CVPR 2025 Papers — Page 14
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis
Jian Han (ByteDance), Xiaobing Liu (ByteDance)
GenerationData SynthesisTransformerImageText
🎯 What it does: Infinity is proposed, a visual autoregressive model based on bit quantization, utilizing an infinite vocabulary classifier and a bit self-correction mechanism to achieve high-resolution (1024×1024) text-to-image generation, leading in both speed and quality.
INFP: Audio-Driven Interactive Head Generation in Dyadic Conversations
Yongming Zhu (Bytedance), Zhipeng Ge (Bytedance)
GenerationTransformerDiffusion modelFlow-based ModelVideoAudio
🎯 What it does: An audio-driven interactive head generation framework named INFP is proposed, which can dynamically switch between speaking and listening states based on the audio from both parties during a two-person conversation, generating videos with realistic facial expressions and head movements.
InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment
Yunhong Lu (Zhejiang University), Min Zhang (Zhejiang University)
GenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageText
🎯 What it does: A direct preference optimization framework called DDIM-InPO is proposed, based on DDIM reparameterization and inverse reasoning, achieving efficient alignment of text-to-image diffusion models by fine-tuning latent variables that are highly correlated with human preferences.
Insight-V: Exploring Long-Chain Visual Reasoning with Multimodal Large Language Models
Yuhao Dong (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIVision Language ModelTextMultimodality
🎯 What it does: The Insight-V system is proposed, specifically designed to enhance the long-chain visual reasoning capabilities of multimodal large language models (MLLMs), utilizing a two-stage data generation and multi-agent training process.
InsightEdit: Towards Better Instruction Following for Image Editing
Yingjing Xu (Zhejiang University), Qiang Liu (01.AI)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes InsightEdit, an end-to-end instruction-based image editing framework based on a multimodal large language model, and constructs a high-quality large-scale dataset called AdvancedEdit.
Insightful Instance Features for 3D Instance Segmentation
Wonseok Roh (Korea University), Sangpil Kim (Korea University)
Object DetectionSegmentationConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud
🎯 What it does: The IKNE framework is proposed, improving instance candidate features in 3D instance segmentation;
Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning
Hanxun Yu (Zhejiang University), Jianke Zhu (Udeer.ai)
RecognitionObject DetectionSegmentationRetrievalComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodalityPoint Cloud
🎯 What it does: A unified instance-aware large multimodal model, Inst3D-LMM, is proposed, capable of performing various tasks such as 3D visual localization, question answering, and dense description without the need for separate fine-tuning for each task.
InsTaG: Learning Personalized 3D Talking Head from Few-Second Video
Jiahe Li (Beihang University), Lin Gu
GenerationData SynthesisGaussian SplattingVideoAudio
🎯 What it does: The InsTaG framework is proposed, which can quickly learn personalized 3D talking heads from just a few seconds of video and achieve high-quality real-time rendering.
Instance-wise Supervision-level Optimization in Active Learning
Shinnosuke Matsuo (Kyushu University), Masahiro Nomura (CyberAgent, Inc.)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: An instance-level supervised layer optimization active learning framework, ISO, has been developed to dynamically decide the full annotation or weak annotation for each instance under a fixed budget.
InstanceCap: Improving Text-to-Video Generation via Instance-aware Structured Caption
Tiehan Fan (Nanjing University), Ying Tai (Nanjing University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodalityChain-of-Thought
🎯 What it does: This paper proposes the InstanceCap framework, which achieves instance-aware structured video subtitle generation and applies it to text-to-video generation.
InstanceGaussian: Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception
Haijie Li (Peking University), Jian Zhang (Peking University)
Object DetectionSegmentationRetrievalGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes InstanceGaussian, which jointly learns the appearance and semantic features of 3D Gaussian points and achieves category-free instance segmentation through bottom-up aggregation.
Instant Adversarial Purification with Adversarial Consistency Distillation
Chun Tong Lei (City University of Hong Kong), Chun Pong Lau (City University of Hong Kong)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: A single-step controlled adversarial purification framework OSCP is proposed for one-step purification of adversarial samples in diffusion models.
Instant Gaussian Stream: Fast and Generalizable Streaming of Dynamic Scene Reconstruction via Gaussian Splatting
Jinbo Yan (Peking University), Ronggang Wang (Peking University)
TransformerGaussian SplattingOptical FlowVideo
🎯 What it does: A streaming dynamic scene reconstruction framework named Instant Gaussian Stream (IGS) is proposed, capable of completing reconstruction in about 2 seconds per frame and rendering in real-time;
Instant3dit: Multiview Inpainting for Fast Editing of 3D Objects
Amir Barda (Tel Aviv University), Thibault Groueix (Adobe Research)
RestorationGenerationDiffusion modelNeural Radiance FieldGaussian SplattingPoint CloudMesh
🎯 What it does: We propose Instant3dit, which transforms 3D editing into multi-view image inpainting, supporting rapid local generation for NeRF, Gaussian Splat, and meshes.
Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning
Sherry X. Chen (University of California), Pradeep Sen (University of California)
GenerationData SynthesisDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes a self-supervised Instruct-CLIP method that utilizes contrastive learning to semantically align image pairs (original images and edited images) with their corresponding editing instructions, thereby improving and expanding the existing instruction-driven image editing dataset. Based on this, InstructPix2Pix is fine-tuned, significantly enhancing editing quality.
Instruction-based Image Manipulation by Watching How Things Move
Mingdeng Cao (University of Tokyo), Zhihao Xia (Adobe)
Image TranslationGenerationPose EstimationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelOptical FlowImageVideoMultimodality
🎯 What it does: This paper proposes an instruction-based image editing model called InstructMove, which is trained using frame pairs extracted from videos and editing instructions generated by a multimodal large language model, supporting complex transformations such as non-rigid editing, poses, expressions, and perspectives.
Integral Fast Fourier Color Constancy
Wenjun Wei (University of Science and Technology of China), Yi Jin (University of Science and Technology of China)
Computational EfficiencyImage
🎯 What it does: This paper proposes the Integral Fast Fourier Color Constancy (IFFCC) algorithm, which achieves real-time automatic white balance in multi-light source scenarios.
InterAct: Advancing Large-Scale Versatile 3D Human-Object Interaction Generation
Sirui Xu, Liang-Yan Gui
GenerationData SynthesisOptimizationDiffusion modelVideoTextMultimodality
🎯 What it does: A large-scale 3D human-object interaction dataset, InterAct, has been constructed, containing 21.81 hours (extended to 30.70 hours) of data, along with a unified optimization and multi-task generation framework.
InteractAnything: Zero-shot Human Object Interaction Synthesis via LLM Feedback and Object Affordance Parsing
Jinlu Zhang (Peking University), Siyuan Huang (Peking University)
GenerationData SynthesisPose EstimationTransformerLarge Language ModelDiffusion modelImagePoint Cloud
🎯 What it does: This paper proposes InteractAnything, which utilizes feedback from large language models (LLM) and a 2D diffusion model to analyze the feasibility of objects, achieving zero-shot human-object interaction (HOI) synthesis for any 3D object.
InteractionMap: Improving Online Vectorized HDMap Construction with Interaction
Kuang Wu (Langge Technology), Zhanbin Li (Langge Technology)
Object DetectionAutonomous DrivingRecurrent Neural NetworkTransformerImage
🎯 What it does: This paper proposes InteractionMap, designing an online vectorized HD map construction framework based on DETR, which combines spatial relationship embedding, keyframe hierarchical temporal fusion, and geometric-aware alignment to achieve precise predictions of map elements such as lanes and pedestrian crossing lines.
Interactive Medical Image Analysis with Concept-based Similarity Reasoning
Ta Duc Huy (Australian Institute for Machine Learning), Vu Minh Hieu Phan (Australian Institute for Machine Learning)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkContrastive LearningImageBiomedical Data
🎯 What it does: A Concept Similarity Reasoning Network (CSR) is proposed, achieving concept-level explanations and interactivity in medical image classification.
Interactive Medical Image Segmentation: A Benchmark Dataset and Baseline
Junlong Cheng (Sichuan University), Junjun He (Shanghai AI Laboratory)
SegmentationTransformerPrompt EngineeringImageBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: A large-scale interactive medical image segmentation benchmark dataset, IMed-361M, has been proposed, and based on this, a baseline model called IMIS-Net has been constructed to support various interaction methods (points, boxes, text).
InteractVLM: 3D Interaction Reasoning from 2D Foundational Models
Sai Kumar Dwivedi (Max Planck Institute for Intelligent Systems), Dimitrios Tzionas (University of Amsterdam)
Object DetectionSegmentationPose EstimationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImagePoint Cloud
🎯 What it does: This paper proposes InteractVLM, which predicts 3D contact points between humans and objects from a single outdoor image and achieves joint 3D reconstruction based on these contact points.
InterDyn: Controllable Interactive Dynamics with Video Diffusion Models
Rick Akkerman (Max Planck Institute for Intelligent Systems), Victoria Fernández Abrevaya (Max Planck Institute for Intelligent Systems)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: This paper presents InterDyn, a framework that utilizes large video diffusion models to achieve controllable interactive dynamic generation, capable of synthesizing continuous videos containing object interactions given an initial image and driving action signals.
Interleaved-Modal Chain-of-Thought
Jun Gao (Soochow University), Wenjie Li (Hong Kong Polytechnic University)
GenerationData SynthesisTransformerPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: This paper proposes an Interleaved Chain of Thought for Multimodal Reasoning (ICoT), which allows visual language models to generate intermediate reasoning chains that simultaneously contain visual segments and textual reasoning steps during the reasoning process.
