CVPR 2026 Papers — Page 18
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Improving Vision-language Models with Perception-centric Process Reward Models
Yingqian Min (Renmin University of China), Ji-Rong Wen (Renmin University of China)
RetrievalReinforcement Learning from Human FeedbackReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose Perceval, a perception-centric process reward model, to achieve word-by-word error localization and correction in the reasoning chain of visual language models;
IMS3: Breaking Distributional Aggregation in Diffusion-Based Dataset Distillation
Chenru Wang, Chi Zhang (Westlake University)
Knowledge DistillationTransformerSupervised Fine-TuningDiffusion modelImage
🎯 What it does: Propose the IM-S3 framework, which utilizes diffusion inversion matching for fine-tuning and achieves untrained class discrimination through subgroup center similarity during the sampling stage, thereby enhancing the distribution coverage and discriminability of diffusion models in dataset distillation.
IMU-HOI: A Symbiotic Framework for Coherent Human-Object Interaction and Motion Capture via Contact-Conscious Inertial Fusion
Lizhou Lin (Shanghai Jiao Tong University), Ling Pei (Shanghai Jiao Tong University)
Pose EstimationRecurrent Neural NetworkTime Series
🎯 What it does: Propose a framework named IMU-HOI that utilizes sparse IMU to simultaneously capture full-body human pose and object 6 degrees of freedom (DoF) trajectory, and achieves cooperative fusion of both through hand-object contact probability;
In Pursuit of Pixel Supervision for Visual Pre-training
Lihe Yang (FAIR, Meta), Hu Xu (FAIR, Meta)
SegmentationDepth EstimationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes Pixio, a self-supervised visual pre-training method based on Masked AutoEncoder, which utilizes the self-supervised screening strategy MetaCLIP-S on 2B web-crawled images. It further improves the decoder depth, occlusion block size, and the number of class tokens in MAE to learn richer spatial structural representations.
Incentivizing Generative Zero-Shot Learning via Outcome-Reward Reinforcement Learning with Visual Cues
Wenjin Hou (Zhejiang University), Hehe Fan (Zhejiang University)
GenerationData SynthesisTransformerReinforcement LearningDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Designed and implemented a generative zero-shot learning framework named RLVC, where the generative model is treated as a reinforcement learning strategy. The framework uses reward signals based on task outcomes and class-level visual hints to guide feature synthesis, and employs a cold start training strategy to improve training stability.
Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning
Xiaodong Wang (Peking University), Peixi Peng (Peking University)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Propose the VideoReasoner framework, which enhances video reasoning capabilities on baseline multimodal large language models by leveraging multi-task cold start and multi-task reinforcement learning (GRPO), integrating three multimodal reasoning pipelines: event localization, keyframe detection, and direct answering.
Inconsistency-aware Multimodal Schrodinger Bridge for Deepfake Localization
Jiayu Xiong (Huaqiao University), Jun Xue (Wuhan University)
Object DetectionAnomaly DetectionTransformerVision Language ModelDiffusion modelMultimodality
🎯 What it does: Proposed a cross-modal deepfake localization framework IaMSB based on the Schrödinger Bridge (Schrödinger Bridge), which can directly output precise time intervals in audio-visual dual-modal data.
IncreFA: Breaking the Static Wall of Generative Model Attribution
Haotian Qin, Zhanyu Ma (Beijing University Of Posts And Telecommunications)
GenerationExplainability and InterpretabilityTransformerVision Language ModelImageBenchmark
🎯 What it does: Proposes the IncreFA framework to achieve continuous and scalable image attribution for emerging AI-generated models.
Incremental Object Detection via Future-Aware Decoupled Cross-Head Distillation
Chenfeng Yin (Xidian University), Xinbo Gao (Xidian University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposes two techniques, Future-Aware Decoupled Cross-Head Distillation (FaCHD) and Region Prototype Semantic Compensation (RPSC), to address the knowledge distillation conflicts and semantic drift issues caused by the coupling between the detection head and backbone network in incremental object detection.
Inference-time Physics Alignment of Video Generative Models with Latent World Models
Jianhao Yuan (FAIR, Meta Superintelligence Labs), Adriana Romero-Soriano (FAIR, Meta Superintelligence Labs)
GenerationData SynthesisReinforcement LearningDiffusion modelWorld ModelVideoPhysics Related
🎯 What it does: The WMReward method is proposed to enhance the physical plausibility of video generation by aligning pre-trained diffusion models with latent world models during the inference phase.
Inferring Compositional 4D Scenes without Ever Seeing One
Ahmet Berke Gökmen (INSAIT), Danda Pani Paudel (INSAIT)
GenerationTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: Propose the COM4D method for unsupervised reconstruction of 4D scenes with multiple objects from monocular videos.
InfiniBench: Infinite Benchmarking for Visual Spatial Reasoning with Customizable Scene Complexity
Haoming Wang (University of Pittsburgh), Wei Gao (University of Pittsburgh)
GenerationData SynthesisOptimizationLarge Language ModelAgentic AIVision Language ModelVideoMultimodalityMeshBenchmark
🎯 What it does: By integrating LLM agents, clustering-based layout optimization, and task-aware camera trajectory generation methods, InfiniBench can automatically generate infinitely customizable, physically feasible, and highly complex 3D scenes, rendering them into multi-frame videos suitable for vision-language models (VLMs).
