ICCV 2025 Papers — Page 15
IEEE/CVF International Conference on Computer Vision · 2701 papers
LLaVA-CoT: Let Vision Language Models Reason Step-by-Step
Guowei Xu (Tsinghua University), Li Yuan (Peking University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: LLaVA-CoT is constructed, a visual language model that achieves autonomous structured reasoning through a four-stage process (summary, image description, reasoning, conclusion), and proposes the SWIRES method for re-tracking search during the testing phase.
LLaVA-KD: A Framework of Distilling Multimodal Large Language Models
Yuxuan Cai (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a method to transfer knowledge from large-scale multimodal language models (l-MLLM) to small-scale multimodal language models (s-MLLM) through knowledge distillation, achieving a significant improvement in the performance of lightweight models.
LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models
Yuzhang Shang (University of Central Florida), Yan Yan (University of Illinois Chicago)
CompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoMultimodality
🎯 What it does: This paper proposes PruMerge, an adaptive visual token compression method for large multimodal models (LMM) that significantly reduces the number of visual tokens without a substantial loss in performance.
LLaVA-SP: Enhancing Visual Representation with Visual Spatial Tokens for MLLMs
Haoran Lou (Beijing University of Posts and Telecommunications), Xinliang Wang (Beihang University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Based on the CLIP-ViT visual encoder, six visual spatial tokens are added to the visual tokens to extract multi-scale spatial information through cropping or pooling, and fine-grained features are enhanced by cross-attention fusion, improving the visual understanding capability of the multimodal large language model (MLLM).
LLM Thought Divergence and Convergence for Dialogue-Based Image Generation Control
Hui Li (Hong Kong Polytechnic University)
GenerationTransformerLarge Language ModelDiffusion modelImageTextChain-of-Thought
🎯 What it does: A dialogue-based image generation control framework LTDC based on LLM is proposed, achieving step-by-step reasoning and evaluation from dialogue to image through the divergence and convergence of thought branches.
LLM-assisted Entropy-based Adaptive Distillation for Unsupervised Fine-grained Visual Representation Learning
Jianfeng Dong (Zhejiang Gongshang University), Xun Wang (Zhejiang Gongshang University)
RetrievalKnowledge DistillationRepresentation LearningConvolutional Neural NetworkLarge Language ModelContrastive LearningImageMultimodality
🎯 What it does: In unsupervised fine-grained visual representation learning, the LEAD framework is proposed, which first uses a large language model to generate fine-grained category descriptions, then utilizes CLIP to generate prototype-driven multimodal knowledge, and adapts the distillation strength through information entropy to achieve a dynamic fusion of teacher model knowledge and self-supervised contrastive learning.
LLM-Assisted Semantic Guidance for Sparsely Annotated Remote Sensing Object Detection
Wei Liao (Nanjing University of Science and Technology), Zhen Cui
Object DetectionTransformerLarge Language ModelImage
🎯 What it does: A sparse annotation remote sensing object detection framework based on large language models (LLM) is proposed, utilizing LLM to generate semantic priors and guide pseudo-label assignment.
LLM-enhanced Action-aware Multi-modal Prompt Tuning for Image-Text Matching
Mengxiao Tian (Beijing Institute of Technology), Shuo Yang (Shenzhen MSU-BIT University)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelContrastive LearningImageText
🎯 What it does: This paper improves the fine-grained action understanding in image-text matching by allowing CLIP to receive action triplets and action state prompts generated by a large language model, proposing an action-aware multimodal prompt tuning method.
LMM-Det: Make Large Multimodal Models Excel in Object Detection
Jincheng Li (360 AI Research), Yuhui Yin (360 AI Research)
Object DetectionTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: This paper proposes LMM-Det, a method that enables large multimodal models to have end-to-end object detection capabilities without adding dedicated detection modules.
LMM4LMM: Benchmarking and Evaluating Large-multimodal Image Generation with LMMs
Jiarui Wang (Shanghai Jiao Tong University), Xiongkuo Min (Shanghai Jiao Tong University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningImageMultimodalityBenchmark
🎯 What it does: A large-scale multimodal image generation evaluation database, EvalMi-50K (20 fine-grained tasks, 50,400 generated images, 2 million+ MOS, and 50,000 QA), and a unified evaluation model based on LMM, LMM4LMM, are proposed for simultaneously assessing perceptual quality, text-image correspondence, and task-specific accuracy.
Local Dense Logit Relations for Enhanced Knowledge Distillation
Liuchi Xu (Northeastern University), Jun Cheng (Shenzhen Institutes of Advanced Technology)
Knowledge DistillationTransformerImage
🎯 What it does: Proposes a Local Dense Relationship Logit Distillation (LDRLD) method that recursively decouples and reorganizes logit information, further enhancing the distillation of key category relationships through Adaptive Decay Weights (ADW).
Local Scale Equivariance with Latent Deep Equilibrium Canonicalizer
Md Ashiqur Rahman (Purdue University), Raymond A. Yeh (Purdue University)
ClassificationSegmentationImage
🎯 What it does: This paper proposes a Deep Equivariant Normalizer (DEC) that normalizes local scale transformations in the network's latent space to enhance the model's equivariance and performance regarding local scales.
