Xiang (Tsinghua University), Jiaolong Yang (USTC)
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderMesh
🎯 What it does: Propose a native 3D representation called O-Voxel based on sparse voxels, and build a sparse compressed VAE (SC-VAE) on it to learn a compact and structured 3D latent space. Subsequently, a large flow-matching DiT model is used to generate high-resolution 3D assets with arbitrary topology and physically rendered materials in this latent space.
Natural Human Motion Recovery by Aligning High-Order Temporal Dynamics from Monocular Videos
Dingkun Wei (Zhejiang University), Xiaowei Zhou (Zhejiang University)
Pose EstimationOptimizationTransformerVideo
🎯 What it does: By estimating high-order temporal dynamics (velocity and acceleration) from monocular videos and applying them as soft constraints to the existing HMR pipeline, global motion trajectory post-processing is achieved, making the recovered human motion more natural and smooth.
NavForesee: A Unified Vision-Language World Model for Hierarchical Planning and Dual-Horizon Navigation Prediction
Fei Liu (Alibaba Group), Mu Xu (Alibaba Group)
Autonomous DrivingTransformerLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelWorld ModelMultimodality
🎯 What it does: Propose NavForesee, integrating hierarchical language planning with a dual-temporal world model based on VLM, to achieve synchronized planning and prediction for navigation tasks.
NEAF: Natural Image Editing with Attention Fusion for Generalizable Test-time Optimization in Text-Guided Image Editing
Jisoo Kim (Hansung University), Heeseok Oh (Hansung University)
GenerationPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes a zero-shot, no-training, no-fine-tuning text-driven image editing framework called NEAF, which can directly convert any pre-trained diffusion model into an editing model;
NeAR: Coupled Neural Asset-Renderer Stack
Hong Li (Beijing University of Aeronautics and Astronautics), Hao Zhao (Tsinghua University)
GenerationData SynthesisTransformerRectified FlowGaussian SplattingImageMesh
🎯 What it does: Proposes a coupled neural asset-renderer stack NeAR, capable of generating relightable 3D assets from a single image and achieving real-time, visually consistent multi-view rendering.
NEC-Diff: Noise-Robust Event-RAW Complementary Diffusion for Seeing Motion in Extreme Darkness
Haoyue Liu (Huazhong University of Science and Technology), Luxin Yan (Huazhong University of Science and Technology)
RestorationDiffusion modelMultimodality
🎯 What it does: Designed and implemented a diffusion-model-based event-RAW composite imaging framework, NEC-Diff, for high-quality image reconstruction in extremely dark environments.
Negative Binomial Variational Autoencoders for Overdispersed Latent Modeling
Yixuan Zhang (Southeast University), Feng Zhou (Renmin University of China)
GenerationRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Propose a variational autoencoder using the negative binomial distribution (NegBio-VAE), modeling overdispersed latent neural spike counts by introducing a discrete tunable dispersion parameter.