InfiniDepth: Arbitrary-Resolution and Fine-Grained Depth Estimation with Neural Implicit Fields
Hao Yu (Zhejiang University), Sida Peng (Zhejiang University)
Depth EstimationTransformerNeural Radiance FieldImagePoint Cloud
🎯 What it does: This paper proposes InfiniDepth, which converts monocular depth estimation into depth prediction in a continuous space using neural implicit fields, thereby achieving depth estimation at arbitrary resolutions with fine-grained details.
Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout
Hidir Yesiltepe (Virginia Tech), Pinar Yanardag (Virginia Tech)
GenerationData SynthesisTransformerDiffusion modelVideoTextBenchmark
🎯 What it does: Proposed a framework named ∞-RoPE that can generate infinitely long, controllable, cinematic videos during the inference phase.
InfinityHuman: Towards Long-Term Audio-Driven Human Animation
Xiaodi Li (ByteDance), Bingyue Peng (ByteDance)
GenerationData SynthesisPose EstimationSuper ResolutionTransformerReinforcement LearningDiffusion modelFlow-based ModelAuto EncoderImageVideoAudio
🎯 What it does: Built an audio-driven full-body animation generation framework named InfinityHuman, capable of generating high-resolution, long-duration, identity-consistent, naturally animated, and emotion-controllable full-body videos.
Information-Theoretic Decomposition for Multimodal Interaction Learning
Zequn Yang (Renmin University of China), Di Hu (Renmin University of China)
Representation LearningConvolutional Neural NetworkTransformerMultimodality
🎯 What it does: Proposed a multi-modal interactive learning framework DMIL based on information theory, which can adaptively decompose and reinforce redundant, unique, and synergistic information;
InnoAds-Composer: Efficient Condition Composition for E-Commerce Poster Generation
Yuxin Qin (JD.com, Inc), Ching Law (JD.com, Inc)
GenerationTransformerDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: Propose a single-stage multi-condition e-commerce poster generation framework, InnoAds-Composer, which can simultaneously control background style, main objects, and text.
InsCal: Calibrated Multi-Source Fully Test-Time Prompt Tuning for Object Detection
Xiaofan Que (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
Object DetectionDomain AdaptationPrompt EngineeringImage
🎯 What it does: This paper proposes InsCal, a fully test-time adaptation (FTTA) method for text-driven object detection, achieving cross-domain adaptation through multi-source prompt tuning, text-guided image enhancement, and instance-level calibrated entropy minimization.
INSID3: Training-Free In-Context Segmentation with DINOv3
Claudia Cuttano (Politecnico di Torino), Stefan Roth (TU Darmstadt)
SegmentationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Proposes INSID3, a fully unsupervised, no-training, no-decoder scene segmentation method that directly performs pixel-level segmentation of arbitrary concepts (objects, parts, personalized instances) under one-shot examples by leveraging frozen DINOv3 self-supervised visual features.
Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings
Yunxiang Peng (University of Delaware), Xi Peng (University of Delaware)
ClassificationExplainability and InterpretabilityTransformerImage
🎯 What it does: This paper proposes to evaluate model generalization ability by leveraging the internal circuit structure of Vision Transformers, and introduces two novel metrics: DDB for model pre-selection and CSS for post-deployment performance monitoring.
INSIGHT Bench: Towards Grounded IN-SItu Guidance for Robotic ManipulaTion
Seonho Kim (Hanyang University), Yoonseon Oh (Korea University)
Data SynthesisRobotic IntelligenceVision-Language-Action ModelImageMultimodalityBenchmark
🎯 What it does: Proposed the INSIGHT Bench, focusing on the 'on-site guidance normalization' task involving text/symbol guidance directly embedded on objects for robots, and constructed the corresponding simulation environment and data generation framework.
Instance-level Visual Active Tracking with Occlusion-Aware Planning
Haowei Sun (South China University of Technology), Mingkui Tan (South China University of Technology)
Object TrackingData SynthesisConvolutional Neural NetworkTransformerDiffusion modelImageVideo
🎯 What it does: Proposed a complete three-module pipeline OA-VAT to address instance-level distinction and occlusion recovery in visual active tracking.
InstantRetouch: Efficient and High-Fidelity Instruction-Guided Image Retouching with Bilateral Space
Jiarui Wu (Shanghai AI Laboratory), Tianfan Xue (CPII under InnoHK)
Image HarmonizationKnowledge DistillationConvolutional Neural NetworkPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageText
🎯 What it does: Propose an efficient, faithful instruction-driven image retouching method based on bilateral space
InstantViR: Real-Time Video Inverse Problem Solver with Distilled Diffusion Prior
Weimin Bai (Peking University), He Sun (Peking University)
RestorationKnowledge DistillationTransformerDiffusion modelAuto EncoderVideo
🎯 What it does: Achieved real-time video inverse problem solving, providing high-quality video restoration in real-time.
InstAP: Instance-Aware Vision-Language Pre-Train for Spatial-Temporal Understanding
Ashutosh Kumar, Quan Kong (Toyota)
RetrievalTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: Propose the InstAP framework, which simultaneously learns global scene alignment and instance-level alignment during video-language pretraining, and constructs a large-scale dual-granularity dataset called InstVL.
Instruction-Guided Lesion Segmentation for Chest X-rays with Automatically Generated Large-Scale Dataset
Geon Choi (KAIST), Edward Choi (KAIST)
SegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextBiomedical Data
🎯 What it does: Constructed the MIMIC-ILS large-scale instruction-oriented chest X-ray lesion segmentation dataset, and trained the ROSALIA model on it to achieve precise segmentation and text generation based on natural language instructions.