InterMimic: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions
Sirui Xu (University of Illinois Urbana-Champaign), Liang-Yan Gui (University of Illinois Urbana-Champaign)
Knowledge DistillationRobotic IntelligenceTransformerReinforcement LearningMultimodalityPhysics Related
🎯 What it does: Proposes the InterMimic framework, which utilizes reinforcement learning and teacher-student distillation to train a unified physical simulation full-body interaction control strategy, achieving realistic human-object interaction.
Interpretable Generative Models through Post-hoc Concept Bottlenecks
Akshay Kulkarni (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
GenerationExplainability and InterpretabilityDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Two low-cost post-hoc concept bottleneck methods are proposed—Concept Bottleneck Autoencoder (CB-AE) and Concept Controller (CC)—to transform pre-trained generative models (GAN or diffusion models) into interpretable models and achieve concept-level image editing.
Interpretable Image Classification via Non-parametric Part Prototype Learning
Zhijie Zhu (University of New South Wales), Yang Song (University of New South Wales)
ClassificationExplainability and InterpretabilityTransformerContrastive LearningImage
🎯 What it does: This paper proposes a part-level interpretable image classification framework based on non-parametric prototype learning, utilizing ViT feature clustering to obtain diversified part prototypes for each category, which are directly used for classification.
Interpreting Object-level Foundation Models via Visual Precision Search
Ruoyu Chen (Institute of Information Engineering, Chinese Academy of Sciences), Xiaochun Cao (Sun Yat-sen University)
Object DetectionSegmentationOptimizationExplainability and InterpretabilitySupervised Fine-TuningImage
🎯 What it does: A Visual Precision Search method for object-level foundational models is proposed to generate instance-level interpretable heatmaps.
Inversion Circle Interpolation: Diffusion-based Image Augmentation for Data-scarce Classification
Yanghao Wang (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
ClassificationGenerationData SynthesisVision Language ModelDiffusion modelImage
🎯 What it does: This paper proposes an augmentation method based on diffusion models, Diff-II, which utilizes category concept learning, image inverse interpolation, and two-stage denoising to generate synthetic training samples that are both faithful and diverse, thereby improving classification performance in data-scarce scenarios.
Investigating the Role of Weight Decay in Enhancing Nonconvex SGD
Tao Sun (National University of Defense Technology), Bao Wang (University of Utah)
OptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Analyze the impact of weight decay on the convergence and generalization of non-convex SGD, and provide corresponding theoretical and experimental validation.
Invisible Backdoor Attack against Self-supervised Learning
Hanrong Zhang (Zhejiang University), Shiqing Ma (University of Massachusetts Amherst)
Representation LearningAdversarial AttackConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Designed and implemented a stealthy backdoor attack in self-supervised learning (SSL) that can induce downstream classifiers to produce specified target labels while maintaining normal model performance.
IRGS: Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing
Chun Gu (Fudan University), Li Zhang (Fudan University)
OptimizationGaussian SplattingImage
🎯 What it does: This paper proposes an inverse rendering framework IRGS based on two-dimensional Gaussian light scattering, which can accurately model indirect lighting and visibility using the complete rendering equation without making approximations.
IRIS: Inverse Rendering of Indoor Scenes from Low Dynamic Range Images
Chih-Hao Lin (Meta), Changil Kim (Meta)
RestorationOptimizationNeural Radiance FieldImage
🎯 What it does: The IRIS framework is proposed, which utilizes multi-view LDR images to recover spatially varying HDR lighting, physical material properties, and camera response functions, supporting light reconfiguration and object insertion.
Is `Right' Right? Enhancing Object Orientation Understanding in Multimodal Large Language Models through Egocentric Instruction Tuning
Ji Hyeok Jung (Sogang University), Buru Chang (Korea University)
RecognitionObject DetectionTransformerLarge Language ModelSupervised Fine-TuningImageMultimodalityBenchmark
🎯 What it does: This paper proposes a method called 'egocentric instruction tuning', which refines the understanding of direction in multimodal large language models (MLLMs) to interpret the orientation of objects in images from a user-centered perspective. It also creates the EgoOrientBench benchmark to evaluate the direction recognition capabilities of MLLMs across three tasks (Choose, Verify, Freeform) and five datasets.
Is this Generated Person Existed in Real-world? Fine-grained Detecting and Calibrating Abnormal Human-body
Zeqing Wang (Sun Yat-sen University), Yonghong Tian (Peking University)
RestorationAnomaly DetectionTransformerVision Language ModelDiffusion modelImage
🎯 What it does: This paper proposes a framework called HumanCalibrator for detecting and repairing fine-grained abnormal body parts (such as missing or redundant hands, ears, legs, etc.) in human images generated by AIGC.
Is Your World Simulator a Good Story Presenter? A Consecutive Events-Based Benchmark for Future Long Video Generation
Yiping Wang (University of Washington), Yelong Shen (Microsoft)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextBenchmark
🎯 What it does: A benchmark called StoryEval is proposed to evaluate the integrity of text-to-video models in presenting multi-event stories.
iSegMan: Interactive Segment-and-Manipulate 3D Gaussians
Yian Zhao (Peking University), Jie Chen (Peking University)
SegmentationGaussian SplattingPoint Cloud
🎯 What it does: An interactive 3D Gaussian segmentation and editing framework, iSegMan, has been designed to allow users to precisely control regions through 2D click interactions and perform various editing functions.
