ICCV 2025 Papers — Page 7
IEEE/CVF International Conference on Computer Vision · 2701 papers
Disentangling Instance and Scene Contexts for 3D Semantic Scene Completion
Enyu Liu (Huazhong University of Science and Technology), Wenbing Tao (Huazhong University of Science and Technology)
SegmentationAutonomous DrivingPoint CloudBenchmark
🎯 What it does: This paper proposes a dual-stream structure called DISC, which learns instance and scene context separately through instance queries and scene queries in the BEV space, thereby achieving more refined 3D semantic scene completion.
Disrupting Model Merging: A Parameter-Level Defense Without Sacrificing Accuracy
Wei Junhao (RIKEN AIP), Jun Sakuma (RIKEN AIP)
ClassificationGenerationTransformerDiffusion modelImageText
🎯 What it does: A proactive defense method called PaRaMS is proposed to prevent other models from merging at the parameter level without affecting the original functionality of the model.
Dissecting Generalized Category Discovery: Multiplex Consensus under Self-Deconstruction
Luyao Tang (Xiamen University), Yue Huang (Xiamen University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: Proposes the ConGCD method, which enhances general category discovery performance through self-deconstructing visual primitives and multi-route consensus.
DiST-4D: Disentangled Spatiotemporal Diffusion with Metric Depth for 4D Driving Scene Generation
Jiazhe Guo (Tsinghua University), Hao Zhao (Tsinghua University)
GenerationData SynthesisDepth EstimationAutonomous DrivingDiffusion modelImageVideoPoint Cloud
🎯 What it does: The DiST-4D framework is proposed, achieving 4D driving scene generation without the need for per-scene optimization, combining temporal extrapolation and spatial novel view synthesis;
DISTA-Net: Dynamic Closely-Spaced Infrared Small Target Unmixing
Shengdong Han (Nanjing University of Posts and Telecommunications), Yimian Dai (Nankai University)
RestorationObject DetectionSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A dynamic iterative threshold shrinkage network (DISTA-Net) has been designed and implemented for separating densely arranged infrared small targets, achieving sub-pixel level demixing and radiance intensity estimation, and an open-source ecosystem has been released.
DISTIL: Data-Free Inversion of Suspicious Trojan Inputs via Latent Diffusion
Hossein Mirzaei (Ecole Polytechnique Federale de Lausanne), Mackenzie W. Mathis (Ecole Polytechnique Federale de Lausanne)
Explainability and InterpretabilityAdversarial AttackDiffusion modelImageBenchmark
🎯 What it does: A no-training-data, diffusion model-based backdoor trigger inversion method called DISTIL is proposed, which can reconstruct and identify backdoor triggers without using clean samples.
DistillDrive: End-to-End Multi-Mode Autonomous Driving Distillation by Isomorphic Hetero-Source Planning Model
Rui Yu (East China University of Science and Technology), Meng Wang (East China University of Science and Technology)
Autonomous DrivingOptimizationKnowledge DistillationTransformerReinforcement LearningGaussian SplattingMultimodality
🎯 What it does: An end-to-end multimodal autonomous driving framework called DistillDrive is constructed, which enhances planning and decision-making performance through teacher model knowledge distillation, reinforcement learning, and generative models.
Distilling Diffusion Models to Efficient 3D LiDAR Scene Completion
Shengyuan Zhang (Zhejiang University), Lingyun Sun (Zhejiang University)
GenerationAutonomous DrivingComputational EfficiencyKnowledge DistillationDiffusion modelPoint Cloud
🎯 What it does: This paper proposes a sparse LiDAR scene completion method named ScoreLiDAR, which significantly improves sampling speed while maintaining or even enhancing generation quality by distilling existing diffusion models.
Distilling Parallel Gradients for Fast ODE Solvers of Diffusion Models
Beier Zhu (Nanyang Technological University), Chi Zhang (Westlake University)
GenerationData SynthesisOptimizationKnowledge DistillationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: An Ensemble Parallel Direction (EPD) solver is proposed to reduce truncation errors in ODE sampling of diffusion models through parallel gradient evaluations, along with a plugin version called EPD-Plugin.
DisTime: Distribution-based Time Representation for Video Large Language Models
Yingsen Zeng (Meituan Inc), Yang Liu (Peking University)
GenerationRetrievalTransformerLarge Language ModelVideo
🎯 What it does: Proposes the DisTime framework, which achieves continuous time representation using a single time marker, and implements event boundary localization through a distributed time decoder and encoder.
DiT4SR: Taming Diffusion Transformer for Real-World Image Super-Resolution
Zheng-Peng Duan (Nankai University), Chongyi Li (Nankai University)
RestorationSuper ResolutionTransformerDiffusion modelImage
🎯 What it does: The DiT4SR method is proposed, which directly applies large-scale Diffusion Transformer to the task of super-resolution for real scene images.
Dita: Scaling Diffusion Transformer for Generalist Vision-Language-Action Policy
Zhi Hou (Shanghai AI Lab), Yuntao Chen (HKISI)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelDiffusion modelImageVideoMultimodality
🎯 What it does: A general robot learning framework called Dita based on diffusion transformers is proposed, capable of achieving zero-shot and 10-shot adaptation for long-sequence tasks using only third-person cameras.