InstructMix2Mix: Consistent Sparse-View Editing Through Multi-View Model Personalization
Daniel Gilo (Technion), Or Litany (Technion)
Image HarmonizationGenerationKnowledge DistillationPrompt EngineeringDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes I-Mix2Mix, a framework for multi-view consistency editing from sparse-view images.
Inter-Edit: First Benchmark for Interactive Instruction-Based Image Editing
Delong Liu (Beijing University Of Posts And Telecommunications), Fei Su (Beijing University Of Posts And Telecommunications)
Image TranslationLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelFlow-based ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose the Interactive Instruction-Guided Fuzzy Spatial Guidance (I³E) image editing task, construct a 1.1M-scale Inter-Edit training set and a human-annotated test set with 6,250 pairs, and provide location-aware evaluation metrics along with three baseline models.
Inter-Photon-Limited Videography
Andrew Xie (University of Toronto), Kiriakos N. Kutulakos (University of Toronto)
RestorationConvolutional Neural NetworkNeural Radiance FieldVideoPhysics Related
🎯 What it does: A method is proposed to reconstruct high-quality videos from single-photon detector timestamps or binary photon counts under conditions where the photon arrival rate is insufficient to track scene changes.
Interact2Ar: Full-Body Human-Human Interaction Generation via Autoregressive Diffusion Models
Pablo Ruiz-Ponce (Huawei Noah's Ark Lab), Rolandos Alexandros Potamias (Huawei Noah's Ark Lab)
GenerationTransformerVision-Language-Action ModelDiffusion modelMultimodality
🎯 What it does: Propose Interact2Ar, a text-conditioned autoregressive diffusion model for generating full-body and human-human interactions, capable of finely generating hand movements.
Interactive Episodic Memory with User Feedback
Nikesh Subedi (University of Utah), Ziad Al-Halah (University of Utah)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelVideo
🎯 What it does: Proposed an interactive retrospective memory task EM-QnF, allowing users to provide natural language feedback on model predictions and help the model iteratively improve answers.
Interactive Tracking: A Human-in-the-Loop Paradigm with Memory-Augmented Adaptation
Yuqing Huang (Pengcheng Laboratory), Ming-Hsuan Yang (UC Merced)
Object TrackingConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Proposed an interactive visual tracking framework with human-computer collaboration and constructed a specialized interactive tracking benchmark dataset called InteractTrack.
InterAgent: Physics-based Multi-agent Command Execution via Diffusion on Interaction Graphs
Bin Li (ShanghaiTech University), Jingya Wang (ShanghaiTech University)
Robotic IntelligenceGraph Neural NetworkTransformerVision-Language-Action ModelDiffusion modelMultimodalityPhysics Related
🎯 What it does: Developed an end-to-end physics-based text-driven multi-agent humanoid control framework called InterAgent to generate realistic, multi-agent interactive actions.
InternData-A1: Pioneering High-Fidelity Synthetic Data for Pre-training Generalist Policy
Yang Tian (Peking University), Jiangmiao Pang (Shanghai AI Laboratory)
Data SynthesisRobotic IntelligenceVision-Language-Action ModelFlow-based ModelMultimodalityBenchmark
🎯 What it does: Proposed and made publicly available InternData-A1, a high-fidelity synthetic dataset comprising 630k trajectories, 7433 hours, 4 robots, 18 skills, 70 tasks, and 227 scenarios; simultaneously demonstrated that a VLA model pre-trained on this dataset achieves performance comparable to or even superior to models trained on real-world data in both simulated and real tasks.
InternVideo-Next: Towards World-Understanding Video Models
Chenting Wang (Shanghai Jiao Tong University), Limin Wang (Shanghai Jiao Tong University)
RecognitionRetrievalVision Language ModelDiffusion modelContrastive LearningWorld ModelVideoMultimodalityBenchmark
🎯 What it does: Propose InternVideo-Next, a two-stage self-supervised video pre-training framework that combines pixel reconstruction with latent prediction, leveraging a semantic-guided diffusion decoder and frozen teacher's latent prediction to learn world knowledge.
InterPhys: Physics-aware Human Motion Synthesis in a Dynamic Scene
Chaoyue Xing (Australian National University), Miaomiao Liu (Australian National University)
GenerationData SynthesisTransformerDiffusion modelVideoMeshSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: Propose a human motion generation framework for dynamic scenes, utilizing a continuous contact force model and integrating physical constraints in a two-stage diffusion network to achieve physically consistent interactions with dynamic objects and static scenes.
Interpretable and Steerable Concept Bottleneck Sparse Autoencoders
Akshay Kulkarni (University of California San Diego), Kowshik Thopalli (Lawrence Livermore National Laboratory)
Explainability and InterpretabilityRepresentation LearningAuto EncoderImage
🎯 What it does: Propose the Concept Bottleneck Sparse Autoencoder (CB-SAE) framework, combining sparse autoencoders with a concept bottleneck to enhance the explainability and controllability of visual models.
Interpretable Cross-Domain Few-Shot Learning with Rectified Target-Domain Local Alignment
Yaze Zhao (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Domain AdaptationExplainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageAgriculture Related
🎯 What it does: This paper addresses the weak alignment of local fine-grained features in CLIP models for cross-domain few-shot learning by proposing a self-supervised regularization framework called CC-CDFSL based on cyclic consistency, and combines a semantic anchor mechanism to enhance local alignment and interpretability.