It's a (Blind) Match! Towards Vision-Language Correspondence without Parallel Data
Dominik Schnaus (TU Munich), Daniel Cremers (TU Munich)
OptimizationVision Language ModelImageText
🎯 What it does: This paper studies 'blind' matching without aligned training data by utilizing the similarity matrix of visual and language models to solve the quadratic assignment problem, and constructs an improved Hahn-Grant iterative solver;
ITA-MDT: Image-Timestep-Adaptive Masked Diffusion Transformer Framework for Image-Based Virtual Try-On
Ji Woo Hong (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes a Transformer-based denoising diffusion model—ITA-MDT—for virtual try-on of clothing images.
Iterative Predictor-Critic Code Decoding for Real-World Image Dehazing
Jiayi Fu (Nankai University), Chongyi Li (Nankai University)
RestorationGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a code decoding framework based on an iterative predictor-critic, utilizing a high-quality VQGAN codebook to progressively dehaze real scene fog images.
IterIS: Iterative Inference-Solving Alignment for LoRA Merging
Hongxu Chen (University of Science and Technology of China), Long Chen (Hong Kong University of Science and Technology)
OptimizationFederated LearningSafty and PrivacyTransformerLarge Language ModelVision Language ModelDiffusion modelImageText
🎯 What it does: A LoRA merging algorithm called IterIS based on an iterative reasoning-solving framework is designed to merge multi-task LoRA into a unified adapter while ensuring data privacy.
Jailbreaking the Non-Transferable Barrier via Test-Time Data Disguising
Yongli Xiang (University of Sydney), Tongliang Liu (University of Sydney)
Domain AdaptationAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: A black-box attack method for non-transferable learning models, called JailNTL, is proposed, which uses data camouflage during testing to bypass non-transferability barriers.
JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba
Xiaoyong Lu (Southeast University), Songlin Du (Southeast University)
Pose EstimationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A lightweight local feature matcher called JamMa is proposed, which utilizes a joint Mamba and JEGO (Joint-Efficient-Global-Omnidirectional) scan-merge strategy to achieve efficient global mutual interaction and complete coarse-to-fine matching.
Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation
Chengyue Wu (DeepSeek-AI), Ping Luo (University of Hong Kong)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Designed and trained Janus, a unified multimodal model capable of simultaneous image understanding and image generation. The key point is the separation of visual encoding into two independent channels (understanding encoder and generation encoder), using the same autoregressive Transformer to uniformly process text and visual information.
JanusFlow: Harmonizing Autoregression and Rectified Flow for Unified Multimodal Understanding and Generation
Yiyang Ma (Tsinghua University), Chong Ruan (Peking University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningRectified FlowImageTextMultimodality
🎯 What it does: We propose JanusFlow, a unified multimodal framework that combines autoregressive language models with rectified flow for image understanding and generation.
JarvisIR: Elevating Autonomous Driving Perception with Intelligent Image Restoration
Yunlong Lin (Xiamen University), Xinghao Ding
RestorationAutonomous DrivingTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImage
🎯 What it does: Proposes JarvisIR, an intelligent system based on VLM that uses a VLM controller to automatically schedule various specialized restoration models to handle multiple weather-degraded images in real environments;
JiSAM: Alleviate Labeling Burden and Corner Case Problems in Autonomous Driving via Minimal Real-World Data
Runjian Chen (University of Hong Kong), Ping Luo (University of Hong Kong)
Object DetectionAutonomous DrivingPoint Cloud
🎯 What it does: Train a LiDAR 3D detection model by combining simulation data with a very small amount of real annotations.
Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video
Hoang Chuong Nguyen (Australian National University), Miaomiao Liu (Australian National University)
Depth EstimationOptimizationNeural Radiance FieldVideo
🎯 What it does: This paper proposes a joint optimization framework for monocular video-based non-prior NeRF and camera motion.
Joint Out-of-Distribution Filtering and Data Discovery Active Learning
Sebastian Schmidt (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
ClassificationAnomaly DetectionData-Centric LearningImage
🎯 What it does: A joint filtering framework for out-of-distribution (OOD) and active learning for discovering new categories, named Joda, is proposed, which can efficiently select samples for labeling from a data pool containing unknown categories and out-of-domain samples.
Joint Scheduling of Causal Prompts and Tasks for Multi-Task Learning
Chaoyang Li (Harbin Institute of Technology), Qing Liao (Harbin Institute of Technology)
OptimizationRepresentation LearningTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: A joint scheduling framework based on JSCPT is proposed, utilizing causal prompts and task scheduling to achieve efficient fine-tuning of multi-task visual-language models.