DiTaiListener: Controllable High Fidelity Listener Video Generation with Diffusion
Maksim Siniukov (Institute for Creative Technologies), Mohammad Soleymani (Institute for Creative Technologies)
GenerationData SynthesisDiffusion modelVideoMultimodalityAudio
🎯 What it does: An end-to-end video diffusion model called DiTaiListener is proposed, which can generate high-fidelity, natural listener facial videos from the speaker's audio, facial actions, and optional text prompts, supporting long-term continuous generation.
DiTFastAttnV2: Head-wise Attention Compression for Multi-Modality Diffusion Transformers
Hanling Zhang (Tsinghua University), Yu Wang (Tsinghua University)
GenerationCompressionTransformerDiffusion modelImageMultimodality
🎯 What it does: A head-based attention compression framework DiTFastAttnV2 is proposed for the Multi-Modal Diffusion Transformer (MMDiT), which includes head-level Arrow Attention, head-wise caching, and an efficient kernel that integrates multiple strategies.
DIVE: Taming DINO for Subject-Driven Video Editing
Yi Huang (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Shifeng Chen (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
Image TranslationGenerationTransformerDiffusion modelImageVideoText
🎯 What it does: The DIVE framework is proposed, utilizing the semantic features of DINOv2 to achieve subject-driven video editing: accurately replacing the target subject (specified by reference images or text prompts) while retaining motion trajectories in the source video, maintaining consistency with the background and overall video.
Diversity-Enhanced Distribution Alignment for Dataset Distillation
Hongcheng Li (Institute of Information Engineering, Chinese Academy of Sciences), Weiping Wang (Institute of Information Engineering, Chinese Academy of Sciences)
Knowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Diversity Enhanced Distribution Alignment (DEDA) method, which matches the Gaussian distributions (mean and covariance) of each class of the original data in the feature space of a pre-trained network. It adds variance maximization and non-diagonal covariance minimization regularization to the final layer features, thereby enhancing the diversity of distilled data and alleviating the gradient starvation problem.
Divide-and-Conquer for Enhancing Unlabeled Learning, Stability, and Plasticity in Semi-supervised Continual Learning
Yue Duan (Nanjing University), Yinghuan Shi (Southeast University)
ClassificationKnowledge DistillationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A divide-and-conquer framework named USP is proposed to simultaneously enhance unlabeled learning, memory stability, and learning plasticity in semi-supervised continual learning (SSCL).
Diving into the Fusion of Monocular Priors for Generalized Stereo Matching
Chengtang Yao (Beijing Institute of Technology), Yunde Jia (Beijing Institute of Technology)
Depth EstimationDomain AdaptationAutonomous DrivingRecurrent Neural NetworkOptical FlowImage
🎯 What it does: Integrate the unbiased monocular depth prior provided by the Vision Foundation Model with stereo matching to enhance cross-domain generalization.
DLF: Extreme Image Compression with Dual-generative Latent Fusion
Naifu Xue (Communication University of China), Yan Lu (Microsoft Research Asia)
CompressionGenerative Adversarial NetworkImage
🎯 What it does: A dual generative latent fusion (DLF) ultra-low bitrate image compression framework is proposed, which utilizes a dual-branch approach to split the latent representation into semantic and detail components for compression.
DLFR-Gen: Diffusion-based Video Generation with Dynamic Latent Frame Rate
Zhihang Yuan (Tsinghua University Infinigence AI), Yu Wang (Tsinghua University)
GenerationCompressionTransformerDiffusion modelVideo
🎯 What it does: A non-trained dynamic latent frame rate compression module is introduced to the Diffusion Transformer (DiT) video generation model, which adaptively merges or retains frames in the latent space based on the motion frequency of the video content, significantly reducing the computational load during inference.
DM-EFS: Dynamically Multiplexed Expanded Features Set Form for Robust and Efficient Small Object Detection
Aashish Sharma (KLASS Engineering and Solutions)
Object DetectionImage
🎯 What it does: The DM-EFS method is proposed, which achieves efficient detection of small targets by dynamically multiplexing the extended feature set.
DMesh++: An Efficient Differentiable Mesh for Complex Shapes
Sanghyun Son (University of Maryland), Yi Zhou (University of Maryland)
GenerationComputational EfficiencyPoint CloudMesh
🎯 What it does: This paper presents DMesh++, a differentiable mesh representation that utilizes continuous point features, and implements an algorithm for efficiently reconstructing complex 2D/3D shapes from point clouds and multi-view images.
DMQ: Dissecting Outliers of Diffusion Models for Post-Training Quantization
Dongyeun Lee (KAIST), Junmo Kim (KAIST)
GenerationData SynthesisCompressionDiffusion modelImage
🎯 What it does: A post-training quantization framework DMQ is proposed, which utilizes learned equivalent scaling and per-channel binary scaling to address the outlier problem in the quantization of diffusion models.
DNF-Intrinsic: Deterministic Noise-Free Diffusion for Indoor Inverse Rendering
Rongjia Zheng (Sun Yat-Sen University), Wei-Shi Zheng (Sun Yat-Sen University)
RestorationGenerationDepth EstimationTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderImage
🎯 What it does: Using a denoising diffusion model to predict five intrinsic properties of objects (albedo, metallic, roughness, normal, depth) from a single indoor RGB image.