Interpretable Debiasing of Vision-Language Models for Social Fairness
Na Min An (KAIST AI), Hyunjung Shim (KAIST AI)
Explainability and InterpretabilityVision Language ModelAuto EncoderMultimodality
🎯 What it does: Proposes an interpretable VLM/LVLM debiasing framework called DEBIASLENS, which identifies social attribute neurons using a sparse autoencoder and selectively disables them during inference to achieve internal model debiasing;
Interpretable Prompts made Edit-Friendly: Token-to-Token Similarity Reduction in dLLMs for Edit-Friendly Hard Prompt Inversion
Naresh Kumar Devulapally (University at Buffalo, SUNY), Vishnu Suresh Lokhande (Adobe Research)
GenerationPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Using a discrete diffusion language model combined with CLIP-guided hard reverse prompt inversion methods to generate editable and interpretable prompts for text-to-image (T2I) generation.
InterPrior: Scaling Generative Control for Physics-Based Human-Object Interactions
Sirui Xu (University of Illinois Urbana Champaign), Liang-Yan Gui (University of Illinois Urbana Champaign)
GenerationRobotic IntelligenceReinforcement LearningVision-Language-Action ModelAuto EncoderMultimodality
🎯 What it does: This paper proposes InterPrior, a scalable generative controller that enables robots to achieve physically feasible human-object interactions based on sparse goals by performing variational distillation from a large-scale expert with full reference and reinforcement learning fine-tuning on it;
InterRVOS: Interaction-Aware Referring Video Object Segmentation
Woojeong Jin (KAIST AI), Seungryong Kim (KAIST AI)
SegmentationTransformerLarge Language ModelSupervised Fine-TuningVideoTextMultimodality
🎯 What it does: Propose an interactive perception visual segmentation task called InterRVOS, construct a large-scale interactive expression dataset named InterRVOS-127K, and design a ReVIOSa model capable of segmenting actors and targets separately.
Intervention-Aware Multiscale Representation Learning from Imaging Phenomics and Perturbation Transcriptomics
Jiayuan Chen, Ping Zhang (The Ohio State University)
Knowledge DistillationRepresentation LearningDrug DiscoveryTransformerAuto EncoderContrastive LearningImageBiomedical Data
🎯 What it does: Propose an intervention-aware distillation framework (TIDE) that leverages transcriptomic information to guide microscopic image feature learning, achieving mechanistic representation of drug interventions.
Intra-class Distribution-guided Generative Hashing with Neighbor Refinement for Cross-modal Retrieval
Hao Sun (Qufu Normal University), Lei Huang (Ocean University of China)
Data SynthesisRetrievalTransformerContrastive LearningMultimodality
🎯 What it does: Propose an adaptive synthetic sample generation framework based on intra-class distribution (IDGH) to enhance the discriminativeness of hash codes in cross-modal retrieval.
InTrain: Intrinsic Trainability for Zero-Cost Neural Architecture Search
Qinqin Zhou (Fuzhou University), Weiwei Cai (Harbin Medical University)
Neural Architecture SearchBenchmark
🎯 What it does: Proposes InTrain, a zero-cost neural architecture search (NAS) proxy that combines geometric capacity and optimization elasticity to evaluate network trainability.
Intrinsic Concept Extraction Based on Compositional Interpretability
Hanyu Shi (Guangdong University of Technology), Pan Pan (VIPSHOP)
Explainability and InterpretabilityVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed the CI-ICE task and introduced the HyperExpress method to extract composable and interpretable ontological concepts from a single image.
Intrinsic Geometry-Appearance Consistency Optimization for Sparse-View Gaussian Splatting
Kaiqiang Xiong (Peking University), Ronggang Wang (Peking University)
GenerationOptimizationGaussian SplattingImage
🎯 What it does: Proposed the ICO-GS framework, achieving high-quality novel view synthesis under sparse views through geometric regularization and virtual view-based appearance optimization.
Intrinsic Image Fusion for Multi-View 3D Material Reconstruction
Peter Kocsis (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationConvolutional Neural NetworkDiffusion modelImagePoint Cloud
🎯 What it does: Propose the Intrinsic Image Fusion framework by combining single-view material priors with inverse path tracing, achieving high-quality reconstruction of room-scale 3D PBR materials.
IntrinsicWeather: Controllable Weather Editing in Intrinsic Space
Yixin Zhu (Nanjing University), Beibei Wang (Nanjing University)
Image TranslationImage HarmonizationGenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelImage
🎯 What it does: Achieve controllable weather editing in the intrinsic space through inverse rendering and forward rendering; inverse rendering decomposes the input image into three intrinsic maps: material, geometry, and illumination (including weather), while forward rendering re-renders new images based on these intrinsic maps and text-based weather prompts.
IntroSVG: Learning from Rendering Feedback for Text-to-SVG Generation via an Introspective Generator-Critic Framework
Feiyu Wang (Fudan University), Junyu Gao (Institute of Artificial Intelligence (TeleAI), China Telecom)
Image TranslationGenerationTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Proposed the IntroSVG framework, which employs a unified Vision-Language Model to generate SVG code while self-evaluating and iteratively improving through visual feedback.