Joint Vision-Language Social Bias Removal for CLIP
Haoyu Zhang (National University of Singapore), Mohan Kankanhalli (National University of Singapore)
ClassificationRetrievalTransformerContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a method to jointly remove multimodal social biases in images and texts within the CLIP model, aiming to eliminate unfair inferences caused by sensitive attributes such as gender, age, and race.
JTD-UAV: MLLM-Enhanced Joint Tracking and Description Framework for Anti-UAV Systems
Yifan Wang (Beihang University), Xuelong Li (Beijing University of Posts and Telecommunications)
Object DetectionObject TrackingTransformerLarge Language ModelVideo
🎯 What it does: Proposed the Unmanned Aerial Vehicle Tracking and Intent Understanding (UTIU) task and implemented the JTD-UAV framework.
Just Dance with pi! A Poly-modal Inductor for Weakly-supervised Video Anomaly Detection
Snehashis Majhi (INRIA), Francois Bremond (INRIA)
Anomaly DetectionKnowledge DistillationContrastive LearningOptical FlowVideoMultimodality
🎯 What it does: A weakly supervised video anomaly detection framework PI-VAD has been developed, which enhances RGB representation during the training phase using pseudo-modal generation and cross-modal induction with five additional modalities, while inference only uses RGB.
K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs
Ziheng Ouyang (Nankai University), Qibin Hou (Nankai University)
GenerationTransformerDiffusion modelImage
🎯 What it does: A training-independent LoRA fusion method called K-LoRA is designed, which utilizes Top-K selection and time step scaling to dynamically select content and style LoRA in each attention layer, achieving high-quality fusion of objects and styles.
K-Sort Arena: Efficient and Reliable Benchmarking for Generative Models via K-wise Human Preferences
Zhikai Li (Institute of Automation Chinese Academy of Sciences), Zhen Dong (University of California Berkeley)
GenerationRecommendation SystemComputational EfficiencyReinforcement LearningPrompt EngineeringImageVideoTextMultimodalityBenchmark
🎯 What it does: We propose K-Sort Arena, a visual generation model evaluation platform based on K-wise (K>2) human preference comparisons. It utilizes Bayesian probability modeling and UCB exploration-exploitation matching strategies to achieve rapid and reliable leaderboard updates.
KAC: Kolmogorov-Arnold Classifier for Continual Learning
Yusong Hu (Nankai University), Ming-Ming Cheng (Nankai University)
ClassificationTransformerPrompt EngineeringImage
🎯 What it does: A novel continuous learning classifier named Kolmogorov-Arnold Classifier (KAC) is proposed, which effectively mitigates catastrophic forgetting while maintaining stable feature distribution.
Keep the Balance: A Parameter-Efficient Symmetrical Framework for RGB+X Semantic Segmentation
Jiaxin Cai (Fuzhou University), Wenxi Liu (Fuzhou University)
SegmentationTransformerSupervised Fine-TuningImageMultimodality
🎯 What it does: A symmetric parameter-efficient fine-tuning framework is proposed for multi-modal (RGB+X) semantic segmentation, fully utilizing the feature representation capabilities of pre-trained models.
KeyFace: Expressive Audio-Driven Facial Animation for Long Sequences via KeyFrame Interpolation
Antoni Bigata (Imperial College London), Maja Pantic (Imperial College London)
GenerationData SynthesisDiffusion modelVideoMultimodalityAudio
🎯 What it does: Using a two-stage keyframe generation and interpolation method, long-duration audio-driven facial animations are generated.
Keyframe-Guided Creative Video Inpainting
Yuwei Guo (Chinese University of Hong Kong), Bo Dai (Hong Kong University)
RestorationGenerationDiffusion modelVideo
🎯 What it does: Proposes VideoRepainter, which adopts a two-stage framework: first, it uses existing image-level inpainting methods to process keyframes, and then utilizes an image-to-video diffusion model to propagate the modified content throughout the entire video;
Kiss3DGen: Repurposing Image Diffusion Models for 3D Asset Generation
Jiantao Lin (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImageMesh
🎯 What it does: The Kiss3DGen framework is proposed, which transforms a pre-trained 2D diffusion model into a 3D generative model, enabling various tasks such as text-to-3D, image-to-3D, editing, and enhancement.
KMD: Koopman Multi-modality Decomposition for Generalized Brain Tumor Segmentation under Incomplete Modalities
Tianyi Liu (Xi'an Jiaotong-Liverpool University), Kaizhu Huang (Duke Kunshan University)
SegmentationMultimodalityBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: A pluggable Koopman Multimodal Decomposition (KMD) module is proposed for brain tumor segmentation tasks under varying modal missing rates (balanced or unbalanced), capable of decomposing the features of each modality into shared and exclusive information, and constructing inter-modal relationships.
Knowledge Bridger: Towards Training-Free Missing Modality Completion
Guanzhou Ke (Beijing Jiaotong University), Hexing Su (Harbin Institute of Technology)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: A training-free multimodal missing completion framework (Knowledge Bridger) is proposed, which automatically extracts knowledge graphs through large-scale multimodal models (LMM) to guide the generation and ranking of missing modalities.