Do It Yourself: Learning Semantic Correspondence from Pseudo-Labels
Olaf Dünkel, Adam Kortylewski (University of Freiburg)
Object DetectionRepresentation LearningDiffusion modelContrastive LearningImage
🎯 What it does: Improved semantic correspondence features based on the foundational model through self-supervised pseudo-label learning.
DocThinker: Explainable Multimodal Large Language Models with Rule-based Reinforcement Learning for Document Understanding
Wenwen Yu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodalityChain-of-Thought
🎯 What it does: A rule-based reinforcement learning framework called DocThinker is proposed, which can dynamically improve the reasoning process of multimodal large language models (MLLMs) during inference and output interpretable intermediate steps (inference trajectories, restated questions, areas of focus, final answers).
Does Your Vision-Language Model Get Lost in the Long Video Sampling Dilemma?
Tianyuan Qu (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
Computational EfficiencyTransformerVision Language ModelVideoTextBenchmark
🎯 What it does: A long video high necessary sampling density task benchmark LSDBench is proposed, along with an inference-driven hierarchical sampling framework RHS and a semantic-guided frame selector SGFS, to improve the sampling efficiency and accuracy of long video visual language models.
DOGR: Towards Versatile Visual Document Grounding and Referring
Yinan Zhou (Xi'an Jiaotong University), Li Zhu (Xi'an Jiaotong University)
RecognitionObject DetectionRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: The DOGR-Engine is proposed to generate high-quality document parsing and instruction tuning data, based on which DOGR-Bench and the DOGR model are constructed, significantly enhancing document localization, citation, and dialogue capabilities.
DOLLAR: Few-Step Video Generation via Distillation and Latent Reward Optimization
Zihan Ding (Princeton University), Yuchen Liu (Adobe Research)
GenerationKnowledge DistillationDiffusion modelScore-based ModelVideo
🎯 What it does: This paper proposes the DOLLAR method, which combines three techniques: variational score distillation, consistency distillation, and latent reward optimization, to achieve high-quality and diverse video generation with only 4 steps of inference.
Domain Generalizable Portrait Style Transfer
Xinbo Wang (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Image TranslationDomain AdaptationDiffusion modelImage
🎯 What it does: A domain-generalizable portrait style transfer framework is proposed, utilizing semantic correspondence and a dual conditional diffusion model to achieve high-quality, semantically aligned style transfer.
Domain-aware Category-level Geometry Learning Segmentation for 3D Point Clouds
Pei He (Xidian University), Wenping Ma (Xidian University)
SegmentationDomain AdaptationPoint Cloud
🎯 What it does: A category-level geometric learning framework is proposed to enhance the domain generalization ability of point cloud semantic segmentation through category geometric embedding and geometric consistency learning.
DONUT: A Decoder-Only Model for Trajectory Prediction
Markus Knoche (RWTH Aachen University), Bastian Leibe (RWTH Aachen University)
Autonomous DrivingTransformerTime SeriesBenchmark
🎯 What it does: A trajectory prediction model called DONUT is proposed, which is based on a single decoder module and can uniformly handle historical and future trajectories to achieve autoregressive prediction.
Doodle Your Keypoints: Sketch-Based Few-Shot Keypoint Detection
Subhajit Maity (University of Central Florida), Yi-Zhe Song (University of Surrey)
Pose EstimationDomain AdaptationConvolutional Neural NetworkContrastive LearningImageMultimodality
🎯 What it does: A source-free few-shot keypoint detection framework based on hand-drawn sketches is proposed, capable of locating new keypoints in photos with only a small number of annotated sketches provided.
DoppDrive: Doppler-Driven Temporal Aggregation for Improved Radar Object Detection
Yuval Haitman (General Motors), Oded Bialer (General Motors)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: A time aggregation method utilizing Doppler information, DoppDrive, has been developed to reduce scattering and significantly improve radar point cloud density by dynamically moving points radially and limiting the aggregation duration specific to each point.
Doppler-Aware LiDAR-RADAR Fusion for Weather-Robust 3D Detection
Yujeong Chae (Korea Advanced Institute of Science and Technology), Kuk-Jin Yoon (Korea Advanced Institute of Science and Technology)
Object DetectionAutonomous DrivingMultimodalityPoint Cloud
🎯 What it does: This paper proposes a 3D object detection framework DLRFusion based on 4D RADAR and LiDAR, capable of achieving robust detection under various weather conditions.
DPoser-X: Diffusion Model as Robust 3D Whole-body Human Pose Prior
Junzhe Lu (Tsinghua University), Ziwei Liu (Nanyang Technological University)
RestorationGenerationPose EstimationDiffusion modelScore-based ModelMeshStochastic Differential Equation
🎯 What it does: We propose DPoser-X, an unconditional full-body human pose prior based on diffusion models, which is unified for various inverse problems such as pose generation, human mesh recovery, and pose completion.
