ICCV 2025 Papers — Page 13
IEEE/CVF International Conference on Computer Vision · 2701 papers
Importance-Based Token Merging for Efficient Image and Video Generation
Haoyu Wu (Stony Brook University), Dimitris Samaras (University of North Carolina at Charlotte)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideo
🎯 What it does: This paper proposes a token merging method based on importance scoring to reduce computational load during inference in diffusion models while maintaining image quality.
Improved Noise Schedule for Diffusion Training
Tiankai Hang (Southeast University), Baining Guo (Microsoft Research Asia)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: Designed and evaluated a new noise scheduling strategy using importance sampling to focus on moderate noise levels, aiming to enhance the training efficiency and effectiveness of diffusion models.
Improving Large Vision and Language Models by Learning from a Panel of Peers
Jefferson Hernandez (Rice University), Kushal Kafle (Adobe Research)
GenerationOptimizationTransformerReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the Panel-of-Peers (PoP) framework, which allows a group of visually language models with similar capabilities to generate answers for each other and evaluate one another, thereby achieving self-improvement without human annotation.
Improving Multimodal Learning via Imbalanced Learning
Shicai Wei (University of Electronic Science and Technology of China), Yang Luo (University of Electronic Science and Technology of China)
Representation LearningVideoMultimodalityAudio
🎯 What it does: A strategy named Asymmetric Representation Learning (ARL) is proposed to improve multimodal learning;
Improving Noise Efficiency in Privacy-preserving Dataset Distillation
Runkai Zheng (Carnegie Mellon University), Fernando De La Torre
Safty and PrivacyComputational EfficiencyKnowledge DistillationImage
🎯 What it does: In privacy-preserving dataset distillation, two main modules, DOS and SER, are proposed to enhance noise efficiency.
Improving Rectified Flow with Boundary Conditions
Xixi Hu (University of Texas at Austin), Qiang Liu (University of Texas at Austin)
GenerationData SynthesisTransformerFlow-based ModelRectified FlowImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes the Boundary-enforced Rectified Flow Model (Boundary RF Model), which enhances the generation quality and sampling stability of Rectified Flow by explicitly enforcing boundary conditions in the parameterization of the velocity field.
Improving SAM for Camouflaged Object Detection via Dual Stream Adapters
Jiaming Liu (Shanghai Jiao Tong University), Guihai Chen (Shanghai Jiao Tong University)
Object DetectionKnowledge DistillationTransformerImage
🎯 What it does: A dual-stream adapter SAM-DSA based on the Segment Anything Model (SAM) is proposed to improve camouflage object detection under RGB-D views.
Incremental Few-Shot Semantic Segmentation via Multi-Level Switchable Visual Prompts
Maoxian Wan (Beihang University), Zhong Zhou (Beihang University)
SegmentationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes an incremental few-shot semantic segmentation framework based on Multi-layer Switchable Visual Prompts (MSVP), which utilizes visual language models and text semantics to separate foreground from background. It achieves learning of new categories while retaining memory of old categories through a three-layer knowledge base consisting of task-continuous prompts, stage-specific prompts, and region-unique prompts.
Inference-Time Diffusion Model Distillation
Geon Yeong Park (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
GenerationKnowledge DistillationDiffusion modelScore-based ModelImage
🎯 What it does: This paper proposes a method for distilling diffusion models during the inference phase, utilizing the teacher model to guide the student model early in the sampling process, thereby enhancing the visual quality and text alignment of the student model.
InfGen: A Resolution-Agnostic Paradigm for Scalable Image Synthesis
Tao Han (Hong Kong University of Science and Technology), Lei Bai (Shanghai Artificial Intelligence Laboratory)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: We propose InfGen, a generator that can decode fixed-size latent vectors at any resolution, enabling existing VAE-based diffusion models to achieve high-resolution generation while significantly reducing inference time.
InfiniCube: Unbounded and Controllable Dynamic 3D Driving Scene Generation with World-Guided Video Models
Yifan Lu (NVIDIA), Jiahui Huang (NVIDIA)
GenerationData SynthesisAutonomous DrivingTransformerDiffusion modelVideoPoint Cloud
🎯 What it does: This paper proposes a complete pipeline for generating large-scale, controllable, dynamic 3D driving scenes from HD maps, vehicle bounding boxes, and textual prompts, ultimately resulting in high-fidelity 3D Gaussian fields (3DGS) and corresponding voxel worlds that can be reconstructed in a short time for simulation and rendering purposes.