Knowledge Memorization and Rumination for Pre-trained Model-based Class-Incremental Learning
Zijian Gao (National University of Defense Technology), Huaimin Wang (National University of Defense Technology)
ClassificationTransformerImage
🎯 What it does: In replay-free class-incremental learning, researchers utilize pre-trained models and propose the MoAL method, which integrates momentum adapter weight interpolation, knowledge memory, and a reminiscence mechanism to continuously enhance model adaptability and alleviate catastrophic forgetting.
Knowledge-Aligned Counterfactual-Enhancement Diffusion Perception for Unsupervised Cross-Domain Visual Emotion Recognition
Wen Yin (University of Electronic Science and Technology of China), Tao He (Monash University)
RecognitionDomain AdaptationMixture of ExpertsVision Language ModelDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes an unsupervised cross-domain visual emotion recognition (UCDVER) task and designs a knowledge-aligned diffusion perception framework (KCDP) to achieve emotion recognition between the source domain (real images) and the target domain (stickers/abstract images).
Koala-36M: A Large-scale Video Dataset Improving Consistency between Fine-grained Conditions and Video Content
Qiuheng Wang (Shenzhen University), Di Zhang (Kuaishou Technology)
SegmentationGenerationConvolutional Neural NetworkLarge Language ModelVideoText
🎯 What it does: A large high-quality video dataset, Koala-36M, has been constructed, and an improved video segmentation, structured subtitle generation, VTSS filtering, and metric condition processing workflow has been proposed.
KVQ: Boosting Video Quality Assessment via Saliency-guided Local Perception
Yunpeng Qu (Tsinghua University), Jian Wang (Tsinghua University)
RecognitionOptimizationTransformerVideo
🎯 What it does: This paper proposes a no-reference video quality assessment framework called KVQ based on visual attention and local perception.
L-SWAG: Layer-Sample Wise Activation with Gradients Information for Zero-Shot NAS on Vision Transformers
Sofia Casarin (Free University of Bozen-Bolzano), Oswald Lanz (Free University of Bozen-Bolzano)
Neural Architecture SearchTransformerImageBenchmark
🎯 What it does: A zero-cost NAS method for Vision Transformers, L-SWAG, is proposed and validated, along with a multi-agent fusion algorithm, LIBRA-NAS, and a cross-task benchmark containing 2000 ViT models is constructed.
Label Shift Meets Online Learning: Ensuring Consistent Adaptation with Universal Dynamic Regret
Yucong Dai (National University of Defense Technology), Chenping Hou (National University of Defense Technology)
Domain AdaptationOptimizationImage
🎯 What it does: An online label distribution drift adaptive method named LASU is proposed, which combines the OSCM-L estimator and the optimized Optimistic Online Mirror Descent.
LAL: Enhancing 3D Human Motion Prediction with Latency-aware Auxiliary Learning
Xiaoning Sun (Nanjing University of Science and Technology), Shengxiang Hu (Nanjing University of Science and Technology)
Pose EstimationGraph Neural NetworkTransformerVideo
🎯 What it does: This paper proposes a 'Delayed Perception' framework (LAL) based on auxiliary learning, which utilizes previously overlooked delayed period data to enhance the accuracy of effective prediction intervals in human motion prediction.
LamRA: Large Multimodal Model as Your Advanced Retrieval Assistant
Yikun Liu (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)
RetrievalTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: Proposes the LamRA framework, which utilizes the LoRA module to enable large multimodal models with general retrieval and re-ranking capabilities.
Language Guided Concept Bottleneck Models for Interpretable Continual Learning
Lu Yu (Tianjin University of Technology), Changsheng Xu (University of Science and Technology of China)
Explainability and InterpretabilityLarge Language ModelVision Language ModelImage
🎯 What it does: This paper proposes an explainable continual learning framework based on a language-guided Concept Bottleneck Model (CBM), aimed at alleviating catastrophic forgetting while enhancing model interpretability.
Language-Assisted Debiasing and Smoothing for Foundation Model-Based Semi-Supervised Learning
Na Zheng (National University of Singapore), Roger Zimmermann (National University of Singapore)
ClassificationSegmentationTransformerLarge Language ModelPrompt EngineeringImage
🎯 What it does: A method combining language knowledge for debiasing and smoothing (LADaS) is proposed for semi-supervised learning based on pre-trained foundational models, significantly improving the quality and utilization of pseudo-labels.
Language-Guided Audio-Visual Learning for Long-Term Sports Assessment
Huangbiao Xu (Fuzhou University), Wenzhong Guo (Fuzhou University)
TransformerPrompt EngineeringVideoTextMultimodalityAudio
🎯 What it does: A multi-dimensional language-guided audio-visual learning framework MLAVL is proposed for quality assessment of long-duration sports videos.
