ICCV 2025 Papers — Page 6
IEEE/CVF International Conference on Computer Vision · 2701 papers
DAViD: Modeling Dynamic Affordance of 3D Objects Using Pre-trained Video Diffusion Models
Hyeonwoo Kim, Hanbyul Joo
GenerationData SynthesisRobotic IntelligenceDiffusion modelVideoMesh
🎯 What it does: Using a pre-trained video diffusion model to synthesize multi-view HOI images/videos and elevate them to 3D, then fine-tuning the human motion diffusion model and conditional diffusion model with LoRA, training a generative dynamic affordance model DAViD to achieve dynamic interaction generation for any 3D object.
DC-AE 1.5: Accelerating Diffusion Model Convergence with Structured Latent Space
Junyu Chen (NVIDIA), Han Cai (NVIDIA)
GenerationCompressionDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes DC-AE 1.5, a deep compression autoencoder with a structured latent space and enhanced diffusion training, aimed at accelerating the convergence of latent diffusion models with a large number of channels and improving the quality of high-resolution image generation.
DC-AR: Efficient Masked Autoregressive Image Generation with Deep Compression Hybrid Tokenizer
Yecheng Wu (Massachusetts Institute of Technology), Song Han (NVIDIA)
GenerationCompressionTransformerDiffusion modelImage
🎯 What it does: This paper proposes DC-AR, an efficient masked autoregressive text-to-image generation framework;
DC-ControlNet: Decoupling Inter- and Intra-Element Conditions in Image Generation with Diffusion Models
Hongji Yang (University of Macau), Jianbing Shen (University of Macau)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Designed the DC-ControlNet framework to separately control the content and layout of various elements in an image, achieving more refined multi-condition generation.
DC-TTA: Divide-and-Conquer Framework for Test-Time Adaptation of Interactive Segmentation
Jihun Kim (KAIST), Kuk-Jin Yoon (KAIST)
SegmentationDomain AdaptationImage
🎯 What it does: A divide-and-conquer testing-time adaptation framework DC-TTA is proposed to improve the performance of the interactive segmentation model SAM.
DCHM: Depth-Consistent Human Modeling for Multiview Detection
Jiahao Ma (Australian National University), Chuong Nguyen (CSIRO Data61)
Object DetectionDepth EstimationGaussian SplattingPoint Cloud
🎯 What it does: A depth-consistent human modeling method based on monocular depth estimation and Gaussian Splatting is proposed for multi-view pedestrian detection, achieving the integration of sparse view information into a global 3D point cloud for pedestrian localization.
DCT-Shield: A Robust Frequency Domain Defense against Malicious Image Editing
Aniruddha Bala (Samsung Research and Development Institute), Siddharth Roheda (Samsung Research and Development Institute)
Image TranslationCompressionAdversarial AttackDiffusion modelAuto EncoderImage
🎯 What it does: A method for imperceptible adversarial perturbation of images in the DCT domain is designed, utilizing the JPEG quantization process to protect against malicious editing of diffusion models.
DDB: Diffusion Driven Balancing to Address Spurious Correlations
Aryan Yazdan Parast (University of Melbourne), Naveed Akhtar (University of Melbourne)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Automatically generate minority class samples using text inversion, language segmentation, and local inpainting of Stable Diffusion to rebalance the training set and reduce the model's reliance on spurious correlations.
Debiased Curriculum Adaptation for Safe Transfer Learning in Chest X-ray Classification
Mingyang Liu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
ClassificationDomain AdaptationImageBiomedical Data
🎯 What it does: A Debiased Curriculum Adaptation framework is proposed to achieve safe unsupervised domain adaptation in chest X-ray image classification tasks.
Debiased Teacher for Day-to-Night Domain Adaptive Object Detection
Yiming Cui (Hangzhou Dianzi University), Chenggang Yan (Hangzhou Dianzi University)
Object DetectionDomain AdaptationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A Debiased Teacher framework is proposed, utilizing physical priors to generate night-style images, bidirectional domain mixed compensation features, and dynamic thresholds with joint distillation to correct pseudo-labels, thereby addressing distribution bias, training bias, and confirmation bias in day-night object detection.
Debiasing Trace Guidance: Top-down Trace Distillation and Bottom-up Velocity Alignment for Unsupervised Anomaly Detection
Xingjian Wang (Zhejiang University), Jiming Chen (Zhejiang University)
Anomaly DetectionDiffusion modelFlow-based ModelImage
🎯 What it does: The Debiasing Trace Guidance (DTG) framework is proposed, which improves the generative process of unsupervised multi-class anomaly detection through upsampling trace distillation and downsampling speed alignment.
DecAD: Decoupling Anomalies in Latent Space for Multi-Class Unsupervised Anomaly Detection
Xiaolei Wang (University of Liverpool), Jimin Xiao (XJTLU)
Anomaly DetectionTransformerAuto EncoderImage
🎯 What it does: A DecAD framework is proposed, which separates encoded features into normal and abnormal parts in the latent space using reversible neural networks, and only uses the normal part for reconstruction, achieving unsupervised multi-class anomaly detection.
Deciphering Cross-Modal Alignment in Large Vision-Language Models via Modality Integration Rate
Qidong Huang (Alibaba Cloud), Nenghai Yu (University of Science and Technology of China)
TransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This study investigates the evaluation of cross-modal alignment during the pre-training phase of large visual language models (LVLM).