InfiniDreamer: Arbitrarily Long Human Motion Generation via Segment Score Distillation
Wenjie Zhuo, Hehe Fan
GenerationData SynthesisKnowledge DistillationDiffusion modelTextMultimodalitySequential
🎯 What it does: The InfiniDreamer framework is proposed to generate human action sequences of arbitrary length that are continuous and smooth from a series of text prompts.
InfiniteYou: Flexible Photo Recrafting While Preserving Your Identity
Liming Jiang (ByteDance Intelligent Creation), Xin Lu (ByteDance Intelligent Creation)
GenerationData SynthesisTransformerSupervised Fine-TuningDiffusion modelImageText
🎯 What it does: A framework for identity-preserving image generation called InfiniteYou (InfU) based on Diffusion Transformer (FLUX) is proposed, capable of recreating specified character photos according to any text description while maintaining facial identity.
InfoBridge: Balanced Multimodal Integration through Conditional Dependency Modeling
Chenxin Li (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)
ClassificationRecognitionContrastive LearningMultimodality
🎯 What it does: This paper proposes the InfoBridge framework, which addresses the over-fusion problem in multimodal learning by utilizing conditional mutual information maximization to achieve a unified fusion of cross-modal information sharing and modal feature protection.
Information Density Principle for MLLM Benchmarks
Chunyi Li, Guangtao Zhai
TransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: This paper proposes the principle of information density for evaluating multimodal large language models (MLLMs), defining four dimensions (error rate, difficulty, redundancy, diversity) to quantify the quality of benchmarks, and provides an executable evaluation pipeline.
Information-Bottleneck Driven Binary Neural Network for Change Detection
Kaijie Yin (Singapore Management University), Hui Kong (Singapore Management University)
RecognitionAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A binary change detection network named BiCD is proposed, specifically designed for change recognition in street view and remote sensing scenes.
Inpaint4Drag: Repurposing Inpainting Models for Drag-Based Image Editing via Bidirectional Warping
Jingyi Lu (University of Hong Kong), Kai Han (University of Hong Kong)
Image TranslationRestorationDiffusion modelImageBenchmark
🎯 What it does: This paper proposes the Inpaint4Drag framework, which splits drag-and-drop image editing into two steps: pixel-level bidirectional warping and image filling, achieving real-time preview and high-quality editing.
INS-MMBench: A Comprehensive Benchmark for Evaluating LVLMs' Performance in Insurance
Chenwei Lin (Fudan University), Jiebo Luo (University of Rochester)
ClassificationRecognitionData-Centric LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: INS-MMBench has been constructed, a hierarchical multimodal benchmark covering four types of insurance (auto insurance, property insurance, health insurance, and agricultural insurance), which includes 12,252 images, 10,372 question-and-answer pairs, covering 22 basic tasks, 12 meta-tasks, and 5 scenario tasks, and evaluates 11 large audiovisual language models.
InsideOut: Integrated RGB-Radiative Gaussian Splatting for Comprehensive 3D Object Representation
Jungmin Lee (Chung Ang University), Jongwon Choi (Chung Ang University)
GenerationData SynthesisGaussian SplattingImageMultimodalityComputed Tomography
🎯 What it does: The InsideOut framework is proposed, which fuses RGB images with X-ray images into 3D Gaussian splatting (3DGS) to generate a complete 3D voxel model that contains both appearance texture and internal structure.
InstaDrive: Instance-Aware Driving World Models for Realistic and Consistent Video Generation
Zhuoran Yang (University of Science and Technology of China), Yanyong Zhang (University of Science and Technology of China)
GenerationData SynthesisAutonomous DrivingDiffusion modelAuto EncoderWorld ModelVideo
🎯 What it does: Introducing InstaDrive - a video generation world model for driving scenarios, achieving instance-level temporal consistency and spatial geometric accuracy through instance flow guidance and spatial geometric alignment.
Instance-Level Video Depth in Groups Beyond Occlusions
Yuan Liang (South China University of Technology), Shengfeng He (Singapore Management University)
SegmentationDepth EstimationTransformerVideo
🎯 What it does: This paper presents a novel video depth dataset GID for dynamic multi-object scenes and designs a two-stage instance-aware depth estimation framework called InstanceDepth. The framework first obtains a rough scene structure through global depth initialization, and then utilizes instance segmentation, shape priors, and spatial relationships for instance-level depth refinement to address the depth discontinuity caused by occlusion.
Instant GaussianImage: A Generalizable and Self-Adaptive Image Representation via 2D Gaussian Splatting
Zhaojie Zeng (Huazhong University of Science and Technology), Lili Ju (University of South Carolina)
Representation LearningConvolutional Neural NetworkGaussian SplattingImage
🎯 What it does: This paper presents Instant-GI, a general adaptive image representation framework based on 2D Gaussian splatting.