LocalDyGS: Multi-view Global Dynamic Scene Modeling via Adaptive Local Implicit Feature Decoupling
Jiahao Wu (Peking University), Ronggang Wang (Peking University)
Gaussian SplattingPoint Cloud
🎯 What it does: This paper studies a multi-view dynamic scene reconstruction method called LocalDyGS, which can simultaneously handle large-scale and fine-scale dynamic motions.
LOCATEdit: Graph Laplacian Optimized Cross Attention for Localized Text-Guided Image Editing
Achint Soni (University of Waterloo), Sirisha Rambhatla (University of Waterloo)
Image TranslationGenerationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes LOCATEdit, which achieves precise and localized text-guided image editing through graph Laplacian regularization of cross-attention maps.
LoD-Loc v2: Aerial Visual Localization over Low Level-of-Detail City Models using Explicit Silhouette Alignment
Juelin Zhu (National University of Defense Technology), Shen Yan (National University of Defense Technology)
SegmentationPose EstimationOptimizationConvolutional Neural NetworkSupervised Fine-TuningSimultaneous Localization and MappingImage
🎯 What it does: LoD-Loc v2 is proposed, utilizing low LoD urban models (LoD1) to achieve aerial visual positioning with drones.
LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models
Haiwen Huang (University of Tübingen), Dan Zhang (Bosch Center for Artificial Intelligence)
SegmentationDepth EstimationSuper ResolutionKnowledge DistillationTransformerImageVideo
🎯 What it does: A coordinate-based cross-attention transformer (LoftUp) is proposed to upsample low-resolution features of visual foundation models (VFM) to high resolution, and to train using high-resolution pseudo labels generated through self-distillation.
LOMM: Latest Object Memory Management for Temporally Consistent Video Instance Segmentation
Seunghun Lee (DGIST), Sunghoon Im (DGIST)
Object TrackingSegmentationTransformerVideo
🎯 What it does: This paper proposes a temporal consistency video instance segmentation framework based on Latest Object Memory (LOM), which can continuously track and update the latest state of objects in each frame, thereby achieving more robust identity management and segmentation in long temporal videos.
Long Context Tuning for Video Generation
Yuwei Guo (Chinese University of Hong Kong), Lu Jiang (ByteDance)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: This paper proposes Long Context Tuning (LCT), which expands the context window of pre-trained single-shot video diffusion models to achieve unified generation and consistency control for scene-level multi-shot videos.
Long-Context State-Space Video World Models
Ryan Po (Princeton University), Xun Huang (Adobe Research)
GenerationData SynthesisRetrievalComputational EfficiencyDiffusion modelWorld ModelVideo
🎯 What it does: A long-term video world model that combines state space models with local attention is proposed to achieve continuous video generation.
Long-LRM: Long-sequence Large Reconstruction Model for Wide-coverage Gaussian Splats
Chen Ziwen (Oregon State University), Zexiang Xu (Hillbot)
GenerationData SynthesisComputational EfficiencyTransformerGaussian SplattingImage
🎯 What it does: This paper proposes Long‑LRM, a feedforward model capable of quickly generating high-quality 360° scene-level Gaussian reconstructions from 32 images at a resolution of 960×540 within 1 second.
Long-Tailed Classification with Multi-Granularity Semantics
Yuting Liu (Tianjin University), Yu Wang (Tianjin University)
ClassificationConvolutional Neural NetworkLarge Language ModelContrastive LearningImage
🎯 What it does: A contrastive learning framework based on large language models was developed to expand label semantics and construct a multi-granularity semantic knowledge graph, improving long-tail classification performance.
Long-term Traffic Simulation with Interleaved Autoregressive Motion and Scenario Generation
Xiuyu Yang (University of Texas Austin), Philipp Krähenbühl (University of Texas Austin)
Autonomous DrivingTransformerSupervised Fine-TuningTime SeriesSequential
🎯 What it does: We propose InfGen, a unified next-token prediction model that can automatically switch between temporal motion simulation and spatial scene generation, achieving closed-loop traffic simulation for up to 30 seconds.
LONG3R: Long Sequence Streaming 3D Reconstruction
Zhuoguang Chen (Shanghai Artificial Intelligence Laboratory), Hang Zhao (Shanghai Artificial Intelligence Laboratory)
TransformerPoint Cloud
🎯 What it does: Proposed the LONG3R framework for long sequence streaming 3D reconstruction.
LongAnimation: Long Animation Generation with Dynamic Global-Local Memory
Nan Chen (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
GenerationData SynthesisDiffusion modelAuto EncoderVideo
🎯 What it does: An automatic coloring method is proposed for long animations (averaging about 500 frames), which achieves long-term color consistency while maintaining smooth local transitions.
LongSplat: Robust Unposed 3D Gaussian Splatting for Casual Long Videos
Chin-Yang Lin (National Yang Ming Chiao Tung University), Yu-Lun Liu (National Yang Ming Chiao Tung University)
Pose EstimationOptimizationGaussian SplattingVideo
🎯 What it does: LongSplat proposes a framework for jointly optimizing camera pose and 3D Gaussian splatting, enabling high-quality view synthesis from long videos without pose information.
Looking in the Mirror: A Faithful Counterfactual Explanation Method for Interpreting Deep Image Classification Models
Townim Chowdhury, Zhibin Liao (Australian Institute for Machine Learning)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes the Mirror-CFE method for generating trustworthy and interpretable contrastive explanation images that align with the decision boundaries of depth image classifiers.