Language-Guided Image Tokenization for Generation
Kaiwen Zha (Google DeepMind), Xiuye Gu (Google DeepMind)
GenerationCompressionTransformerVision Language ModelDiffusion modelGenerative Adversarial NetworkImageText
🎯 What it does: A text-guided image tokenization framework called TexTok is proposed, which uses image description words to provide semantic conditions for the tokenizer, enhancing compression rate and reconstruction quality.
Language-Guided Salient Object Ranking
Fang Liu (City University of Hong Kong), Rynson W. H. Lau (City University of Hong Kong)
Object DetectionSegmentationTransformerVision Language ModelImageMultimodality
🎯 What it does: A language-guided salient object ranking method LG-SOR is proposed, which utilizes image descriptions generated by a large vision-language model to assist the visual model in completing salient object ranking and segmentation.
Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection
Marc-Antoine Lavoie (University of Toronto), Steven L. Waslander (University of Toronto)
Object DetectionDomain AdaptationAutonomous DrivingTransformerContrastive LearningImage
🎯 What it does: The study focuses on target detection domain adaptation between labeled source domains and unlabeled target domains, proposing the DINO Teacher framework.
Large-scale Multi-view Tensor Clustering with Implicit Linear Kernels
Jiyuan Liu (National University of Defense Technology), Ke Liang (National University of Defense Technology)
OptimizationMultimodalityBenchmark
🎯 What it does: A large-scale multi-view tensor clustering method LMTC is proposed, which eliminates the traditional tensor rotation technique and clusters all samples by embedding their similarities into a low-rank tensor using an implicit linear kernel.
Large-Scale Text-to-Image Model with Inpainting is a Zero-Shot Subject-Driven Image Generator
Chaehun Shin (Seoul National University), Sungroh Yoon (Seoul National University)
GenerationData SynthesisTransformerDiffusion modelImageText
🎯 What it does: A zero-shot, diptych-based text-driven image generation method for painting is proposed, utilizing the dual-generation capability of the FLUX model to align reference images and text prompts.
Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis
Yousef Yeganeh (Technical University of Munich), Ehsan Adeli (Stanford University)
GenerationData SynthesisDiffusion modelImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: Proposes the Latent Drifting technique, which utilizes pre-trained diffusion models to generate and manipulate adversarial samples of medical images.
Latent Space Imaging
Matheus Souza (King Abdullah University of Science and Technology), Wolfgang Heidrich (King Abdullah University of Science and Technology)
ClassificationSegmentationCompressionGenerative Adversarial NetworkImage
🎯 What it does: A Latent Space Imaging (LSI) paradigm is proposed, which directly encodes images into the latent space of StyleGANXL using a single-pixel camera and programmable masks, achieving extremely high compression rates and supporting reconstruction and various downstream tasks.
Latent Space Super-Resolution for Higher-Resolution Image Generation with Diffusion Models
Jinho Jeong (Yonsei University), Seon Joo Kim (Yonsei University)
RestorationGenerationSuper ResolutionDiffusion modelImage
🎯 What it does: A framework for super-resolution re-sampling in latent space, LSRNA, is proposed, which combines region-adaptive noise injection to enhance the generation quality and speed of text-to-high-resolution image conversion.
LatentHOI: On the Generalizable Hand Object Motion Generation with Latent Hand Diffusion.
Muchen Li (University of British Columbia), Shugao Ma (Meta Reality Lab)
GenerationRobotic IntelligenceDiffusion modelAuto EncoderPoint Cloud
🎯 What it does: This paper proposes the LatentHOI framework, which enables the generation of hand-object interaction motions for unseen objects.
LaTexBlend: Scaling Multi-concept Customized Generation with Latent Textual Blending
Jian Jin (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: The LATEXBLEND framework is proposed, which supports the generation of multi-concept customized text-to-image by storing and blending concepts in a latent text space.
LATTE-MV: Learning to Anticipate Table Tennis Hits from Monocular Videos
Daniel Etaat (University of California), S. Shankar Sastry (University of California)
Object DetectionObject TrackingPose EstimationRobotic IntelligenceTransformerVideo
🎯 What it does: In this study, the authors built a scalable monocular video reconstruction system that reconstructed over 73,000 exchanges of 3D performance data using publicly available table tennis match footage. Based on this, they trained a transformer generative model to predict the opponent's future shot trajectories, and then combined the prediction results with confidence intervals for the pre-pose control of a robotic paddle.
LaVin-DiT: Large Vision Diffusion Transformer
Zhaoqing Wang (AIsphere), Tongliang Liu (University of Sydney)
Object DetectionSegmentationDepth EstimationTransformerDiffusion modelAuto EncoderImageVideo
🎯 What it does: A large-scale visual diffusion transformer, LaVin-DiT, is proposed, which can accomplish over 20 visual tasks through context learning without the need for fine-tuning.
Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers
Haoran You (Georgia Institute of Technology), Yingyan Celine Lin
Image TranslationRestorationGenerationCompressionTransformerDiffusion modelImageTextMultimodality
🎯 What it does: A DiffCR dynamic Diffusion Transformer inference framework is proposed, which can adaptively route computation and compress ratios across token, layer, and time step dimensions.