InvAD: Inversion-based Reconstruction-Free Anomaly Detection with Diffusion Models
Shunsuke Sakai (University of Fukui), Tatsuhito Hasegawa (University of Fukui)
Anomaly DetectionDiffusion modelImageBiomedical Data
🎯 What it does: Propose InvAD, a reconstruction-free anomaly detection method based on diffusion models
InvCoSS: Inversion-driven Continual Self-supervised Learning in Medical Multi-modal Image Pre-training
Zihao Luo (University of Electronic Science and Technology of China), Xiaosong Wang (Shanghai Innovation Institute)
ClassificationSegmentationGenerationData SynthesisKnowledge DistillationRepresentation LearningTransformerAuto EncoderMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyAlzheimer's DiseaseElectronic Health Records
🎯 What it does: Propose the InvCoSS framework to achieve continuous self-supervised learning in medical multimodal image pretraining, utilizing model inversion to generate synthetic images for knowledge preservation without data replay.
InverFill: One-Step Inversion for Enhanced Few-Step Diffusion Inpainting
Duc Vu (Qualcomm AI Research), Anh Tran (Qualcomm AI Research)
RestorationData SynthesisTransformerDiffusion modelAuto EncoderImageTextBenchmark
🎯 What it does: Designed and implemented a single-step inverse network called InvertFill for image restoration, which generates semantically aligned noise initial values to improve the quality of few-step diffusion models in restoration tasks.
Investigating Self-Supervised Representations for Audio-Visual Deepfake Detection
Dragos-Alexandru Boldisor (Bitdefender), Elisabeta Oneata (Bitdefender)
Anomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper systematically evaluates the effectiveness and interpretability of various self-supervised features in audio-visual deepfake detection, constructing a multi-dimensional evaluation framework including linear probing, anomaly detection, spatiotemporal explanation, and complementary analysis.
IP-Adapter Is All You Need: Towards Fine-Tuning-Free Diffusion-Based Talking Face Generation
Hao Wu (Information Engineering University), Jinwei Wang (Nankai University)
GenerationData SynthesisDiffusion modelOptical FlowVideo
🎯 What it does: This paper proposes a diffusion model framework called FreeTalkDiff, which combines IP-Adapter with Stable Diffusion and introduces modules such as Structurist, Structure Controller, and Noise Sensor to achieve high-quality, precise lip-sync video generation for speakers;
IPR-1: Interactive Physical Reasoner
Mingyu Zhang (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Jiao Tong University)
TransformerLarge Language ModelReinforcement LearningVision Language ModelAuto EncoderWorld ModelOptical FlowImageVideoTextMultimodalityBenchmarkPhysics Related
🎯 What it does: Explored a general method for acquiring physical and causal reasoning through interactive experience, and proposed the IPR framework.
IR-HGP: Physically-Aware Gaussian Inverse Rendering for High-Illumination Scenes via Generative Priors
Qingan Zhang (Sun Yat Sen University), Chengying Gao (Sun Yat Sen University)
GenerationOptimizationDiffusion modelScore-based ModelGaussian SplattingImageMeshPhysics Related
🎯 What it does: Propose a physics-aware inverse rendering framework IR-HGP based on 3D Gaussian Splatting, which can achieve material separation, illumination reconstruction, and real-time rendering under high illumination scenes;
Iris: Bringing Real-World Priors into Diffusion Model for Monocular Depth Estimation
Xinhao Cai (Nanjing University of Science and Technology), Wenguan Wang (Zhejiang University)
Depth EstimationAutonomous DrivingDiffusion modelImage
🎯 What it does: Propose Iris, a two-stage Priors-to-Geometry structure based on deterministic diffusion models, which first uses real images to guide low-frequency layout and then refines high-frequency geometry with synthetic data, achieving unsupervised monocular depth estimation.
Iris: Integrating Language into Diffusion-based Monocular Depth Estimation
Ziyao Zeng (Yale University), Alex Wong (Yale University)
Depth EstimationConvolutional Neural NetworkVision Language ModelDiffusion modelAuto EncoderImageText
🎯 What it does: This study incorporates textual information as a condition into diffusion models for monocular depth estimation, providing additional semantic constraints in visually ambiguous regions.
IrisFP: Adversarial-Example-based Model Fingerprinting with Enhanced Uniqueness and Robustness
Ziye Geng (University of Houston), Changqing Luo (University of Houston)
Adversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: Designed a model fingerprinting method called IrisFP based on adversarial examples to enhance the uniqueness and robustness of deep network fingerprints.
Is Bin Generation Indispensable? A Bin-Generation-Free Dataset Quantization via Semantic Perspective
Maijie Deng (Huazhong University of Science and Technology), Chenru Ma (Huazhong University of Science and Technology)
CompressionConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Propose a bin-free dataset quantization framework called BGFDQ, which utilizes KNN neighbor identification and neighbor-aware core subset selection, combined with adaptive semantic transfer patch dropping, to achieve efficient and scalable image dataset compression.
Is Parameter Isolation Better for Prompt-Based Continual Learning?
Jiangyang Li (Xi'an Jiaotong University), Yihong Gong (Xi'an Jiaotong University)
Meta LearningTransformerPrompt EngineeringMixture of ExpertsImage
🎯 What it does: Construct a global prompt pool, employ task-aware gating routing for dynamic sparse activation of a small number of prompts, and introduce a history-aware modulator to balance prompt usage and gradient updates, thereby enhancing generalization and parameter efficiency in continual learning.
Is the Modality Gap a Bug or a Feature? A Robustness Perspective
Rhea Chowers (Hebrew University), Yair Weiss (Hebrew University)
ClassificationRetrievalVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper investigates the modality gap in multi-modal contrastive learning models, demonstrating that it arises from initialization and gradient dynamics, and showing the impact of this gap on model robustness.