Decoding Correlation-Induced Misalignment in the Stable Diffusion Workflow for Text-to-Image Generation
Yunze Tong (Zhejiang University), Kun Kuang (Zhejiang University)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: To address the 'object missing' problem in Stable Diffusion when handling related words in text, a de-correlation fine-tuning method is proposed on the self-attention layer of the text encoder to improve the alignment between text and images without introducing external knowledge.
Decouple and Track: Benchmarking and Improving Video Diffusion Transformers For Motion Transfer
Qingyu Shi (Peking University), Xiangtai Li (Nanyang Technological University)
Object TrackingGenerationData SynthesisTransformerDiffusion modelVideoTextBenchmark
🎯 What it does: A motion transfer framework DeT based on Video Diffusion Transformer (DiT) is proposed, achieving separation and transfer of motion and appearance by incorporating shared temporal convolution kernels and dense point tracking loss into DiT; simultaneously, a new evaluation benchmark MTBench is constructed to cover a wider range of motion difficulties.
Decouple to Reconstruct: High Quality UHD Restoration via Active Feature Disentanglement and Reversible Fusion
Yidi Liu (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
RestorationConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImage
🎯 What it does: The D²R‑UHDNet framework is proposed, which achieves high-quality UHD image restoration by decoupling and separately processing background-dominant features and degradation-dominant features.
Decoupled Diffusion Sparks Adaptive Scene Generation
Yunsong Zhou (Shanghai AI Lab), Hongyang Li
GenerationAutonomous DrivingTransformerDiffusion modelVideo
🎯 What it does: A separation diffusion model named Nexus is proposed for controllable and real-time traffic scene generation.
Decoupled Multi-Predictor Optimization for Inference-Efficient Model Tuning
Liwei Luo (Tianjin University), Qinghua Hu (Tianjin University)
OptimizationTransformerSupervised Fine-TuningImage
🎯 What it does: A decoupled multi-predictor optimization (DMPO) method based on early exit is proposed to achieve efficient fine-tuning of inference for large-scale pre-trained models.
Deep Adaptive Unfolded Network via Spatial Morphology Stripping and Spectral Filtration for Pan-sharpening
Hebaixu Wang (Wuhan University), Jiayi Ma (Wuhan University)
Image TranslationRestorationSegmentationImageMultimodalityBenchmark
🎯 What it does: A deep adaptive unfolding network (DAPNet) is proposed, which achieves multi-modal image fusion (full-resolution spectral image reconstruction) through spatial morphology stripping and spectral filtering.
Deep Incomplete Multi-view Clustering with Distribution Dual-Consistency Recovery Guidance
Jiaqi Jin (National University of Defense Technology), Kunlun He (Academy of Military Sciences)
OptimizationFlow-based ModelAuto EncoderImageTextMultimodalityAudio
🎯 What it does: This paper proposes a deep incomplete multi-view clustering method named BURG, which utilizes view-specific normalized flow models for distribution transfer recovery and guides the reconstruction of missing views through dual constraints of neighborhood consistency and prototype consistency, ultimately achieving clustering of incomplete multi-view data.
Deep Space Weather Model: Long-Range Solar Flare Prediction from Multi-Wavelength Images
Shunya Nagashima (Keio University), Komei Sugiura (Keio University)
ClassificationConvolutional Neural NetworkRecurrent Neural NetworkAuto EncoderImageTime SeriesBenchmark
🎯 What it does: Using multi-spectral solar images to predict the maximum flare category within 24 hours, and proposing an end-to-end Deep SWM model with the FlareBench dataset.
Deeply Supervised Flow-Based Generative Models
Inkyu Shin (ByteDance), Liang-Chieh Chen (ByteDance)
GenerationData SynthesisComputational EfficiencyTransformerFlow-based ModelImageTextOrdinary Differential Equation
🎯 What it does: A flow-based generative model called DeepFlow is proposed, which significantly improves training efficiency and generation quality by implementing deep supervision and internal velocity representation alignment between multiple layers of the Transformer.
DeepMesh: Auto-Regressive Artist-mesh Creation with Reinforcement Learning
Ruowen Zhao (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationData SynthesisOptimizationTransformerReinforcement LearningMesh
🎯 What it does: A framework named DeepMesh is proposed, which generates artistic style triangular meshes through an autoregressive method, addressing the challenges of existing methods regarding face count limitations and mesh incompleteness.
DeepShield: Fortifying Deepfake Video Detection with Local and Global Forgery Analysis
Yinqi Cai (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
ClassificationRecognitionAnomaly DetectionTransformerContrastive LearningVideo
🎯 What it does: A deepfake video detection framework named DeepShield is proposed, which enhances the model's robustness to unknown forgeries by utilizing a dual strategy of local patch supervision and global forgery diversification.
DeFSS: Image-to-Mask Denoising Learning for Few-shot Segmentation
Zishu Qin (Fudan University), Weifeng Ge (Fudan University)
SegmentationTransformerDiffusion modelImage
🎯 What it does: This paper addresses the few-shot semantic segmentation task by transforming the image-to-mask decoding process into a two-stage framework of image erosion and mask denoising, thereby reducing the over-reliance on support samples.