LookOut: Real-World Humanoid Egocentric Navigation
Boxiao Pan (Stanford University), Leonidas J. Guibas (Stanford University)
Pose EstimationAutonomous DrivingRobotic IntelligenceConvolutional Neural NetworkContrastive LearningVideoMultimodalityPoint Cloud
🎯 What it does: This paper proposes the task of predicting future 6D head poses from perspective videos and trains a model to achieve collision-free navigation in dynamic environments.
LoRA-FAIR: Federated LoRA Fine-Tuning with Aggregation and Initialization Refinement
Jieming Bian (University of Florida), Jie Xu (University of Florida)
Federated LearningTransformerSupervised Fine-TuningImage
🎯 What it does: The LoRA-FAIR method is proposed in the context of federated learning, combining LoRA parameter-efficient fine-tuning to achieve model iteration without data sharing among multiple clients.
LoRA.rar: Learning to Merge LoRAs via Hypernetworks for Subject-Style Conditioned Image Generation
Donald Shenaj (Samsung Research and Development Institute UK), Umberto Michieli (University of Padova)
GenerationLarge Language ModelDiffusion modelImage
🎯 What it does: A hypernetwork-based LoRA merging method, LoRA.rar, is proposed for real-time fusion of any subject and style LoRA without the need for any additional optimization, enabling the rapid generation of high-quality images that meet both content and style constraints.
LoRAverse: A Submodular Framework to Retrieve Diverse Adapters for Diffusion Models
Mert Sonmezer (Middle East Technical University), Pinar Yanardag (Virginia Tech)
GenerationRetrievalOptimizationVision Language ModelDiffusion modelImageText
🎯 What it does: The LoRAverse framework is proposed, which uses submodular optimization to select both relevant and diverse models from a vast number of LoRA adapters to enhance the creative diversity of diffusion models.
Loss Functions for Predictor-based Neural Architecture Search
Han Ji (Sichuan University), Yanan Sun (Sichuan University)
Neural Architecture SearchGraph Neural NetworkTransformerBenchmark
🎯 What it does: The system studied eight loss functions (regression, ranking, and weighted) for performance predictors and proposed the PWLNAS method based on piecewise loss to enhance NAS effectiveness.
LOTA: Bit-Planes Guided AI-Generated Image Detection
Hongsong Wang, Jie Gui
ClassificationAnomaly DetectionComputational EfficiencyConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: An AI-generated image detection method called LOTA is proposed, which is based on low-bit plane noise extraction and maximum gradient block selection.
LOTS of Fashion! Multi-Conditioning for Image Generation via Sketch-Text Pairing
Federico Girella (University of Verona), Marco Cristani (University of Verona)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: The LOTS method is proposed, which generates fashion images conditionally through local sketch-text, addressing the attribute confusion problem.
Low-Light Image Enhancement Using Event-Based Illumination Estimation
Lei Sun (Sofia University St. Kliment Ohridski), Luc Van Gool (Sofia University St. Kliment Ohridski)
RestorationImage
🎯 What it does: Utilizing the temporal mapping of events from event cameras to estimate illumination, combined with the Retinex theory to enhance low-light images;
LUDVIG: Learning-Free Uplifting of 2D Visual Features to Gaussian Splatting Scenes
Juliette Marrie (Univ. Grenoble Alpes), Julien Mairal (Univ. Grenoble Alpes)
Object DetectionSegmentationContrastive LearningGaussian SplattingImage
🎯 What it does: This paper proposes a general method for elevating 2D visual features to 3D Gaussian splatting scenes based on aggregation and graph diffusion without learning.
Lumina-Image 2.0: A Unified and Efficient Image Generative Framework
Qi Qin (Shanghai AI Laboratory), Peng Gao (Shenzhen Institutes of Advanced Technology)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper presents Lumina-Image 2.0, a unified and efficient text-to-image generation framework that integrates the unified Next-DiT model, the unified Captioner UniCap, and multi-stage progressive training and inference acceleration strategies.
LUSD: Localized Update Score Distillation for Text-Guided Image Editing
Worameth Chinchuthakun (Vistec), Supasorn Suwajanakorn (Vistec)
GenerationData SynthesisDiffusion modelScore-based ModelImageText
🎯 What it does: This paper proposes an image editing framework LUSD based on score distillation, focusing on text-guided object insertion tasks.
LUT-Fuse: Towards Extremely Fast Infrared and Visible Image Fusion via Distillation to Learnable Look-Up Tables
Xunpeng Yi (Wuhan University), Jiayi Ma (Southeast University)
Image TranslationComputational EfficiencyKnowledge DistillationImageMultimodality
🎯 What it does: The paper proposes a fast infrared-visible image fusion framework called LUT-Fuse based on a learnable lookup table (MM-LUT), utilizing low-order approximation coding and high-order context coding, and achieving efficient fusion through distillation learning.
LV-MAE: Learning Long Video Representations through Masked-Embedding Autoencoders
Ilan Naiman (Amazon), Gerard Medioni (Amazon)
Representation LearningAuto EncoderVideoMultimodality
🎯 What it does: A long video self-supervised learning framework LV-MAE has been developed, which learns long-term temporal representations across segments by performing masked reconstruction on multimodal embeddings of short video segments.