Is your VLM Sky-Ready? A Comprehensive Spatial Intelligence Benchmark for UAV Navigation
Lingfeng Zhang (Tsinghua University), Wenbo Ding (Tsinghua University)
Autonomous DrivingTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityPoint CloudBenchmark
🎯 What it does: Constructed the SpatialSky-Bench benchmark from a UAV perspective, and trained a specialized Sky-VLM using the SpatialSky-Dataset with 1M samples to evaluate and enhance UAV spatial intelligence tasks.
iSHIFT: Lightweight Slow-Fast GUI Agent with Adaptive Perception
Sarthak Mehrotra (Indian Institute of Technology Bombay), Vineeth N. Balasubramanian
TransformerLarge Language ModelAgentic AIVision Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: Propose a lightweight GUI agent called iSHIFT, which can adaptively switch between slow (high-precision) and fast (quick) inference modes based on task requirements, and dynamically activate the visual perception module when needed.
IsoCLIP: Decomposing CLIP Projectors for Efficient Intra-modal Alignment
Simone Magistri (University of Florence), Andrew D. Bagdanov (University of Florence)
ClassificationRetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: The paper analyzes the intra-modal alignment insufficiency caused by the CLIP projector, proposing IsoCLIP which enhances intra-modal similarity through spectral decomposition and projection into an isotropic subspace in a training-free manner.
iSplat: Iterative Learning for Fine-Grained Gaussian Splatting
Haifeng Wu (University of Electronic Science and Technology of China), Wen Li (University of Electronic Science and Technology of China)
Depth EstimationOptimizationRecurrent Neural NetworkGaussian SplattingImage
🎯 What it does: Propose iSplat, an iterative feed-forward 3D Gaussian splatting framework based on GRU loops, which simultaneously optimizes geometry and appearance through multi-step self-correction.
It Takes Two: A Duet of Periodicity and Directionality for Burst Flicker Removal
Lishen Qu (Nankai International Advanced Research Institute), Jufeng Yang (Nankai International Advanced Research Institute)
RestorationTransformerImage
🎯 What it does: This paper proposes a Transformer-based architecture named Flickerformer for eliminating flicker artifacts generated during shooting.
It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models
Anne Harrington (UC Berkeley), Alexei A. Efros (UC Berkeley)
GenerationDiffusion modelImage
🎯 What it does: This paper proposes a method that directly optimizes the initial noise via gradient optimization on a pre-trained diffusion model to eliminate mode collapse and enhance the diversity of generated images.
Iterative Closed-Loop Motion Synthesis for Scaling the Capabilities of Humanoid Control
Weisheng Xu (Hong Kong University of Science and Technology), Renjing Xu (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerReinforcement LearningVision Language ModelDiffusion modelMultimodalitySequentialChain-of-Thought
🎯 What it does: Propose a closed-loop automated framework for motion data generation and iterative training (CLAIMS), which generates high-difficulty, semantically labeled action data using multi-domain professional action templates and a language diffusion model, and repeatedly trains a physical controller to enhance its performance in extreme actions.
IVAAN: Instance-level Vision-Language Alignment via Attribute-Guided Text Prompts Generation for Nuclei Analysis
Jaehoon Jeong (DGIST), Sang Hyun Park (DGIST)
ClassificationSegmentationTransformerPrompt EngineeringVision Language ModelContrastive LearningBiomedical Data
🎯 What it does: Proposed an instance-level vision-language framework that automatically generates attribute-guided pseudo-text prompts and aligns nuclear visual features with corresponding text through contrastive learning, achieving nuclear segmentation and classification.
Jailbreaking Vision-Language Models via Dissonance-Guided Suffix Optimization and Image-Phrase Injection
Jiacheng Pi (University of Science and Technology of China), Wenjie Ruan (University of Science and Technology of China)
Adversarial AttackPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Propose an attack framework that simultaneously targets text suffixes and image phrases to achieve efficient jailbreak in visual language models.
JANUS: A Lightweight Framework for Jailbreaking Text-to-Image Models via Distribution Optimization
Haolun Zheng (Zhejiang University), Kui Ren (Zhejiang University)
GenerationAdversarial AttackReinforcement LearningText
🎯 What it does: Proposed a lightweight text-to-image model jailbreak framework called JANUS.
JarvisEvo: Towards a Self-Evolving Photo Editing Agent with Synergistic Editor-Evaluator Optimization
Yunlong Lin (Xiamen University), Qinglin Lu (Xiamen University)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: JarvisEvo is a unified photo editing agent capable of performing multi-modal chain-of-thought reasoning, tool calling, result evaluation, and self-reflection during the editing process, achieving end-to-end self-improvement.
Joint Learning of General and Diverse Patterns with Mixture of Memory Experts for Weakly-Supervised Video Anomaly Detection
Bo Sun (University Of Chinese Academy Of Sciences), Yaowei Wang (Peking University)
Anomaly DetectionTransformerLarge Language ModelMixture of ExpertsVision Language ModelVideo
🎯 What it does: Designed and implemented a sparse mixture of memory experts (MoME) framework that jointly learns general and diverse anomaly patterns through internal expert memory and shared external memory, achieving weakly supervised video anomaly detection.