DeGauss: Dynamic-Static Decomposition with Gaussian Splatting for Distractor-free 3D Reconstruction
Rui Wang (ETH Zurich), Siyu Tang (ETH Zurich)
RestorationGenerationOptimizationGaussian SplattingVideo
🎯 What it does: A self-supervised framework called DeGauss is proposed, based on dynamic-static Gaussian splatting, for achieving interference-free 3D reconstruction in real-world dynamic videos.
Degradation-Modeled Multipath Diffusion for Tunable Metalens Photography
Jianing Zhang (Fudan University), Xiaoyun Yuan (Shanghai Jiao Tong University)
RestorationDiffusion modelImage
🎯 What it does: A framework for metal lens photography image restoration based on Spatially Variable Distortion-Aware Attention (SVDA) and multi-path diffusion is proposed to address the issues of spatially varying distortion in metal lens imaging and the lack of paired data.
Demeter: A Parametric Model of Crop Plant Morphology from the Real World
Tianhang Cheng (University of Illinois Urbana-Champaign), Shenlong Wang (University of Illinois Urbana-Champaign)
SegmentationGenerationOptimizationGaussian SplattingVideoPoint CloudAgriculture Related
🎯 What it does: A plant morphology parameterization model named Demeter is proposed, which enables data-driven low-dimensional modeling of plant topology, shape, joints (posture), and deformation, trained using real field 3D reconstruction data.
Democratizing High-Fidelity Co-Speech Gesture Video Generation
Xu Yang (South China University of Technology), Changxing Ding (South China University of Technology)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideoAudio
🎯 What it does: A lightweight audio-to-skeleton prediction framework is proposed, which combines existing human video generation models to achieve synchronized, realistic collaborative speaking video generation with audio, facial expressions, and body gestures.
Democratizing Text-to-Image Masked Generative Models with Compact Text-Aware One-Dimensional Tokens
Dongwon Kim (ByteDance), Liang-Chieh Chen (ByteDance)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImageText
🎯 What it does: This paper proposes the TA-TiTok one-dimensional text-aware tokenizer and the MaskGen text-to-image generation framework, which can achieve efficient and reproducible text generation tasks on public datasets.
Denoising Token Prediction in Masked Autoregressive Models
Ting Yao (HiDream.ai Inc), Tao Mei (HiDream.ai Inc)
GenerationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: A framework for denoising prediction in masked autoregressive models (De-MAR) is proposed, achieving higher quality and faster speed in image generation tasks.
Dense Policy: Bidirectional Autoregressive Learning of Actions
Yue Su (Shanghai Jiao Tong University), Lixin Yang (Shanghai Jiao Tong University)
GenerationRobotic IntelligenceTransformerReinforcement LearningVision-Language-Action ModelImageVideo
🎯 What it does: This paper proposes Dense Policy, a bidirectional autoregressive action generation model based on an Encoder-only Transformer, which can generate sparse keyframes from a single frame and recursively expand to a complete action sequence through bidirectional interpolation.
Dense2MoE: Restructuring Diffusion Transformer to MoE for Efficient Text-to-Image Generation
Youwei Zheng (Sun Yat-sen University), Xiaohua Xie (ByteDance Intelligent Creation)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationTransformerMixture of ExpertsDiffusion modelImage
🎯 What it does: Transforming the dense Diffusion Transformer (DiT) into a sparse Mixture of Experts (MoE) structure, and introducing Mixture of Blocks (MoB) for hierarchical sparsification, resulting in Dense2MoE.
DepR: Depth Guided Single-view Scene Reconstruction with Instance-level Diffusion
Qingcheng Zhao (ShanghaiTech University), Zhuowen Tu (University of California San Diego)
GenerationDepth EstimationOptimizationDiffusion modelAuto EncoderImage
🎯 What it does: The DepR framework is proposed, which achieves instance-level 3D scene reconstruction from a single RGB image using depth information, and assembles the reconstructed objects into a complete scene.
Depth Any Event Stream: Enhancing Event-based Monocular Depth Estimation via Dense-to-Sparse Distillation
Jinjing Zhu (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
Depth EstimationKnowledge DistillationTransformerSupervised Fine-TuningMultimodality
🎯 What it does: This paper proposes EventDAM, a method for transferring the RGB depth model DAM to event cameras through dense-to-sparse knowledge distillation, achieving unlabelled event depth estimation.
Depth AnyEvent: A Cross-Modal Distillation Paradigm for Event-Based Monocular Depth Estimation
Luca Bartolomei (Advanced Research Center on Electronic System), Stefano Mattoccia (Advanced Research Center on Electronic System)
Depth EstimationDomain AdaptationKnowledge DistillationRecurrent Neural NetworkSupervised Fine-TuningMultimodality
🎯 What it does: Proposes a cross-modal distillation paradigm that uses dense depth labels generated by the Vision Foundation Model to train a monocular depth estimation network for event cameras, and presents a recursive model.
DEPTHOR: Depth Enhancement from a Practical Light-Weight dToF Sensor and RGB Image
Jijun Xiang (Huazhong University of Science and Technology), Xin Yang (Huazhong University of Science and Technology)
RestorationDepth EstimationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a depth enhancement method for dToF sensors based on deep completion, called DepthOR, which includes a noise-robust training strategy and a two-stage network that incorporates monocular depth estimation.