LVAgent: Long Video Understanding by Multi-Round Dynamical Collaboration of MLLM Agents
Boyu Chen (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Yali Wang (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
RetrievalTransformerLarge Language ModelAgentic AIVision Language ModelVideoMultimodality
🎯 What it does: Proposes LVAgent, a multi-round dynamic collaboration multimodal large language model agent framework for long video understanding.
LVBench: An Extreme Long Video Understanding Benchmark
Weihan Wang (Zhipu AI), Jie Tang (Tsinghua University)
RecognitionRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: A long video understanding benchmark LVBench has been constructed, collecting approximately 103 segments of videos longer than 30 minutes, and generating 1,549 multiple-choice questions to systematically evaluate the performance of multimodal models in long-term understanding.
LVFace: Progressive Cluster Optimization for Large Vision Models in Face Recognition
Jinghan You (ByteDance Inc), Jiao Ran (ByteDance Inc)
RecognitionOptimizationTransformerImage
🎯 What it does: This paper proposes LVFace, a face recognition model based on Vision Transformer, and introduces a Progressive Cluster Optimization (PCO) training framework.
Lyra: An Efficient and Speech-Centric Framework for Omni-Cognition
Zhisheng Zhong (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
RecognitionGenerationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageVideoTextMultimodalityAudio
🎯 What it does: Designed and trained Lyra, an efficient, speech-centered multimodal large language model that supports cross-modal understanding and generation of speech, images, videos, and text.
M-Net: MRI Brain Tumor Sequential Segmentation Network via Mesh-Cast
Jiacheng Lu (Capital Normal University), Guoping Huo (China University of Mining and Technology Beijing)
SegmentationConvolutional Neural NetworkRecurrent Neural NetworkTransformerMultimodalitySequentialBiomedical DataMagnetic Resonance Imaging
🎯 What it does: M-Net is proposed, utilizing the Mesh-Cast mechanism and a Two-Phase Sequential training strategy to perform spatiotemporal serialization segmentation of MRI brain tumor slice sequences.
M-SpecGene: Generalized Foundation Model for RGBT Multispectral Vision
Kailai Zhou (Nanjing University), Xun Cao (Nanjing University)
Object DetectionSegmentationTransformerContrastive LearningImageMultimodality
🎯 What it does: This paper proposes M-SpecGene, a general foundational model for RGB-T multispectral vision, which utilizes self-supervised pre-training to learn cross-modal invariant representations, eliminating the need for traditional task-specific handcrafted modules.
M2EIT: Multi-Domain Mixture of Experts for Robust Neural Inertial Tracking
Yan Li (Sun Yat-sen University), Erwei Yin (Defense Innovation Institute, Academy of Military Sciences)
Object TrackingDomain AdaptationConvolutional Neural NetworkRecurrent Neural NetworkMixture of ExpertsSimultaneous Localization and MappingTime SeriesSequential
🎯 What it does: A multi-domain mixture expert model M EIT2 is proposed for inertial tracking, integrating spatial, temporal, and frequency domain features, and achieving adaptive fusion through a multi-domain alignment router.
M2SFormer: Multi-Spectral and Multi-Scale Attention with Edge-Aware Difficulty Guidance for Image Forgery Localization
Ju-Hyeon Nam (Inha University), Sang-Chul Lee (Inha University)
SegmentationAnomaly DetectionTransformerImage
🎯 What it does: This paper proposes a multi-spectral multi-scale attention framework based on Transformer, called M2SFormer, for precise localization of image forgery regions.
MA-CIR: A Multimodal Arithmetic Benchmark for Composed Image Retrieval
Jaeseok Byun (Seoul National University), Taesup Moon (Seoul National University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: The MA-CIR benchmark is proposed to evaluate the combinatorial understanding ability of synthetic image retrieval models in scenarios involving three arithmetic types (addition, elimination, replacement) and seven complex semantic categories (left-right, up-down, spatial reasoning, color, size, action, object reasoning). High-quality retrieval triplets are constructed through manual review and hard negative samples; at the same time, improvements are made to the text encoder based on large language models (LLM), which are fine-tuned on MA-CIR.
MAESTRO: Task-Relevant Optimization via Adaptive Feature Enhancement and Suppression for Multi-task 3D Perception
Changwon Kang (Hanyang University), Jun Won Choi (Seoul National University)
Object DetectionSegmentationAutonomous DrivingOptimizationConvolutional Neural NetworkPoint Cloud
🎯 What it does: The MAESTRO framework is proposed, which combines Class-wise Prototype Generator, Task-Specific Feature Generator, and Scene Prototype Aggregator to generate task-specific features for 3D object detection, BEV map segmentation, and 3D occupancy prediction on the basis of a shared backbone, significantly reducing task conflicts.
Magic Insert: Style-Aware Drag-and-Drop
Nataniel Ruiz, Shlomi Fruchter
Image TranslationDomain AdaptationDiffusion modelImage
🎯 What it does: This paper proposes and implements Magic Insert, a style-aware drag-and-drop insertion method that allows any subject to be dragged from one image to another with a drastically different style while maintaining style consistency and subject identity.
MagicCity: Geometry-Aware 3D City Generation from Satellite Imagery with Multi-View Consistency
Xingbo Yao (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelGaussian SplattingImage
🎯 What it does: Generate geometrically consistent high-quality 3D city models from satellite images by first extracting geometric and texture features, then using a dual-encoder to encode and generate multi-view images through a city-scale multi-view diffusion model, and finally reconstructing the final 3D city using 3D Gaussian splatting.