Joint Spectral Image Reconstruction and Semantic Segmentation with Cooperative Unfolding
Zijun He (Zhejiang University), Xin Yuan (Westlake University)
RestorationSegmentationOptimizationTransformerImage
🎯 What it does: Propose a deep split network named CRSDUN for joint spectral image reconstruction and semantic segmentation in CASSI measurements
Joint-Aligned Latent Action: Towards Scalable VLA Pretraining in the Wild
Hao Luo (Peking University), Zongqing Lu (Peking University)
Representation LearningRobotic IntelligenceTransformerVision-Language-Action ModelFlow-based ModelVideoTextMultimodality
🎯 What it does: Propose a new Vision-Language-Action (VLA) pre-training framework called JALA, which constructs a predictable and information-rich potential action space by jointly aligning the intermediate hidden states of the Transformer with potential action vectors derived from inverse dynamics. This framework is trained together with a large-scale mixed human video dataset, UniHand-Mix, to enhance VLA generalization and robotic manipulation performance in real-world environments.
JoPPO: Hierarchical Photography Assessment via Contrastive Joint Conditional Probabilistic Reinforcement Learning
Yifan Yang (OPPO AI Center), Dan Meng (OPPO AI Center)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImage
🎯 What it does: Proposed a multi-attribute comprehensive aesthetic evaluation method called JoPPO based on vision-language models. It first injects compositional priors through SFT, then achieves consistent optimization of attributes and overall scores using joint conditional probability reinforcement learning, capable of outputting attribute scores, overall scores, and natural language explanations.
JRM: Joint Reconstruction Model for Multiple Objects without Alignment
Qirui Wu (Meta Reality Labs Research), Henry Howard-Jenkins (Meta Reality Labs Research)
GenerationTransformerDiffusion modelFlow-based ModelImageTextMultimodalityPoint Cloud
🎯 What it does: Propose the Joint Reconstruction Model (JRM), which achieves joint reconstruction of multiple objects by implicitly aggregating misaligned observations in the latent space of a 3D generative model.
JUMP-Hand: Learning Joint-wise Uncertainty to Gate Mixture of View Experts for Multi-View 3D Hand Reconstruction
Haohong Kuang (Huazhong University of Science and Technology), Joey Tianyi Zhou (Huazhong University of Science and Technology)
Pose EstimationConvolutional Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: Proposes a multi-view 3D hand reconstruction method called JUMP-Hand, which utilizes a joint uncertainty gating mechanism to integrate expert information from different views.
Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers
Wenhao Sun (University Of Electronic Science And Technology Of China), Zhaoqiang Liu (University Of Electronic Science And Technology Of China)
GenerationComputational EfficiencyTransformerFlow-based ModelAuto EncoderImageOrdinary Differential Equation
🎯 What it does: Proposes a Just-in-Time (JiT) training-agnostic spatial acceleration framework that significantly reduces computational cost while preserving generation quality by dynamically selecting sparse anchors during the Diffusion Transformer generation process.
Kaleidoscopic Scintillation Event Imaging
Alex Bocchieri (University of Wisconsin Madison), Andreas Velten (University of Wisconsin Madison)
OptimizationImagePhysics Related
🎯 What it does: Designed a kaleidoscope-shaped scintillator with a mirror structure, capturing scintillation events and their mirror images using a single-photon avalanche diode (SPAD) camera, and achieving three-dimensional event localization under low-photon conditions through a Gaussian Mixture Model combined with the EM algorithm.
KaLOS finds Consensus: A Meta-Algorithm for Evaluating Inter-Annotator Agreement in Complex Vision Tasks
David Tschirschwitz (Bauhaus-Universität Weimar), Volker Rodehorst (Bauhaus-Universität Weimar)
Explainability and InterpretabilityData-Centric LearningImageBiomedical Data
🎯 What it does: This paper proposes the K LOS Meta-Algorithm for unified, interpretable evaluation of annotation consistency across multiple annotators in complex visual tasks (e.g., object detection, instance segmentation).
KAMP: Knowledge-Anchored Multimodal Pretraining Framework for Medical Image Representation
Feiyu Huang (Hong Kong University Of Science And Technology (Guangzhou)), Lei Chen (Hong Kong University Of Science And Technology)
Representation LearningTransformerLarge Language ModelReinforcement LearningContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Developed a multimodal pre-training framework KAMP based on generating personalized knowledge from large language models (LLMs) as semantic anchors to enhance medical image representation learning.
KASALv2: Fully Automatic 3D Rotational Symmetry Classification and Axis Localization
Mengxin Zhang (Southeast University), Yijun Zhou (Southeast University)
ClassificationPose EstimationPoint Cloud
🎯 What it does: Propose KASALv2, a fully automated 3D rotational symmetry analysis framework that can classify symmetry types, identify rotation orders, and locate complete axes without geometric references.
Keep It Frozen: Domain-Routed Conditional Residual Modulation for Multi-Domain Vision Transformers
Ufaq Khan (Mohamed bin Zayed University of Artificial Intelligence), Muhammad Haris Khan (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationSegmentationDomain AdaptationMeta LearningTransformerImageBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: Propose DCRM-ViT, which utilizes a frozen Vision Transformer and incorporates a residual modulation block based on a domain router and parameter synthesis network, achieving multi-domain adaptation without updating during inference.
Keep it SymPL: Symbolic Projective Layout for Allocentric Spatial Reasoning in Vision-Language Models
Jaeyun Jang (Kyung Hee University), Hyoseok Hwang (Kyung Hee University)
Object DetectionPose EstimationDepth EstimationVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: This paper addresses the issue of visual-language models (VLMs) performing poorly in spatial reasoning from an allocentric perspective. It proposes the SymPL framework, which transforms allocentric problems into symbolic-layout problems that VLMs are naturally adept at, thereby achieving more accurate spatial reasoning.