DepthSync: Diffusion Guidance-Based Depth Synchronization for Scale- and Geometry-Consistent Video Depth Estimation
Yue-Jiang Dong (Tsinghua University), Song-Hai Zhang (Tsinghua University)
Depth EstimationDiffusion modelVideo
🎯 What it does: A training-agnostic framework named DepthSync is proposed, which achieves cross-window scale synchronization and intra-window geometric alignment by guiding the forward and backward processes in the diffusion of video depth estimation, thereby enhancing the scale and geometric consistency of long video depth estimation.
DeRIS: Decoupling Perception and Cognition for Enhanced Referring Image Segmentation through Loopback Synergy
Ming Dai (Southeast University), Wankou Yang
Object DetectionSegmentationTransformerSupervised Fine-TuningImageMultimodality
🎯 What it does: Decouples the RIS task into two main modules: perception and cognition, proposing the DeRIS framework, and enhances segmentation accuracy and text-image understanding through iterative query interactions with feedback loops.
Derm1M: A Million-scale Vision-Language Dataset Aligned with Clinical Ontology Knowledge for Dermatology
Siyuan Yan (Monash University), Zongyuan Ge (Monash University)
ClassificationRecognitionRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityAudio
🎯 What it does: This paper constructs Derm1M, the first large-scale audiovisual language dataset containing 1,029,761 pairs of skin images and text, 390 types of skin diseases, and 130 clinical concepts, and pre-trains the DermLIP visual-language model based on this dataset.
Describe Anything: Detailed Localized Image and Video Captioning
Long Lian (NVIDIA), Yin Cui (NVIDIA)
RecognitionObject DetectionSegmentationTransformerPrompt EngineeringVision Language ModelImageVideoBenchmark
🎯 What it does: A detailed local description model for user-specified areas in images and videos is proposed - Describe Anything Model (DAM).
Describe, Adapt and Combine: Empowering CLIP Encoders for Open-set 3D Object Retrieval
Zhichuan Wang (Huazhong Agricultural University), Xiang Bai (Huazhong University of Science and Technology)
RetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageMultimodalityPoint Cloud
🎯 What it does: This paper proposes the DAC framework, which utilizes multi-view images in combination with CLIP and a multimodal large language model to achieve open-set 3D object retrieval.
Describe, Don't Dictate: Semantic Image Editing with Natural Language Intent
En Ci (Nanjing University), Ying Tai (Nanjing University)
Image TranslationGenerationDiffusion modelImageTextBenchmark
🎯 What it does: This paper proposes a descriptive editing framework called DescriptiveEdit, which redefines image editing tasks as text-to-image generation based on descriptions, and injects reference image features into the generation process through Cross-Attentive UNet without altering the weights of the original pre-trained T2I model.
DeSPITE: Exploring Contrastive Deep Skeleton-Pointcloud-IMU-Text Embeddings for Advanced Point Cloud Human Activity Understanding
Thomas Kreutz (Technical University Darmstadt), Alejandro Sanchez Guinea (Technical University Darmstadt)
RecognitionRetrievalRecurrent Neural NetworkTransformerContrastive LearningTextMultimodalityPoint Cloud
🎯 What it does: The DeSPITE model is proposed, which maps LiDAR point clouds, skeletons, IMU, and text into a shared embedding space to achieve cross-modal matching, retrieval, and pre-training for human action recognition.
Details Matter for Indoor Open-vocabulary 3D Instance Segmentation
Sanghun Jung (University of Washington), Cheng-Hao Kuo (Amazon Lab126)
Object DetectionSegmentationVision Language ModelPoint Cloud
🎯 What it does: A two-stage open vocabulary 3D instance segmentation method is proposed, first generating and aggregating 3D candidate boxes using a visual language model, and then filtering and classifying them with Alpha-CLIP and SMS;
Detect Anything 3D in the Wild
Hanxue Zhang (OpenDriveLab at Shanghai AI Laboratory), Zetong Yang (GAC R&D Center)
Object DetectionDepth EstimationAutonomous DrivingTransformerContrastive LearningImagePoint Cloud
🎯 What it does: This paper proposes DetAny3D, a foundational model that enables 3D detection on any monocular image through prompts such as boxes, points, and text.
Detection, Pose Estimation and Segmentation for Multiple Bodies: Closing the Virtuous Circle
Miroslav Purkrabek (Czech Technical University in Prague), Jiri Matas (Czech Technical University in Prague)
Object DetectionSegmentationPose EstimationTransformerImage
🎯 What it does: This paper proposes a closed-loop iterative framework (BBox‑Mask‑Pose, BMP) that achieves self-improvement in detection, instance segmentation, and pose estimation in multi-person scenes through the mutual conditioning of three major models: detection, segmentation, and pose estimation.
Deterministic Object Pose Confidence Region Estimation
Jinghao Wang (National University of Defense Technology), Qifeng Yu (National University of Defense Technology)
Pose EstimationPoint Cloud
🎯 What it does: A deterministic method is proposed to estimate the confidence interval of 6D pose by directly regressing the Gaussian distribution of keypoints and utilizing ICP and the implicit function theorem.
Devil is in the Uniformity: Exploring Diverse Learners within Transformer for Image Restoration
Shihao Zhou (Nankai University), Jufeng Yang (Nankai University)
RestorationTransformerImage
🎯 What it does: For the image restoration task, a Transformer-based model called HINT is proposed, which achieves image denoising, dehazing, deraining, desnowing, and low-light enhancement by improving the multi-head attention mechanism.