DRaM-LHM: A Quaternion Framework for Iterative Camera Pose Estimation
Chen Lin (Flatiron Institute), Sonya M. Hanson (Flatiron Institute)
Pose EstimationSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes a closed-form orthogonal projection initialization based on quaternion adjoint matrices (DRaM) and combines it with LHM orthogonal iteration to construct a camera pose estimation framework with low iteration counts and robustness to noise;
Draw Your Mind: Personalized Generation via Condition-Level Modeling in Text-to-Image Diffusion Models
Hyungjin Kim (Inha University), Young-Duk Seo (Inha University)
GenerationRecommendation SystemTransformerDiffusion modelImageText
🎯 What it does: The DrUM method is proposed, utilizing core sampling, Transformer adapter (PeCA), and hierarchical softmax guidance to perform conditional-level modeling in the latent space, achieving personalized text-to-image generation without the need for fine-tuning.
Drawing Developmental Trajectory from Cortical Surface Reconstruction
Wenxuan Wu (South China University of Technology), Xin Zhang (South China University of Technology)
Knowledge DistillationImageBiomedical DataMagnetic Resonance ImagingOrdinary Differential Equation
🎯 What it does: A dual-task framework is proposed, utilizing age-conditioned ODE and self-supervised learning to directly reconstruct the cortical surface from single time-point MRI and generate developmental trajectories.
Dream-to-Recon: Monocular 3D Reconstruction with Diffusion-Depth Distillation from Single Images
Philipp Wulff (Technical University of Munich), Daniel Cremers (Technical University of Munich)
Depth EstimationAutonomous DrivingKnowledge DistillationDiffusion modelImage
🎯 What it does: A method for dense 3D reconstruction using only a single image is proposed, which completes view completion through a diffusion model and generates high-quality scene geometry, which is then distilled into a feedforward model;
DreamActor-M1: Holistic, Expressive and Robust Human Image Animation with Hybrid Guidance
Yuxuan Luo (Bytedance), Tianshu Hu (Bytedance)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: A novel human image animation framework called DreamActor-M1 based on Diffusion Transformer (DiT) is proposed, capable of synchronously generating fine-grained facial expressions and full-body movements at multiple scales (portrait, half-body, full-body), and supports long-term temporal consistency.
DreamCube: RGB-D Panorama Generation via Multi-plane Synchronization
Yukun Huang (University of Hong Kong), Xihui Liu (University of Hong Kong)
GenerationData SynthesisDepth EstimationDiffusion modelImage
🎯 What it does: This paper presents DreamCube, a multi-plane synchronized RGB-D cube map diffusion model that can generate panoramic RGB-D (cube) images from single-view RGB-D images and further create 3D scenes.
DreamDance: Animating Human Images by Enriching 3D Geometry Cues from 2D Poses
Yatian Pang (Peking University), Li Yuan (Peking University)
GenerationData SynthesisPose EstimationDiffusion modelVideoBenchmark
🎯 What it does: This paper proposes DreamDance, which utilizes a two-stage diffusion model to create high-quality animations of human figures based solely on skeletal pose sequences.
DreamFuse: Adaptive Image Fusion with Diffusion Transformer
Junjia Huang (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
Image HarmonizationGenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: An adaptive image fusion framework called DreamFuse based on Diffusion Transformer (DiT) is proposed, and a dataset of 80k high-quality fused images is constructed through an iterative human-machine interactive data generation pipeline.
DreamLayer: Simultaneous Multi-Layer Generation via Diffusion Model
Junjia Huang (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: The DreamLayer framework is proposed to achieve text-driven multi-layer image generation and decomposition, supporting hierarchical editing and untrained image-to-layer splitting.
DreamRelation: Relation-Centric Video Customization
Yujie Wei (Fudan University), Hongming Shan (Fudan University)
RecognitionGenerationTransformerDiffusion modelContrastive LearningVideo
🎯 What it does: Proposes the DreamRelation method to address the video generation problem of customizing relationships between two subjects from a small number of example videos.
DreamRenderer: Taming Multi-Instance Attribute Control in Large-Scale Text-to-Image Models
Dewei Zhou (Zhejiang University), Yi Yang (Harvard University)
GenerationData SynthesisTransformerVision Language ModelImage
🎯 What it does: This paper presents DreamRenderer, a training-free, pluggable controller for achieving fine-grained attribute control in multi-instance image generation.
DriveArena: A Closed-loop Generative Simulation Platform for Autonomous Driving
Xuemeng Yang (Shanghai Artificial Intelligence Laboratory), Yu Qiao (Shanghai Artificial Intelligence Laboratory)
GenerationAutonomous DrivingDiffusion modelImageText
🎯 What it does: This paper presents DRIVEARENA—a closed-loop high-fidelity simulation platform for vehicular networks, integrating Traffic Manager (traffic flow simulation) and World Dreamer (diffusion-based image generation), achieving interactive and controllable multi-view, sustainably generated driving scenes on any OSM map.
DriveX: Omni Scene Modeling for Learning Generalizable World Knowledge in Autonomous Driving
Chen Shi (Chinese University of Hong Kong), Li Jiang (Chinese University of Hong Kong)
Autonomous DrivingTransformerWorld ModelMultimodalityPoint Cloud
🎯 What it does: This paper proposes a self-supervised world model called DriveX, which utilizes Omni Scene Modeling to learn generalizable scene dynamics and overall representations in the BEV latent space, and applies the prediction results uniformly to various autonomous driving tasks through Future Spatial Attention.