MagicColor: Multi-Instance Sketch Colorization
Yinhan Zhang (Hong Kong University of Science and Technology), Zeyu Wang (Hong Kong University of Science and Technology)
Image TranslationGenerationDiffusion modelImageVideo
🎯 What it does: The MagicColor framework is proposed to achieve one-time multi-instance line art coloring, maintaining color consistency and improving efficiency.
MagicDrive-V2: High-Resolution Long Video Generation for Autonomous Driving with Adaptive Control
Ruiyuan Gao (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
GenerationAutonomous DrivingTransformerLarge Language ModelDiffusion modelVideoText
🎯 What it does: Achieve high-resolution (up to 848×1600) long-sequence (up to 241 frames) multi-view video generation in autonomous driving scenarios, supporting dual controllability of geometry (lane, obstacle positions, camera trajectory) and text (weather, time, etc.).
MagicHOI: Leveraging 3D Priors for Accurate Hand-object Reconstruction from Short Monocular Video Clips
Shibo Wang (Hong Kong University of Science and Technology), Jie Song (Hong Kong University of Science and Technology)
Object DetectionGenerationPose EstimationDiffusion modelScore-based ModelVideo
🎯 What it does: This paper proposes the MagicHOI framework, which achieves complete 3D reconstruction of hands and objects in short monocular hand-object interaction videos by integrating a zero-shot diffusion model for perspective synthesis priors.
MagicID: Hybrid Preference Optimization for ID-Consistent and Dynamic-Preserved Video Customization
Hengjia Li (Zhejiang University), Deng Cai (Zhejiang University)
GenerationOptimizationVision Language ModelDiffusion modelOptical FlowVideo
🎯 What it does: Proposes the MagicID framework, which utilizes mixed preference optimization and two-stage sampling to achieve personalized video generation while maintaining identity consistency and natural dynamics.
MagicMirror: ID-Preserved Video Generation in Video Diffusion Transformers
Yuechen Zhang (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
GenerationData SynthesisTransformerDiffusion modelImageVideo
🎯 What it does: This paper proposes the MagicMirror framework, which generates identity-preserving videos with dynamic expressions from a single reference image.
MagicMotion: Controllable Video Generation with Dense-to-Sparse Trajectory Guidance
Quanhao Li, Zuxuan Wu
GenerationData SynthesisTransformerDiffusion modelVideoTextBenchmark
🎯 What it does: A controllable video generation framework called MagicMotion is proposed, which transitions from dense to sparse trajectory guidance. It supports three control methods: mask, bounding box, and sparse box, and can generate long videos from a single image.
MaGS: Reconstructing and Simulating Dynamic 3D Objects with Mesh-adsorbed Gaussian Splatting
Shaojie Ma (Zhejiang University), Yi Yang (Zhejiang University)
Gaussian SplattingVideoMesh
🎯 What it does: Proposes the Mesh-adsorbed Gaussian Splatting (MaGS) framework, which jointly accomplishes 3D reconstruction and dynamic simulation from monocular video.
MagShield: Towards Better Robustness in Sparse Inertial Motion Capture Under Magnetic Disturbances
Yunzhe Shao (Tsinghua University), Feng Xu (Tsinghua University)
Pose EstimationRecurrent Neural NetworkTime Series
🎯 What it does: This paper proposes an IMU attitude estimation method named MagShield, specifically designed to address attitude errors caused by magnetic interference in sparse inertial motion capture systems.
Make Me Happier: Evoking Emotions Through Image Diffusion Models
Qing Lin (Nanyang Technological University), Mengmi Zhang (Nanyang Technological University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes an emotion-driven image generation model named EmoEditor, which can transform the emotion of an image into a target emotion while preserving the semantic structure of the source image.
Make Your Training Flexible: Towards Deployment-Efficient Video Models
Chenting Wang (Shanghai Jiao Tong University), Limin Wang (Nanjing University)
OptimizationComputational EfficiencyKnowledge DistillationTransformerVideo
🎯 What it does: Designed the Flux framework to implement Token Optimization, enabling video models to train and infer flexibly under different computing powers and resolutions.
Mamba-3VL: Taming State Space Model for 3D Vision Language Learning
Yuan Wang (Tsinghua University), Zhipeng Zhang (Anyverse Intelligence)
RecognitionObject DetectionSegmentationRobotic IntelligenceTransformerPrompt EngineeringVision Language ModelMultimodalityPoint Cloud
🎯 What it does: Mamba-3VL is proposed, a 3D vision-language framework based on the state space model (Mamba), designed to unify the handling of interactive tasks between 3D vision and natural language (such as visual localization, dense description, question answering, segmentation, etc.)
MambaML: Exploring State Space Models for Multi-Label Image Classification
Xuelin Zhu (Hong Kong Polytechnic University), Bing Wang (Hong Kong Polytechnic University)
ClassificationRecognitionTransformerImage
🎯 What it does: A multi-label image classification framework called MambaML is proposed, which generates label-specific visual representations through the Mamba cross-attention module and bidirectional Mamba decoder, achieving efficient multi-label classification.