DexH2R: A Benchmark for Dynamic Dexterous Grasping in Human-to-Robot Handover
Youzhuo Wang (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
Robotic IntelligenceReinforcement LearningDiffusion modelMultimodalityPoint CloudBenchmark
🎯 What it does: Developed the DexH2R human-robot handover dataset and proposed the DynamicGrasp three-stage dynamic grasping framework.
DexVLG: Dexterous Vision-Language-Grasp Model at Scale
Jiawei He (Beijing Academy of Artificial Intelligence), He Wang (Beijing Academy of Artificial Intelligence)
Pose EstimationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelFlow-based ModelPoint CloudMesh
🎯 What it does: A large-scale simulated grasping dataset, DexGraspNet 3.0, was constructed, and a graspable hand pose model guided by language, DexVLG, was trained based on this dataset.
DGTalker: Disentangled Generative Latent Space Learning for Audio-Driven Gaussian Talking Heads
Xiaoxi Liang (Peking University), Ge Li (Peking University)
GenerationData SynthesisGaussian SplattingVideoAudio
🎯 What it does: This paper proposes a 3D audio-driven head animation framework called DGTalker based on 3D Gaussian Splatting, which can generate high-fidelity talking heads in real-time from monocular video, viewable from any angle.
DH-FaceVid-1K: A Large-Scale High-Quality Dataset for Face Video Generation
Donglin Di (Li Auto), Xun Yang (University of Science and Technology of China)
GenerationData SynthesisTransformerDiffusion modelVideoMultimodality
🎯 What it does: A large-scale high-quality facial video dataset DH-FaceVid-1K has been proposed and released, and a systematic evaluation of various text-to-video and image-to-video generation models has been conducted based on this dataset.
Di[M]O: Distilling Masked Diffusion Models into One-step Generator
Yuanzhi Zhu (Polytechnic Institute), Vicky Kalogeiton (Polytechnic Institute)
GenerationKnowledge DistillationDiffusion modelImageText
🎯 What it does: A method called Di[M]O is proposed, which distills a multi-step Masked Diffusion Model (MDM) into a single-step generator, achieving high-quality image generation in a single step.
DIA: The Adversarial Exposure of Deterministic Inversion in Diffusion Models
Seunghoo Hong (Sungkyunkwan University), Simon S. Woo (Sungkyunkwan University)
Image TranslationAdversarial AttackDiffusion modelFlow-based ModelImage
🎯 What it does: This paper proposes an attack method called DIA based on the DDIM reverse trajectory, aimed at disrupting real image editing based on DDIM reverse.
Diagnosing Pretrained Models for Out-of-distribution Detection
Haipeng Xiong (National University of Singapore), Angela Yao (National University of Singapore)
Anomaly DetectionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper systematically diagnoses the deficiencies of torchvision pre-trained models in OOD detection and proposes two improvement methods (AugDelete and AugRevise) to enhance OOD performance.
DialNav: Multi-turn Dialog Navigation with a Remote Guide
Leekyeung Han (Korea University), Paul Hongsuck Seo (Korea University)
Graph Neural NetworkTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper presents DialNav, defining a multi-turn dialogue navigation task where a remote guider collaborates with a navigator, and releases the RAIN dataset containing 2,231 entries of trajectories and dialogues.
DICE: Staleness-Centric Optimizations for Parallel Diffusion MoE Inference
Jiajun Luo (Tsinghua University), Zhi Wang (Southern University of Science and Technology)
GenerationOptimizationComputational EfficiencyMixture of ExpertsDiffusion modelImage
🎯 What it does: The DICE framework is proposed to reduce old activation distortion in MoE diffusion model inference and improve efficiency.
DictAS: A Framework for Class-Generalizable Few-Shot Anomaly Segmentation via Dictionary Lookup
Zhen Qu (Institute of Automation Chinese Academy of Sciences), Guiguang Ding (Tsinghua University)
SegmentationAnomaly DetectionTransformerVision Language ModelContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes the DictAS framework, which rephrases the few-shot anomaly segmentation task as a dictionary lookup problem, enabling the detection of anomalies in unseen categories without retraining on the target data.
Diff2I2P: Differentiable Image-to-Point Cloud Registration with Diffusion Prior
Juncheng Mu (Tsinghua University), Yue Gao (Tsinghua University)
OptimizationDiffusion modelScore-based ModelImagePoint CloudBenchmark
🎯 What it does: A fully differentiable image-to-point cloud (I2P) registration framework called Diff2I2P is proposed, which utilizes controlled side score distillation (CSD) from diffusion models and a differentiable registration solver BPnP during training to learn cross-modal features and correspondences.
DiffDoctor: Diagnosing Image Diffusion Models Before Treating
Yiyang Wang (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
RestorationSegmentationAnomaly DetectionTransformerSupervised Fine-TuningDiffusion modelImage
🎯 What it does: Build a pixel-level defect detector and use it as a diagnostic tool, then fine-tune the diffusion model using the pixel-level feedback provided by the detector to form a complete diagnostic-treatment pipeline.