KLIP: Localized Distribution Shift Detection via KL-Divergence with Diffusion Priors in Inverse Problems
Alireza Kheirandish (Georgia Institute of Technology), Sara Fridovich-Keil (Georgia Institute of Technology)
Anomaly DetectionDiffusion modelImageBiomedical DataComputed Tomography
🎯 What it does: Propose the KLIP metric based on the prior and posterior KL divergence of diffusion models for detecting local distribution shifts in inverse problems.
KnowVal: A Knowledge-Augmented and Value-Guided Autonomous Driving System
Zhongyu Xia (Peking University), Ming-Hsuan Yang (University of California Merced)
Autonomous DrivingOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVision-Language-Action ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose KnowVal, an end-to-end autonomous driving system that integrates open-world visual perception, knowledge retrieval, and value assessment;
Kontinuous Kontext: Continuous Strength Control for Instruction-based Image Editing
Rishubh Parihar (Snap Research), Kuan-Chieh Jackson Wang (Snap Research)
Image TranslationVision Language ModelDiffusion modelImageText
🎯 What it does: Propose Kontinuous Kontext, a unified model that introduces continuous intensity control in instruction-driven image editing;
KV-Tracker: Real-Time Pose Tracking with Transformers
Marwan Taher (Imperial College London), Andrew Davison (Imperial College London)
Object TrackingPose EstimationTransformerSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: This paper proposes KV-Tracker, which realizes real-time 6-DoF pose tracking and online reconstruction by leveraging the key-value cache of multi-view Transformer (π³).
KVSmooth: Mitigating Hallucination in Multi-modal Large Language Models through Key-Value Smoothing
Siyu Jiang (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
Explainability and InterpretabilityTransformerLarge Language ModelMultimodalityBenchmark
🎯 What it does: Proposed a training-agnostic, plug-and-play attention entropy-guided key-value smoothing method called KVSmooth to reduce hallucinations in multi-modal large language models.
L3DR: 3D-aware LiDAR Diffusion and Rectification
Quan Liu (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
Autonomous DrivingConvolutional Neural NetworkTransformerDiffusion modelPoint Cloud
🎯 What it does: Proposes L3DR, a LiDAR generation framework that combines 2D range view diffusion with a 3D residual regression network to address defects such as depth leakage and wavefronts, enhancing the geometric realism of point clouds.
LA-Pose: Latent Action Pretraining Meets Pose Estimation
Zhengqing Wang (Wayve), Yasutaka Furukawa (Wayve)
Pose EstimationAutonomous DrivingTransformerSupervised Fine-TuningVideo
🎯 What it does: Proposes LA-Pose, a two-stage framework that combines inverse dynamics pre-training on large-scale unlabeled driving videos with camera pose estimation, achieving efficient and transferable pose prediction through self-supervised latent action representations.
Label What Matters: Modality-Balanced and Difficulty-Aware Multimodal Active Learning
Yuqiao Zeng (Beijing Jiaotong University), Hui Yu (University of Glasgow)
Data-Centric LearningReinforcement LearningImageVideoTextMultimodalityAudio
🎯 What it does: Propose a reinforcement learning-based multi-modal active learning framework RL-MBA, which can adaptively adjust modality weights and select appropriately difficult samples in each round of active learning.
Label-Free Cross-Task LoRA Merging with Null-Space Compression
Wonyoung Lee (KAIST), Kuk-Jin Yoon (KAIST)
ClassificationGenerationTransformerImageTextMultimodality
🎯 What it does: This paper proposes a model merging method for LoRA adapters called Null-Space Compression (NSC) Merging, which can merge multiple single-task LoRA models into a unified model without relying on labels, while simultaneously handling classification, regression, and sequence generation tasks.
LacTokGen: Latent Consistency Tokenizer for 1024-pixel Image Generation by 256 Tokens
Qingsong Xie (OPPO AI Center), Haonan Lu (OPPO AI Center)
GenerationTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes the Latent Consistency Tokenizer (LacTok) and its autoregressive generation model LacTokGen, which can efficiently reconstruct and generate 1024×1024 pixel images using only 256 discrete tokens;
LaDy: Lagrangian-Dynamic Informed Network for Skeleton-based Action Segmentation via Spatial-Temporal Modulation
Haoyu Ji (Harbin Institute of Technology, Shenzhen Southeast University), Honghai Liu (Harbin Institute of Technology, Shenzhen Southeast University)
SegmentationPose EstimationConvolutional Neural NetworkGraph Neural NetworkTransformerGraphSequential
🎯 What it does: Proposed the LaDy framework, which integrates Lagrangian dynamics with skeleton action segmentation to construct a physics-driven spatiotemporal modulation network, achieving frame-level action segmentation for skeleton sequences.
Lafite: A Generative Latent Field for 3D Native Texturing
Chia-Hao Chen (Tsinghua University), Song-Hai Zhang (Tsinghua University)
GenerationRectified FlowAuto EncoderPoint CloudMesh
🎯 What it does: Propose Lafite, a 3D native texture generation framework based on sparse latent color fields;
LagerNVS: Latent Geometry for Fully Neural Real-time Novel View Synthesis
Stanislaw Szymanowicz (University of Oxford), Andrea Vedaldi (University of Oxford)
GenerationTransformerDiffusion modelImage
🎯 What it does: Propose a Latent Geometry Encoder-Decoder network called LagerNVS, based on 3D perceptual features, for real-time Novel View Synthesis without explicit 3D reconstruction;