MamTiff-CAD: Multi-Scale Latent Diffusion with Mamba+ for Complex Parametric Sequence
Liyuan Deng (Northwestern Polytechnical University), Yilei Shi (Northwestern Polytechnical University)
GenerationData SynthesisTransformerDiffusion modelSequential
🎯 What it does: Proposes the MamTiff-CAD framework, which combines the Mamba+ encoder with a multi-scale Transformer diffusion generator to achieve the encoding, reconstruction, and generation of long sequence parameterized CAD instructions.
MamV2XCalib: V2X-based Target-less Infrastructure Camera Calibration with State Space Model
Yaoye Zhu (Institute for AI Industry Research), Yan Wang (Institute for AI Industry Research)
Autonomous DrivingConvolutional Neural NetworkRecurrent Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: A targetless baseline camera calibration method based on vehicle-side LiDAR, MamV2XCalib, is proposed, which utilizes vehicle motion data to achieve automatic rotational calibration of infrastructure cameras.
Manual-PA: Learning 3D Part Assembly from Instruction Diagrams
Jiahao Zhang (Australian National University), Stephen Gould (Australian National University)
Pose EstimationRobotic IntelligenceTransformerContrastive LearningImagePoint Cloud
🎯 What it does: This paper proposes a 3D part assembly method guided by illustrated assembly manuals.
Marigold-DC: Zero-Shot Monocular Depth Completion with Guided Diffusion
Massimiliano Viola (ETH Zurich), Anton Obukhov (ETH Zurich)
RestorationDepth EstimationDiffusion modelImage
🎯 What it does: This paper proposes a zero-shot depth completion method based on a pre-trained diffusion model, Marigold-DC, which utilizes sparse depth observations to guide the pre-trained monocular depth estimation by optimizing the latent space and affine parameters during inference, thereby obtaining a dense depth map.
MaskControl: Spatio-Temporal Control for Masked Motion Synthesis
Ekkasit Pinyoanuntapong (University of North Carolina at Charlotte), Sergey Tulyakov (Snap Inc.)
GenerationData SynthesisPose EstimationOptimizationTransformerAuto EncoderVideoText
🎯 What it does: Proposes the MaskControl method, which utilizes a trainable masking transformer to achieve high-quality human motion generation driven by text, and provides precise spatiotemporal control for any joint, at any time point, and for any objective function.
MaskHand: Generative Masked Modeling for Robust Hand Mesh Reconstruction in the Wild
Muhammad Usama Saleem (University of North Carolina), Pu Wang (University of North Carolina)
GenerationPose EstimationGraph Neural NetworkTransformerImageMesh
🎯 What it does: Using a single RGB image, we construct a generative mask model called MaskHand, which discretizes hand poses using VQ-MANO and learns the 2D→3D probability distribution through a context-guided mask Transformer, achieving high-precision and robust 3D hand mesh recovery.
MaskSAM: Auto-prompt SAM with Mask Classification for Volumetric Medical Image Segmentation
Bin Xie (Illinois Institute of Technology), Gady Agam (Illinois Institute of Technology)
SegmentationTransformerPrompt EngineeringBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper presents MaskSAM, a prompt-free, mask classification-based Segment Anything Model (SAM) adaptation framework for voxel segmentation in medical imaging; it achieves automatic prompting and predicts semantic labels by generating auxiliary masks, bounding boxes, and classification tags.
Mastering Collaborative Multi-modal Data Selection: A Focus on Informativeness, Uniqueness, and Representativeness
Qifan Yu (Zhejiang University), Yueting Zhuang (Zhejiang University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMultimodality
🎯 What it does: This paper proposes a multimodal data selection framework named DataTailor, which automatically selects the most valuable data subset for instruction fine-tuning based on three principles (information quantity, uniqueness, representativeness), achieving the effect of 'less is more'.
MatchDiffusion: Training-free Generation of Match-Cuts
Alejandro Pardo (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: This paper proposes a training-free method for automatic matching clip generation based on a text-to-video diffusion model, which enables smooth transitions between two scenes with similar structures but different semantics without altering the video content.
MaTe: Images Are All You Need for Material Transfer via Diffusion Transformer
Nisha Huang (Tsinghua University), Xiu Li (National Cheng-Kung University)
Image TranslationGenerationComputational EfficiencyTransformerDiffusion modelImageBenchmark
🎯 What it does: This paper presents MaTe, a completely training-free, text-prompt-free, zero-shot material transfer framework that can directly transfer textures from a single real or synthetic material image to a target object.
MATE: Motion-Augmented Temporal Consistency for Event-based Point Tracking
Han Han (University of Science and Technology of China), Zheng-jun Zha (University of Science and Technology of China)
Object TrackingTransformerVideo
🎯 What it does: A point tracking framework MATE based on event cameras is proposed, which utilizes the high temporal resolution and motion blur-free characteristics of events to achieve continuous and accurate tracking of arbitrary points.
MaterialMVP: Illumination-Invariant Material Generation via Multi-view PBR Diffusion
Zebin He (Shenzhen Campus of Sun Yat-sen University), Wenhan Luo (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelImageMesh
🎯 What it does: MaterialMVP has been developed, a one-stop multi-view PBR material generation model that can generate high-quality albedo, metallic, and roughness maps from 3D meshes and image prompts in one go, ensuring multi-view consistency and geometric alignment.
MaTVLM: Hybrid Mamba-Transformer for Efficient Vision-Language Modeling
Yingyue Li (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)
Computational EfficiencyKnowledge DistillationTransformerVision Language ModelMultimodality
🎯 What it does: Knowledge distillation is performed on a pre-trained vision-language model to construct the MaTVLM hybrid Mamba-Transformer architecture.