Differentiable Room Acoustic Rendering with Multi-View Vision Priors
Derong Jin (University of Maryland), Ruohan Gao (University of Maryland)
Neural Radiance FieldImageAudio
🎯 What it does: A differentiable room acoustic rendering framework AV-DAR is proposed, which uses visual features to guide the acoustic reflection response and achieves efficient, differentiable RIR (Room Impulse Response) estimation through sound beam tracing.
Differential-informed Sample Selection Accelerates Multimodal Contrastive Learning
Zihua Zhao (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
GenerationRetrievalOptimizationComputational EfficiencyTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes an online differential information sample selection method called DISSect, aimed at accelerating multimodal contrastive learning, with a focus on addressing the issue of noisy correspondence (label/noisy correspondence) that arises during training.
Differentially Private Fine-Tuning of Diffusion Models
Yu-Lin Tsai (National Yang Ming Chiao University), Francois Buet-Golfouse (Barclays)
GenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: To address the privacy leakage issue of diffusion models, a method called DP-LoRA is proposed for differential privacy fine-tuning on implicit diffusion models (LDM).
DiffIP: Representation Fingerprints for Robust IP Protection of Diffusion Models
Zhuoling Li (Lancaster University), Hossein Rahmani (Lancaster University)
GenerationRepresentation LearningDiffusion modelText
🎯 What it does: Proposes the DiffIP framework for identifying whether diffusion models have been pirated or derived after fine-tuning.
DiffPCI: Large Motion Point Cloud frame Interpolation with Diffusion Model
Tianyu Zhang (Nankai University), Jin Xie (Nanjing University)
Data SynthesisAutonomous DrivingGraph Neural NetworkTransformerDiffusion modelPoint Cloud
🎯 What it does: A point cloud frame interpolation method based on diffusion models, DiffPCI, is proposed, transforming the interpolation task into a stepwise denoising diffusion process.
DiffRefine: Diffusion-based Proposal Specific Point Cloud Densification for Cross-Domain Object Detection
Sangyun Shin (University of Oxford), Niki Trigoni (University of Oxford)
Object DetectionDomain AdaptationAutonomous DrivingTransformerDiffusion modelPoint Cloud
🎯 What it does: This paper proposes a candidate box-specific point cloud densification module based on a diffusion model, which enhances the surface features of sparse point clouds and improves the detection rate of distant targets in cross-domain 3D object detection.
DiffSim: Taming Diffusion Models for Evaluating Visual Similarity
Yiren Song (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
RetrievalDiffusion modelImage
🎯 What it does: This paper proposes the DiffSim method, which utilizes the attention layers of the pre-trained diffusion model U-Net to extract features and calculates image similarity through the Aligned Attention Score.
DiffTell: A High-Quality Dataset for Describing Image Manipulation Changes
Zonglin Di (University of California), Yang Liu (University of California)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageTextMultimodality
🎯 What it does: A high-quality IDC dataset called DiffTell, covering various types of image differences, has been proposed and released, enhancing the performance of multimodal large models in the task of image difference description.
Diffuman4D: 4D Consistent Human View Synthesis from Sparse-View Videos with Spatio-Temporal Diffusion Models
Yudong Jin (Zhejiang University), Xiaowei Zhou (Zhejiang University)
GenerationData SynthesisPose EstimationDiffusion modelAuto EncoderGaussian SplattingVideo
🎯 What it does: We propose Diffuman4D, a sparse view human video synthesis method based on a 4D latent diffusion model, capable of generating high-fidelity, spatiotemporally consistent panoramic views.
DiffuMatch: Category-Agnostic Spectral Diffusion Priors for Robust Non-rigid Shape Matching
Emery Pierson (Ecole Polytechnique), Maks Ovsjanikov (Technical University of Munich)
TransformerDiffusion modelScore-based ModelMesh
🎯 What it does: By training diffusion models in the spectral domain to learn functional mapping distributions, non-rigid shape matching is achieved without category constraints.
Diffusion Curriculum: Synthetic-to-Real Data Curriculum via Image-Guided Diffusion
Yijun Liang (University of Maryland), Tianyi Zhou (University of Maryland)
ClassificationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes an image-guided synthesis technique based on diffusion models, constructing a data interpolation curve from synthetic to real data, and designing an adaptive curriculum learning strategy based on this curve to enhance model robustness against low-quality or scarce data.
Diffusion Epistemic Uncertainty with Asymmetric Learning for Diffusion-Generated Image Detection
Yingsong Huang (Tencent Inc), Qi Xiong (Tencent Inc)
ClassificationAnomaly DetectionConvolutional Neural NetworkDiffusion modelContrastive LearningImage
🎯 What it does: A method for image authenticity detection based on diffusion models, focusing on epistemic uncertainty and asymmetric learning, is proposed to distinguish between real images and diffusion-generated images.
Diffusion Guided Adaptive Augmentation for Generalization in Visual Reinforcement Learning
Jeong Woon Lee (Kyung Hee University), Hyoseok Hwang (Kyung Hee University)
Domain AdaptationReinforcement LearningDiffusion modelImageVideo
🎯 What it does: This paper proposes Diffusion Guided Adaptive Augmentation (DGA2), which generates multi-domain visual data through Stable Diffusion and retains task-relevant semantics using saliency masks, forming an adaptive domain augmentation method.
Diffusion Image Prior
Hamadi Chihaoui (University of Bern), Paolo Favaro (University of Bern)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: Proposes DIffusion Image Prior (DIIP), a completely blind image restoration method that optimizes input noise using a frozen pre-trained diffusion model, combined with a self-supervised early stopping strategy to achieve high-quality restoration for various degradation tasks.