Driving View Synthesis on Free-form Trajectories with Generative Prior
Zeyu Yang (Fudan University), Li Zhang (Fudan University)
GenerationData SynthesisAutonomous DrivingDiffusion modelGaussian SplattingVideoPoint Cloud
🎯 What it does: Combining video diffusion priors with 3D Gaussian Splatting in single-trajectory videos to reconstruct driving scenes and achieve high-quality view synthesis under free trajectories.
DrivingGPT: Unifying Driving World Modeling and Planning with Multi-modal Autoregressive Transformers
Yuntao Chen (Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
GenerationAutonomous DrivingTransformerLarge Language ModelDiffusion modelImageVideoMultimodality
🎯 What it does: The DrivingGPT model is proposed, unifying driving world modeling and trajectory planning into a multimodal autoregressive sequence prediction task, generating future videos and actions using images and discrete action vocabulary.
DropletVideo: A Dataset and Approach to Explore Integral Spatio-Temporal Consistent Video Generation
Runze Zhang (IEIT SYSTEMS Co., Ltd.), Baoyu Fan (IEIT SYSTEMS Co., Ltd.)
GenerationData SynthesisTransformerOptical FlowVideoText
🎯 What it does: This paper proposes the Integral Spatio-Temporal Consistency video generation framework and builds a large-scale dataset and model based on it.
DSO: Aligning 3D Generators with Simulation Feedback for Physical Soundness
Ruining Li (Visual Geometry Group University of Oxford), Andrea Vedaldi (Visual Geometry Group University of Oxford)
GenerationOptimizationReinforcement LearningDiffusion modelMeshPhysics Related
🎯 What it does: Train a 3D generator to directly output 3D objects that can self-support under the influence of gravity.
Dual Domain Control via Active Learning for Remote Sensing Domain Incremental Object Detection
Jiachen Sun (Xidian University), Nannan Wang (Xidian University)
Object DetectionDomain AdaptationKnowledge DistillationTransformerImage
🎯 What it does: This paper proposes a framework for incremental object detection in the remote sensing domain (Active-DDC), which achieves rapid adaptation to new domains while retaining knowledge of old domains by controlling data distribution shifts and feature transfer through active learning.
Dual Reciprocal Learning of Language-based Human Motion Understanding and Generation
Chen Liang (Zhejiang University), Yi Yang (Zhejiang University)
GenerationData SynthesisPose EstimationTransformerReinforcement LearningTextMultimodality
🎯 What it does: Proposes a Dual Reciprocal Learning framework that collaboratively trains text-to-motion (T2M) and motion-to-text (M2T) tasks in a closed-loop manner.
Dual Recursive Feedback on Generation and Appearance Latents for Pose-Robust Text-to-Image Diffusion
Jiwon Kim (Korea University), Kyong Hwan Jin (Korea University)
GenerationData SynthesisPose EstimationDiffusion modelScore-based ModelImagePoint CloudMesh
🎯 What it does: A training-free dual recursive feedback framework is proposed to enhance the appearance and structural control effects of text-to-image diffusion models with robust pose.
Dual-Expert Consistency Model for Efficient and High-Quality Video Generation
Zhengyao Lv (Nanjing University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationMixture of ExpertsDiffusion modelGenerative Adversarial NetworkVideo
🎯 What it does: A parameter-efficient dual-expert consistency model (DCM) is proposed, which first trains a semantic expert for layout and motion modeling, then freezes it and adds LoRA layers for fine-tuning the detail expert, achieving high-quality video generation and efficient sampling.
Dual-level Prototype Learning for Composite Degraded Image Restoration
Zhongze Wang (East China University of Science and Technology), Kaijie Zhao (East China University of Science and Technology)
RestorationTransformerImage
🎯 What it does: Proposes a Dual Prototype Learning (DPL) framework for dynamic scene embedding and guidance recovery of composite degraded images.
Dual-Process Image Generation
Grace Luo, Trevor Darrell
GenerationKnowledge DistillationTransformerVision Language ModelRectified FlowImageMultimodality
🎯 What it does: A dual-process distillation framework is proposed, combining a visual language model (VLM) with a high-quality image generator, achieving gradient backpropagation through VQA loss;
Dual-Rate Dynamic Teacher for Source-Free Domain Adaptive Object Detection
Qi He (Southwest Jiaotong University), Shuai Li (Meituan Inc.)
Object DetectionDomain AdaptationImage
🎯 What it does: This paper studies source-free domain adaptive object detection and proposes a Dual-Speed Dynamic Teacher (DDT) framework, which utilizes asynchronous EMA to achieve grouped updates of teacher parameters, enhancing the quality of pseudo-labels and model adaptability.
Dual-S3D: Hierarchical Dual-Path Selective SSM-CNN for High-Fidelity Implicit Reconstruction
Luoxi Zhang (University of Tsukuba), Itaru Kitahara (University of Tsukuba)
RestorationGenerationConvolutional Neural NetworkTransformerNeural Radiance FieldPoint Cloud
🎯 What it does: Proposes the Dual-S3D framework, achieving high-fidelity implicit 3D reconstruction from a single view through dual-path feature extraction and staged training.
Dual-Temporal Exemplar Representation Network for Video Semantic Segmentation
Xiaolong Xu (Sichuan University), Hao Song (Sichuan University)
SegmentationConvolutional Neural NetworkVideo
🎯 What it does: This paper proposes a Dual Temporal Exemplar Representation Network (DTERN) to achieve video semantic segmentation by learning local and global temporal exemplars.