MAVFlow: Preserving Paralinguistic Elements with Conditional Flow Matching for Zero-Shot AV2AV Multilingual Translation
Sungwoo Cho (Korea Advanced Institute of Science and Technology), Se-Young Yun (Korea Advanced Institute of Science and Technology)
Image TranslationGenerationData SynthesisTransformerFlow-based ModelGenerative Adversarial NetworkVideoMultimodalityAudio
🎯 What it does: This paper proposes MAVFlow, which utilizes dual-modal guidance (audio speaker embedding + visual emotion embedding) and conditional flow matching (OT-CFM) to achieve zero-shot multilingual audio-visual translation while maintaining speaker identity and emotional consistency.
MAVias: Mitigate any Visual Bias
Ioannis Sarridis (Information Technologies Institute), Christos Diou (Harokopio University of Athens)
ClassificationRecognitionConvolutional Neural NetworkLarge Language ModelVision Language ModelImage
🎯 What it does: This paper proposes a framework named MAVias for open-set visual bias mitigation, which automatically identifies instance-level potential visual biases using a base model and suppresses the impact of these biases during training, preventing the model from becoming overly reliant on different visual features.
MBTI: Masked Blending Transformers with Implicit Positional Encoding for Frame-rate Agnostic Motion Estimation
Jungwoo Huh (Yonsei University), Sanghoon Lee (Yonsei University)
Pose EstimationTransformerAuto EncoderSimultaneous Localization and MappingVideo
🎯 What it does: Proposes the MBTI (Masked Blending Transformers with Implicit Positional Encoding) framework for achieving frame rate-independent human motion estimation in videos with different frame rates.
MC-Bench: A Benchmark for Multi-Context Visual Grounding in the Era of MLLMs
Yunqiu Xu (Zhejiang University), Yi Yang
Object DetectionTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a Multi-context Visual Grounding task and constructs a corresponding benchmark dataset, MC-Bench, to evaluate the capabilities of multimodal large language models in multi-image instance-level grounding.
MCAM: Multimodal Causal Analysis Model for Ego-Vehicle-Level Driving Video Understanding
Tongtong Cheng (Chongqing University), Kai Liu (National University of Defense Technology)
Autonomous DrivingTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: A multi-modal causal analysis model (MCAM) is proposed, which achieves causal structure learning and alignment of visual and language modalities in driving videos by constructing a Directed Acyclic Graph of Driving States (DSDAG);
MCID: Multi-aspect Copyright Infringement Detection for Generated Images
Chuanwei Huang (Peking University), Jie Zhou (Tencent Inc)
ClassificationRetrievalTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Proposes a multi-dimensional copyright infringement detection task and designs a hybrid infringement detection model that can identify four types of generated image infringements: content, style, structure, and IP.
MCOP: Multi-UAV Collaborative Occupancy Prediction
Zefu Lin (University of Chinese Academy of Sciences), Zhaoxiang Zhang (University of Chinese Academy of Sciences)
Object DetectionSegmentationAutonomous DrivingRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningPoint Cloud
🎯 What it does: A multi-UAV collaborative occupancy prediction framework MCOP is proposed, utilizing spatial perception encoders, altitude-aware dimensionality reduction, and dual-mask perception guidance to achieve efficient collaborative occupancy prediction.
MDD: A Dataset for Text-and-Music Conditioned Duet Dance Generation
Prerit Gupta (Purdue University), Aniket Bera (Purdue University)
GenerationData SynthesisTransformerDiffusion modelTextMultimodalityBenchmarkAudio
🎯 What it does: The MDD dataset is proposed, defining two multimodal benchmark tasks: Text-to-Duet and Text-to-Dance Accompaniment.
MDP-Omni: Parameter-free Multimodal Depth Prior-based Sampling for Omnidirectional Stereo Matching
Eunjin Son (Jeonbuk National University), Sang Jun Lee (Electronics and Telecommunications Research Institute)
Depth EstimationConvolutional Neural NetworkMultimodality
🎯 What it does: A parameter-free multimodal deep prior sampling and azimuth-based multi-view volume fusion panoramic stereo matching network MDP-Omni is proposed.
MDP3: A Training-free Approach for List-wise Frame Selection in Video-LLMs
Hui Sun (Nanjing University), Ming Li
Recommendation SystemOptimizationVideoBenchmark
🎯 What it does: A training-independent and model-independent multi-frame selection method MDP 3 is proposed, which can select key frames during video large model inference.
MeasureXpert: Automatic Anthropometric Measurement Extraction from Two Unregistered, Partial, Posed, and Dressed Body Scans
Ran Zhao (Vrije Universiteit Brussel), Adrian Munteanu (Coventry University)
SegmentationData SynthesisPose EstimationPoint CloudMesh
🎯 What it does: A MeasureXpert framework is designed and implemented, which automatically reconstructs naked shapes and accurately extracts body measurements using two unregistered partial scans of point clouds (in arbitrary poses).
Measuring the Impact of Rotation Equivariance on Aerial Object Detection
Xiuyu Wu (Xidian University), Xingchen Hu (National University of Defense Technology)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: A novel detection framework called MessDet is proposed, which strictly maintains rotational equivariance. By improving downsampling, introducing rotational equivariant channel attention, and utilizing a multi-branch head network, it achieves more efficient and accurate object detection in aerial images.