Diffusion Transformer meets Multi-level Wavelet Spectrum for Single Image Super-Resolution
Peng Du (Samsung Research and Development Institute China), Feng Zhu (Samsung Research and Development Institute China)
RestorationSuper ResolutionTransformerDiffusion modelImage
🎯 What it does: The image is decomposed into multi-scale frequency spectra using discrete wavelet transform, and super-resolution reconstruction is performed on this spectrum using a diffusion Transformer.
Diffusion-based 3D Hand Motion Recovery with Intuitive Physics
Yufei Zhang (Rensselaer Polytechnic Institute), Qiang Ji (Rensselaer Polytechnic Institute)
GenerationPose EstimationTransformerDiffusion modelMesh
🎯 What it does: Utilizing a conditional diffusion model combined with intuitive physical constraints to enhance single-frame 3D hand reconstruction, generating temporally consistent and physically feasible hand motion sequences.
Diffusion-Based Extreme High-speed Scenes Reconstruction with the Complementary Vision Sensor
Yapeng Meng (Tsinghua University), Rong Zhao (Tsinghua University)
RestorationGenerationConvolutional Neural NetworkDiffusion modelVideo
🎯 What it does: A high frame rate video reconstruction framework CBRDM based on a dual-channel complementary visual sensor (RGB + high frame rate sparse spatiotemporal difference) is proposed.
Diffusion-Based Imaginative Coordination for Bimanual Manipulation
Huilin Xu (Fudan University), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
Robotic IntelligenceTransformerDiffusion modelVideo
🎯 What it does: A unified diffusion model framework is proposed to jointly predict video (latent representations of future frames) and action sequences, achieving dual-arm collaborative control.
Diffusion-based Source-biased Model for Single Domain Generalized Object Detection
Han Jiang (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
Object DetectionDomain AdaptationDiffusion modelImage
🎯 What it does: A single-source domain generalization object detection framework based on diffusion models, SDG-DiffDet, has been designed, which includes a memory-guided diffusion module and a source-guided denoising module.
DiffVSR: Revealing an Effective Recipe for Taming Robust Video Super-Resolution Against Complex Degradations
Xiaohui Li (Shanghai Jiao Tong University), Yu Qiao (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
RestorationSuper ResolutionDiffusion modelAuto EncoderVideo
🎯 What it does: The DiffVSR framework is designed, proposing an advanced learning strategy (PLS) and inter-layer latent transition (ILT) to achieve high-quality super-resolution of complex degraded videos.
DiGA3D: Coarse-to-Fine Diffusional Propagation of Geometry and Appearance for Versatile 3D Inpainting
Jingyi Pan (Hong Kong University of Science and Technology), Qiong Luo (Hong Kong University of Science and Technology)
RestorationGenerationDiffusion modelGaussian SplattingPoint CloudMesh
🎯 What it does: A multi-view 3D reconstruction pipeline based on diffusion models, DiGA3D, is proposed, supporting tasks such as object removal, re-texturing, and replacement.
DIH-CLIP: Unleashing the Diversity of Multi-Head Self-Attention for Training-Free Open-Vocabulary Semantic Segmentation
Songsong Duan (Xidian University), Nannan Wang (Xidian University)
SegmentationTransformerContrastive LearningImage
🎯 What it does: Proposes a training method for an unsupervised Selective Head Attention mechanism and improves the CLIP model for open vocabulary semantic segmentation;
DIMCIM: A Quantitative Evaluation Framework for Default-mode Diversity and Generalization in Text-to-Image Generative Models
Revant Teotia (New York University), Matthew Muckley (Meta)
Object DetectionGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: The DIMCIM framework is proposed to quantify the default mode diversity (Does‑It) of text-to-image models under no-reference image conditions and the generalization ability under explicit prompts (Can‑It).
DimensionX: Create Any 3D and 4D Scenes from a Single Image with Decoupled Video Diffusion
Wenqiang Sun (Hong Kong University of Science and Technology), Yikai Wang (Tsinghua University)
GenerationData SynthesisDiffusion modelOptical FlowImageVideo
🎯 What it does: Proposes the DimensionX framework, which utilizes decoupled video diffusion technology to generate high-quality 3D and 4D scenes from a single image.
DIMO: Diverse 3D Motion Generation for Arbitrary Objects
Linzhan Mou (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringImageVideo
🎯 What it does: Generate diverse 3D motions and corresponding 4D content from a single image, achieving rich motion effects with a single forward inference.
DiMPLe - Disentangled Multi-Modal Prompt Learning: Enhancing Out-Of-Distribution Alignment with Invariant and Spurious Feature Separation
Umaima Rahman (Mohamed Bin Zayed University of Artificial Intelligence), Dwarikanath Mahapatra (Khalifa University)
ClassificationDomain AdaptationExplainability and InterpretabilityContrastive LearningImageMultimodality
🎯 What it does: A new multimodal learning method called DiMPLe is proposed, which enhances the model's alignment capability on out-of-distribution data by separating invariant features from spurious features.