DualReal: Adaptive Joint Training for Lossless Identity-Motion Fusion in Video Customization
Wenchuan Wang (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: A dual-dimensional adaptive joint training framework called DualReal is proposed for video customization, which can generate realistic motion while maintaining high identity consistency, addressing the identity-motion conflict problem of traditional unidimensional customization methods.
DuCos: Duality Constrained Depth Super-Resolution via Foundation Model
Zhiqiang Yan (National University of Singapore), Gim Hee Lee (National University of Singapore)
RestorationDepth EstimationSuper ResolutionPrompt EngineeringImage
🎯 What it does: A novel deep super-resolution framework called DuCos is proposed, based on Lagrangian duality and deep foundation model prompts.
DuET: Dual Incremental Object Detection via Exemplar-Free Task Arithmetic
Munish Monga (Sony Research), Biplab Banerjee (Indian Institute of Technology Bombay)
Object DetectionDomain AdaptationKnowledge DistillationImage
🎯 What it does: The DuET framework is designed to achieve zero-shot dual incremental object detection, allowing for simultaneous learning of new categories and adaptation to domain changes.
DuoCLR: Dual-Surrogate Contrastive Learning for Skeleton-based Human Action Segmentation
Haitao Tian (University of Ottawa)
RecognitionSegmentationGraph Neural NetworkContrastive LearningVideoSequential
🎯 What it does: By using contrastive learning pre-trained models on trimmed skeleton sequences, the method for Skeleton-based human action segmentation has been improved.
DuoLoRA : Cycle-consistent and Rank-disentangled Content-Style Personalization
Aniket Roy (Johns Hopkins University), Fatih Porikli (Qualcomm AI Research)
Image TranslationGenerationTransformerDiffusion modelImage
🎯 What it does: This paper proposes the DuoLoRA framework, which achieves efficient fusion and personalization of content and style through LoRA adapter rank dimension masking, hierarchical priors, and cyclic consistency loss.
DWIM: Towards Tool-aware Visual Reasoning via Discrepancy-aware Workflow Generation & Instruct-Masking Tuning
Fucai Ke, Manmohan Chandraker
RecognitionObject DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision-Language-Action ModelImageTextMultimodality
🎯 What it does: This paper studies a tool-aware visual reasoning framework named DWIM, which enhances the tool usage and reasoning effectiveness of large language models in visual reasoning tasks by combining an inconsistency-aware workflow generation and instruction-masking fine-tuning.
DyGS-SLAM: Real-Time Accurate Localization and Gaussian Reconstruction for Dynamic Scenes
Xinggang Hu (Dalian University of Technology), Xiangyang Ji (Tsinghua University)
Object DetectionObject TrackingPose EstimationGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: A real-time high-precision localization and high-quality dense reconstruction system named DyGS-SLAM for dynamic scenes is proposed.
Dynamic Dictionary Learning for Remote Sensing Image Segmentation
Xuechao Zou (Beijing Jiaotong University), Congyan Lang (Tsinghua University)
SegmentationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A dynamic dictionary learning framework is proposed, explicitly constructing and iteratively optimizing class ID embeddings to enhance fine-grained recognition in remote sensing image semantic segmentation.
Dynamic Group Detection using VLM-augmented Temporal Groupness Graph
Kaname Yokoyama (Toyota Technological Institute), Norimichi Ukita (Toyota Technological Institute)
Object DetectionObject TrackingTransformerSupervised Fine-TuningVision Language ModelVideo
🎯 What it does: Implement dynamic crowd detection in videos by utilizing a vision-language model (CLIP) to extract groupness probability with spatial and temporal context, and construct a temporal groupness graph for clustering.
Dynamic Multi-Layer Null Space Projection for Vision-Language Continual Learning
Borui Kang (Nanjing University), Wenbin Li (Nanjing University)
TransformerVision Language ModelImageMultimodality
🎯 What it does: A dynamic multi-layer null space projection (DMNSP) strategy is proposed and implemented, specifically addressing the catastrophic forgetting problem in visual-language models during continual learning, utilizing a parameter-efficient adapter architecture and applying the projection only to the visual branch.
Dynamic Multimodal Prototype Learning in Vision-Language Models
Xingyu Zhu (Nanyang Technological University), Hanwang Zhang
ClassificationDomain AdaptationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A training-free multimodal prototype learning framework called ProtoMM is proposed, which dynamically updates visual particles through optimal transport during testing, treating text descriptions and visual features as discrete distributions for joint prototype learning, thereby enhancing the zero-shot classification performance of VLM.
Dynamic Point Maps: A Versatile Representation for Dynamic 3D Reconstruction
Edgar Sucar (Visual Geometry Group University of Oxford), Andrea Vedaldi (Visual Geometry Group University of Oxford)
Object TrackingDepth EstimationAutonomous DrivingTransformerVideoPoint Cloud
🎯 What it does: Dynamic Point Maps (DPM) are proposed as a spatiotemporal invariant representation for unifying the prediction of camera parameters, 3D reconstruction, and dynamic scene tasks.