Layered Image Vectorization via Semantic Simplification
Zhenyu Wang (Shenzhen University), Min Lu (Shenzhen University)
SegmentationGenerationCompressionDiffusion modelScore-based ModelImage
🎯 What it does: A hierarchical vectorization method based on progressively simplified images is designed, generating editable vector layers through a two-stage construction from macro structure to detail with semantic alignment.
Layered Motion Fusion: Lifting Motion Segmentation to 3D in Egocentric Videos
Vadim Tschernezki (Visual Geometry Group University of Oxford), Andrea Vedaldi (Visual Geometry Group University of Oxford)
Object DetectionSegmentationNeural Radiance FieldVideo
🎯 What it does: By integrating 2D motion segmentation results into layered NeRF, the 3D segmentation of dynamic and semi-static objects in first-person videos is enhanced.
LayoutVLM: Differentiable Optimization of 3D Layout via Vision-Language Models
Fan-Yun Sun (Stanford University), Jiajun Wu (Google Research)
GenerationOptimizationTransformerVision Language ModelMultimodalityPoint Cloud
🎯 What it does: This paper proposes a 3D layout generation framework called LAYOUTVLM based on a vision-language model, which can generate physically feasible and semantically consistent indoor scene layouts based on unlabeled 3D assets and natural language instructions.
LC-Mamba: Local and Continuous Mamba with Shifted Windows for Frame Interpolation
Min Wu Jeong (Hanyang University), Chae Eun Rhee (Hanyang University)
Image TranslationRestorationOptical FlowVideo
🎯 What it does: Proposes the LC-Mamba model, specifically designed for video frame interpolation tasks.
LeanGaussian: Breaking Pixel or Point Cloud Correspondence in Modeling 3D Gaussians
Jiamin Wu (International Digital Economy Academy), Lei Zhang (International Digital Economy Academy)
GenerationData SynthesisOptimizationTransformerGaussian SplattingPoint Cloud
🎯 What it does: Utilizing a deformed Transformer to directly regress and iteratively optimize 3D Gaussian spheres for single-view novel view synthesis.
Learnable Infinite Taylor Gaussian for Dynamic View Rendering
Bingbing Hu (Xiamen University), Gim Hee Lee (National University of Singapore)
Gaussian SplattingVideo
🎯 What it does: A new dynamic perspective rendering framework is proposed by using a learnable infinite-order Taylor expansion for the displacement, rotation, scale, and other properties of 3D Gaussian primitives in dynamic scenes.
Learned Binocular-Encoding Optics for RGBD Imaging Using Joint Stereo and Focus Cues
Yuhui Liu (University of Hong Kong), Yifan Peng (University of Hong Kong)
Depth EstimationImage
🎯 What it does: This paper designs a hardware-software collaborative RGBD camera system that utilizes learned rank-2 differential optical elements and joint stereo + focus information to achieve high-frequency texture-rich color images and high-resolution depth maps.
Learned Image Compression with Dictionary-based Entropy Model
Jingbo Lu (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
CompressionConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a cross-attention entropy model based on a shared learnable dictionary (DCAE) to enhance the performance of entropy models in learned image compression.
Learning 4D Panoptic Scene Graph Generation from Rich 2D Visual Scene
Shengqiong Wu (National University of Singapore), Tat-seng Chua (National University of Singapore)
GenerationDomain AdaptationTransformerLarge Language ModelImageVideoMultimodalityChain-of-Thought
🎯 What it does: The paper proposes an end-to-end 4D scene graph generation framework based on a 4D large language model, addressing the open vocabulary problem through chain reasoning.
Learning Affine Correspondences by Integrating Geometric Constraints
Pengju Sun (National University of Defense Technology), Daniel Barath (ETH Zurich)
Pose EstimationDepth EstimationConvolutional Neural NetworkImage
🎯 What it does: An end-to-end network based on dense matching and geometric constraints is proposed, capable of simultaneously extracting high-quality point correspondences and local affine transformations to generate accurate affine correspondences.
Learning Audio-guided Video Representation with Gated Attention for Video-Text Retrieval
Boseung Jeong (POSTECH), Suha Kwak (POSTECH)
RetrievalRepresentation LearningTransformerContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: A gated attention-based audio-guided video representation learning framework called AVIGATE is proposed for video-text retrieval.
Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation
Takeshi Noda (Tsinghua University), Zhizhong Han (Wayne State University)
Data SynthesisOptimizationTransformerPoint CloudMesh
🎯 What it does: This paper proposes an end-to-end method for learning continuous SDF from sparse point clouds through bijective surface parameterization and mesh deformation optimization.
Learning Class Prototypes for Unified Sparse-Supervised 3D Object Detection
Yun Zhu (Nanjing University of Science and Technology), Jian Yang (Nanjing University)
Object DetectionContrastive LearningPoint Cloud
🎯 What it does: This paper proposes a unified sparse supervised 3D object detection framework that can achieve object detection with only one box label per scene in both indoor and outdoor environments.