Diorama: Unleashing Zero-shot Single-view 3D Indoor Scene Modeling
Qirui Wu (Simon Fraser University), Angel X. Chang (Simon Fraser University)
Pose EstimationDepth EstimationRetrievalOptimizationTransformerLarge Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper presents the Diorama system, which can achieve zero-shot modeling of panoramic 3D scenes, including architectural structures, object retrieval, pose estimation, and scene layout optimization, based solely on a single RGB image without the need for training or manual annotation.
DIP: Unsupervised Dense In-Context Post-training of Visual Representations
Sophia Sirko-Galouchenko (Valeo), Nicolas Thome (Sorbonne University)
SegmentationDepth EstimationRetrievalDomain AdaptationRepresentation LearningTransformerDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised post-training method called DIP, which utilizes automatically generated pseudo-context tasks to enhance the dense representation of visual encoders, leading to better performance in retrieval-based scene understanding tasks.
Dirichlet-Constrained Variational Codebook Learning for Temporally Coherent Video Face Restoration
Baoyou Chen (Fudan University), Siyu Zhu (Alibaba Group)
RestorationTransformerAuto EncoderVideo
🎯 What it does: This work proposes a variational codebook learning method based on the Dirichlet distribution, which continuousizes the codebook representation of traditional discrete VQ-VAE to achieve temporal consistency recovery of video faces.
DiSCO-3D : Discovering and Segmenting Sub-Concepts from Open-vocabulary Queries in NeRF
Doriand Petit (Universite Paris-Saclay), Loïc Barthe (IRIT)
Object DetectionSegmentationNeural Radiance FieldContrastive LearningImageBenchmark
🎯 What it does: This paper proposes the 3D Open Vocabulary Sub-concept Discovery (OV-SD) task and designs the DiSCO-3D method, which combines unsupervised semantic segmentation and open vocabulary guidance in NeRF to achieve adaptive 3D semantic segmentation for both scenes and user queries.
DisCo: Towards Distinct and Coherent Visual Encapsulation in Video MLLMs
Jiahe Zhao (University of Chinese Academy of Sciences), Hengshuang Zhao (University of Hong Kong)
GenerationComputational EfficiencyTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes DisCo, a method for achieving visual encapsulation in video multimodal large language models, capable of generating semantically clear and temporally consistent visual tokens.
Discontinuity-aware Normal Integration for Generic Central Camera Models
Francesco Milano (ETH Zurich), Robert Thiel (Meta)
Depth EstimationOptimizationImageBenchmark
🎯 What it does: An explicit discrete point fusion method based on the local plane assumption and light ray direction is proposed to recover 3D depth from surface normal maps;
DisCoPatch: Taming Adversarially-driven Batch Statistics for Improved Out-of-Distribution Detection
Francisco Caetano (Eindhoven University of Technology), Fons van der Sommen (Eindhoven University of Technology)
Anomaly DetectionAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes DisCoPatch, an unsupervised lightweight OOD detection framework that utilizes batch statistics generated by Batch Normalization (BN) and adversarial VAE's reconstruction/generation patches to train the discriminator.
DisCoRD: Discrete Tokens to Continuous Motion via Rectified Flow Decoding
Jungbin Cho (Yonsei University), Youngjae Yu (POSTECH)
GenerationData SynthesisPose EstimationRectified FlowVideoMultimodality
🎯 What it does: The DisCoRD method is proposed, which iteratively decodes discrete action tokens into continuous and natural action sequences through Rectified Flow;
Discovering Divergent Representations between Text-to-Image Models
Lisa Dunlap (University of California), Bryan Russell (Adobe Research)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: Proposes the CompCon method, which automatically discovers and describes the visual attribute differences generated by text-to-image models under input dependencies and the corresponding prompt descriptions.
Discretized Gaussian Representation for Tomographic Reconstruction
Shaokai Wu (Shanghai Jiao Tong University), Hongtao Lu (Shanghai Jiao Tong University)
RestorationOptimizationGaussian SplattingImageComputed Tomography
🎯 What it does: A CT reconstruction framework based on Discrete Gaussian Representation (DGR) is proposed, which can directly generate three-dimensional voxel volumes in an end-to-end manner.
DisenQ: Disentangling Q-Former for Activity-Biometrics
Shehreen Azad (University of Central Florida), Yogesh Singh Rawat (University of Central Florida)
RecognitionTransformerVision Language ModelVideoMultimodality
🎯 What it does: This paper proposes a DisenQ Transformer framework based on multimodal language guidance, capable of identity recognition in various daily activities.
Disentangled Clothed Avatar Generation with Layered Representation
Weitian Zhang (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
GenerationDiffusion modelImage
🎯 What it does: This paper proposes LayerAvatar, a hierarchical UV feature plane representation based on a single-stage diffusion model, capable of quickly (in seconds) generating fully separated clothing avatars for components such as body, hair, shoes, tops, and bottoms, while supporting animation and perspective control.
Disentangled World Models: Learning to Transfer Semantic Knowledge from Distracting Videos for Reinforcement Learning
Qi Wang (Shanghai Jiao Tong University), Wenjun Zeng (Ningbo Institute of Digital Twin)
Knowledge DistillationRepresentation LearningReinforcement LearningWorld ModelVideo
🎯 What it does: In visual reinforcement learning, the authors learn separable semantic representations through a pre-trained action-free video prediction model and transfer this knowledge to downstream tasks' world models via offline-to-online latent distillation, achieving sample-efficient policy learning.