Dynamic Reconstruction of Hand-Object Interaction with Distributed Force-aware Contact Representation
Zhenjun Yu, Cewu Lu
Pose EstimationOptimizationRobotic IntelligenceTransformerPoint Cloud
🎯 What it does: In the visual-haptic interaction between the hand and objects, a DF-Field touch representation based on distributed force perception is proposed, and the ViTaM-D framework for visual dynamic tracking and force perception optimization is implemented.
Dynamic Typography: Bringing Text to Life via Video Diffusion Prior
Zichen Liu (Hong Kong University of Science and Technology), Huamin Qu (Hong Kong University of Science and Technology)
GenerationData SynthesisDiffusion modelScore-based ModelVideoText
🎯 What it does: This paper proposes an automatic typography animation framework called Dynamic Typography based on video diffusion priors, which generates semantic-driven deformation and motion of individual letters according to text prompts.
Dynamic-DINO: Fine-Grained Mixture of Experts Tuning for Real-time Open-Vocabulary Object Detection
Yehao Lu (Zhejiang University), Xi Li (Zhejiang University)
Object DetectionTransformerMixture of ExpertsImage
🎯 What it does: Dynamic-DINO is proposed, a framework that integrates dynamic reasoning of Mixture-of-Experts (MoE) into real-time open-vocabulary object detection, capable of expanding the subnet search space while maintaining the parameter count of a single FFN;
Dynamic-VLM: Simple Dynamic Visual Token Compression for VideoLLM
Han Wang, Can Huang (Bytedance Inc)
Data SynthesisCompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoText
🎯 What it does: A dynamic visual token compression-based large-scale video visual language model (Dynamic-VLM) is proposed, and a large-scale synthetic video-text dataset is constructed.
DynamicFace: High-Quality and Consistent Face Swapping for Image and Video using Composable 3D Facial Priors
Runqi Wang (Xiaohongshu Inc.), Yao Hu (Xiaohongshu Inc.)
Image TranslationGenerationRetrievalTransformerDiffusion modelImageVideo
🎯 What it does: This work proposes DynamicFace, a facial swapping framework based on diffusion models, capable of high-quality and temporally consistent transfer of source facial identities to target faces in images and videos, while preserving non-identity attributes such as target expressions, poses, lighting, and backgrounds.
DynamicID: Zero-Shot Multi-ID Image Personalization with Flexible Facial Editability
Xirui Hu (Xi'an Jiaotong University), Haishun Nan (Western Movie Group)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A parameter-free framework called DynamicID is designed to achieve high-fidelity personalized image generation in both single-identity and multi-identity scenarios, supporting flexible facial editing.
DynFaceRestore: Balancing Fidelity and Quality in Diffusion-Guided Blind Face Restoration with Dynamic Blur-Level Mapping and Guidance
Huu-Phu Do (National Yang Ming Chiao Tung University), Ching-Chun Huang (National Yang Ming Chiao Tung University)
RestorationDiffusion modelImage
🎯 What it does: A dynamic blur mapping and guidance method for blind face restoration, DynFaceRestore, is proposed, which can map unknown degraded images to multi-scale Gaussian blurred images and achieve detail reconstruction through a pre-trained diffusion model.
DynImg: Key Frames with Visual Prompts are Good Representation for Multi-Modal Video Understanding
Xiaoyi Bao (Institute of Automation, Chinese Academy of Sciences), Xingang Wang (Alibaba Group)
Representation LearningTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodality
🎯 What it does: A dynamic image (DynImg) video representation method based on key frames and non-key frame visual cues is proposed, utilizing non-key frames as temporal cues overlaid on key frames to enhance the model's focus on rapidly moving areas.
DyWA: Dynamics-adaptive World Action Model for Generalizable Non-prehensile Manipulation
Jiangran Lyu (Peking University), He Wang (Peking University)
Knowledge DistillationRobotic IntelligenceReinforcement LearningWorld ModelPoint Cloud
🎯 What it does: The DyWA framework is proposed, achieving general non-grasp manipulation through a dynamically adaptive world action model.
E-NeMF: Event-based Neural Motion Field for Novel Space-time View Synthesis of Dynamic Scenes
Yan Liu (Zhejiang University), Gang Pan (Zhejiang University)
GenerationData SynthesisNeural Radiance FieldOptical FlowVideo
🎯 What it does: A new spatial-temporal perspective synthesis method is proposed using event cameras to assist monocular video.
E-SAM: Training-Free Segment Every Entity Model
Weiming Zhang (Hong Kong University of Science and Technology), Lin Wang (Nanyang Technological University)
Object DetectionSegmentationTransformerImage
🎯 What it does: A training-agnostic E-SAM framework is proposed, which enhances the automatic mask generation of SAM for entity segmentation through a three-stage post-processing (MMG, EMR, USR).
EA-KD: Entropy-based Adaptive Knowledge Distillation
Chi-Ping Su (National Yang Ming Chiao Tung University), Shin-Jye Lee (National Yang Ming Chiao Tung University)
Knowledge DistillationImage
🎯 What it does: This paper proposes an entropy-based adaptive knowledge distillation method EA-KD, which allows the student model to focus more on high-information samples during the distillation process.