Medical World Model
Yijun Yang (Hong Kong University of Science and Technology), Jieneng Chen
SegmentationGenerationOptimizationConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelDiffusion modelContrastive LearningWorld ModelImageBiomedical DataComputed Tomography
🎯 What it does: Proposes the Medical World Model (MeWM), which predicts post-processed tumor images from pre-processed CT scans and assists in TACE treatment decision-making through a visual language strategy model, tumor generation dynamics model, and inverse dynamics survival analysis model.
MedSegFactory: Text-Guided Generation of Medical Image-Mask Pairs
Jiawei Mao (University of California Santa Cruz), Yuyin Zhou (University of California San Francisco)
SegmentationGenerationData SynthesisDiffusion modelAuto EncoderImageMultimodalityBiomedical DataComputed TomographyUltrasound
🎯 What it does: This paper presents MedSegFactory, a framework based on a dual-stream diffusion model that can generate medical images and their corresponding segmentation masks solely through text prompts, achieving unsupervised pairing generation of image-mask pairs.
MedVSR: Medical Video Super-Resolution with Cross State-Space Propagation
Xinyu Liu (Chinese University of Hong Kong), Ender Konukoglu (ETH Zurich)
RestorationSuper ResolutionOptical FlowVideoBiomedical Data
🎯 What it does: Proposes the MedVSR framework, which addresses the characteristics of medical videos such as frame skipping and jitter, using Cross-State Space Propagation (CSSP) and Internal State Space Reconstruction (ISSR) to achieve high-quality video super-resolution.
MEGA: Memory-Efficient 4D Gaussian Splatting for Dynamic Scenes
Xinjie Zhang (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
CompressionComputational EfficiencyGaussian SplattingVideo
🎯 What it does: The MEGA framework is proposed to achieve memory-efficient 4D Gaussian Splatting for dynamic scene reconstruction and real-time rendering.
MEH: A Multi-Style Dataset and Toolkit for Advancing Egyptian Hieroglyph Recognition
Maksim Golyadkin (AIRI HSE University), Ilya Makarov (AIRI ISP RAS ITMO University)
RecognitionData SynthesisTransformerDiffusion modelImageText
🎯 What it does: A multi-source Egyptian hieroglyph recognition dataset MEH and its tool pyThoth are proposed, and OCR models are evaluated on this dataset to explore the impact of synthetic pre-training and model scale on performance.
Membership Inference Attacks with False Discovery Rate Control
Chenxu Zhao (Iowa State University), Mengdi Huai (Iowa State University)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an attack framework called MIAFdR that can control the false discovery rate (FDR) for membership inference attacks (MIA), integrating it as a pluggable wrapper into existing MIA methods, while providing theoretical guarantees and experimental validation.
MemDistill: Distilling LiDAR Knowledge into Memory for Camera-Only 3D Object Detection
Donghyeon Kwon (POSTECH), Suha Kwak (Samsung Electronics)
Object DetectionAutonomous DrivingKnowledge DistillationTransformerImagePoint Cloud
🎯 What it does: The MemDistill framework is proposed, which distills 3D knowledge obtained from LiDAR training into a 3D detection model that only uses cameras through a scene-dependent memory module.
MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation
Vladislav Bargatin (Lomonosov Moscow State University), Dmitriy Vatolin (MSU Institute for Artificial Intelligence)
Computational EfficiencyRecurrent Neural NetworkOptical FlowImageVideo
🎯 What it does: MEMFOF is a memory-efficient multi-frame optical flow estimation method that can achieve training and inference at 1080p full resolution, significantly reducing memory requirements.
Memory-Efficient 4-bit Preconditioned Stochastic Optimization
Jingyang Li (National University of Singapore), Pan Zhou (Singapore Management University)
OptimizationConvolutional Neural NetworkLarge Language ModelImageText
🎯 What it does: This paper studies and implements a 4-bit low-precision quantized Shampoo optimizer, utilizing Cholesky decomposition and error feedback to significantly reduce the storage requirements of the preconditioner matrix while maintaining training performance.
Memory-Efficient Generative Models via Product Quantization
Jie Shao (Nanjing University), Jianxin Wu (Nanjing University)
GenerationCompressionDiffusion modelImage
🎯 What it does: A method is proposed to compress diffusion models using extended Product Quantization (PQ) to extremely low bit rates (1-2 bits), further improving compression quality through codebook compression, EM calibration, and reinitialization;
MemoryTalker: Personalized Speech-Driven 3D Facial Animation via Audio-Guided Stylization
Hyung Kyu Kim (Chung Ang University), Hak Gu Kim (Chung Ang University)
GenerationTransformerVideoAudio
🎯 What it does: A two-stage voice-driven 3D facial animation framework named MemoryTalker is proposed, which can reflect the speaker's personalized speaking style and generate realistic facial movements solely through audio.
MergeOcc: Bridge the Domain Gap between Different LiDARs for Robust Occupancy Prediction
Zikun Xu (Tsinghua University), Shaobing Xu (Tsinghua University)
SegmentationDomain AdaptationAutonomous DrivingPoint Cloud
🎯 What it does: MergeOcc, a universal 3D occupancy prediction framework, is proposed to address the domain gap between different LiDAR datasets by simultaneously utilizing multiple LiDAR data sources.