EA-Vit: Efficient Adaptation for Elastic Vision Transformer
Chen Zhu (National University of Singapore), Dawei Yang (Houmo AI)
ClassificationSegmentationTransformerSupervised Fine-TuningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: EA-ViT is proposed, an elastic Vision Transformer framework that can generate multi-size sub-models during the adaptation phase and dynamically select them based on device resource budgets through a lightweight router.
EAMamba: Efficient All-Around Vision State Space Model for Image Restoration
Yu-Cheng Lin (National Tsing Hua University), Chun-Yi Lee (National Taiwan University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A novel visual Mamba model for image restoration, EAMamba, is proposed, which combines a multi-head selective scanning module (MHSSM) and an all-around scanning strategy to achieve efficient long-range dependency modeling and local pixel memory retention.
Early Timestep Zero-Shot Candidate Selection for Instruction-Guided Image Editing
Joowon Kim (Korea Advanced Institute of Science and Technology), Eunho Yang (Korea Advanced Institute of Science and Technology)
Image TranslationGenerationComputational EfficiencyDiffusion modelRectified FlowImageBenchmark
🎯 What it does: A zero-shot seed selection method called ELECT is proposed, which can quickly filter out the best random seeds that maintain background consistency in instruction-driven image editing.
Easi3R: Estimating Disentangled Motion from DUSt3R Without Training
Xingyu Chen (Zhejiang University), Anpei Chen (University of Tubingen)
SegmentationPose EstimationTransformerOptical FlowVideoPoint Cloud
🎯 What it does: This paper proposes an Easi3R method that does not require retraining, which completes dynamic object segmentation, camera motion estimation, and 4D point cloud reconstruction by re-weighting the attention layers of DUSt3R during the inference phase.
Easy3D: A Simple Yet Effective Method for 3D Interactive Segmentation
Andrea Simonelli (Meta Reality Labs), Peter Kontschieder (Meta Reality Labs)
SegmentationDomain AdaptationConvolutional Neural NetworkTransformerGaussian SplattingPoint Cloud
🎯 What it does: A simple and efficient interactive 3D instance segmentation method called Easy3D is proposed, which allows obtaining object masks with just a few 3D clicks and can be iteratively refined.
EasyControl: Adding Efficient and Flexible Control for Diffusion Transformer
Yuxuan Zhang (Chinese University of Hong Kong), Jiaming Liu (Alibaba Group)
Image TranslationGenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: We propose EasyControl, a lightweight and pluggable control framework that enables single/multi-condition control, zero-shot multi-condition combination, and efficient inference on the Diffusion Transformer (DiT).
EC-Flow: Enabling Versatile Robotic Manipulation from Action-Unlabeled Videos via Embodiment-Centric Flow
Yixiang Chen (Institute of Automation Chinese Academy of Sciences), Liang Wang (Institute of Automation Chinese Academy of Sciences)
Robotic IntelligenceConvolutional Neural NetworkDiffusion modelVideo
🎯 What it does: The EC-Flow framework is proposed, which utilizes action-annotated videos to learn robot manipulation, directly predicting the Embodiment-Centric Flow (EC-Flow) and calculating executable actions through URDF;
EDFFDNet: Towards Accurate and Efficient Unsupervised Multi-Grid Image Registration
Haokai Zhu (Zhejiang University), Hui-Liang Shen (Zhejiang University)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an unsupervised image registration framework EDFFDNet, which utilizes an exponential decay free-form deformation model, an adaptive sparse motion aggregator, and a progressive correlation strategy to achieve multi-scale local deformations.
Edicho: Consistent Image Editing in the Wild
Qingyan Bai (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: This paper proposes a training-free, plug-and-play image editing method called Edicho, which achieves consistent multi-image editing during the denoising process of diffusion models by utilizing explicit correspondence.
EDiT: Efficient Diffusion Transformers with Linear Compressed Attention
Philipp Becker, Sourav Bhattacharya
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelImageMultimodality
🎯 What it does: This paper proposes the Efficient Diffusion Transformer (EDiT) and the Multimodal Efficient Diffusion Transformer (MM-EDiT), which enhance the efficiency and high-resolution generation capability of Diffusion Transformers through linear compressed attention.
Edit360: 2D Image Edits to 3D Assets from Any Angle
Junchao Huang (Chinese University of Hong Kong), Li Jiang (Chinese University of Hong Kong)
Image TranslationGenerationDiffusion modelImageVideoMesh
🎯 What it does: Achieved 360° 3D asset editing without the need for fine-tuning, supporting editing of 2D images from any perspective and consistently propagating it to a complete 3D model.
EditCLIP: Representation Learning for Image Editing
Qian Wang (King Abdullah University of Science and Technology), Peter Wonka (King Abdullah University of Science and Technology)
Image TranslationRepresentation LearningDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes EditCLIP, a representation model for learning image editing transformations in the CLIP space;
EDM: Efficient Deep Feature Matching
Xi Li (Realsee), Cihui Pan (Realsee)
Pose EstimationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes an Efficient Deep Feature Matching Network (EDM), which extracts multi-scale features through deep CNNs and introduces global context in the deeper layers for fast matching.
EEdit : Rethinking the Spatial and Temporal Redundancy for Efficient Image Editing
Zexuan Yan (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Computational EfficiencyTransformerDiffusion modelImage
🎯 What it does: An efficient image editing framework named EEdit is proposed, focusing on reducing computational redundancy in image editing based on diffusion models.