InstantEdit: Text-Guided Few-Step Image Editing with Piecewise Rectified Flow
Yiming Gong (University of Illinois at Urbana-Champaign), Minjia Zhang (University of Illinois at Urbana-Champaign)
Image TranslationGenerationDiffusion modelRectified FlowImageBenchmark
🎯 What it does: A real-time text-guided image editing method called InstantEdit based on RectifiedFlow is proposed, capable of completing high-quality edits in just 8 sampling steps.
InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes
Zesong Yang (Zhejiang University), Hujun Bao (Zhejiang University)
SegmentationGenerationDiffusion modelContrastive LearningGaussian SplattingPoint CloudMesh
🎯 What it does: The InstaScene framework is proposed, which can accurately decompose arbitrary instances from cluttered scenes and complete their geometric and appearance reconstruction.
INSTINCT: Instance-Level Interaction Architecture for Query-Based Collaborative Perception
Yunjiang Xu (Soochow University), Benyuan Yang (Soochow University)
Object DetectionAutonomous DrivingTransformerGaussian SplattingPoint Cloud
🎯 What it does: The INSTINCT framework is designed to achieve instance-level interaction and collaborative perception in a LiDAR-V2X environment, significantly reducing communication bandwidth while improving detection accuracy.
Instruction-based Image Editing with Planning, Reasoning, and Generation
Liya Ji (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageMultimodalityChain-of-Thought
🎯 What it does: A multi-modal chain thinking editing framework is proposed, which can achieve complex image editing through natural language instructions.
Instruction-Grounded Visual Projectors for Continual Learning of Generative Vision-Language Models
Hyundong Jin (Chung-Ang University), Eunwoo Kim (Chung-Ang University)
ClassificationGenerationTransformerMixture of ExpertsVision Language ModelImageText
🎯 What it does: A mixed model of visual projectors based on instructions (MVP) is proposed, which dynamically selects visual projection experts according to text instructions during continual learning, thereby retaining knowledge of old tasks while quickly adapting to new tasks.
Instruction-Oriented Preference Alignment for Enhancing Multi-Modal Comprehension Capability of MLLMs
Zitian Wang (Beihang University), Si Liu (Beihang University)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodality
🎯 What it does: The Instruction-Oriented Preference Alignment (IPA) framework is proposed, which enhances the instruction-following and overall understanding capabilities of multimodal large language models (MLLMs) through automated preference construction and advanced preference collection.
InstructSeg: Unifying Instructed Visual Segmentation with Multi-modal Large Language Models
Cong Wei (Tsinghua University), Yujiu Yang (Tsinghua University)
SegmentationTransformerLarge Language ModelSupervised Fine-TuningImageVideoTextMultimodality
🎯 What it does: This paper proposes a unified multimodal large language model framework called InstructSeg, which can simultaneously perform four types of text-guided segmentation tasks (RES, ReasonSeg, R-VOS, ReasonVOS) in the image and video domains.
InsViE-1M: Effective Instruction-based Video Editing with Elaborate Dataset Construction
Yuhui Wu (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
GenerationData SynthesisLarge Language ModelVision Language ModelDiffusion modelOptical FlowImageVideoText
🎯 What it does: A high-quality InsViE‑1M instruction video editing dataset (a total of 1M triplets) was constructed, and the InsViE video editing model was trained based on this;
Integrating Biological Knowledge for Robust Microscopy Image Profiling on De Novo Cell Lines
Jiayuan Chen (Ohio State University), Ping Zhang (Ohio State University)
ClassificationSegmentationTransformerContrastive LearningImageBiomedical Data
🎯 What it does: In cellular microscopy image analysis, external biological knowledge is introduced to construct a perturbation relationship graph, combined with cell line transcriptome embeddings, to enhance perturbation prediction performance on unseen cell lines.
Integrating Task-Specific and Universal Adapters for Pre-Trained Model-based Class-Incremental Learning
Yan Wang (Nanjing University), Han-Jia Ye (Nanjing University)
ClassificationRecognitionTransformerImage
🎯 What it does: A class-incremental learning (CIL) framework is proposed for pre-trained models (PTM), which involves training task-specific adapters and merging them into a universal adapter. During inference, the most suitable task adapter is selected using entropy, and then fused with the output of the universal adapter.
Integrating Visual Interpretation and Linguistic Reasoning for Geometric Problem Solving
Zixian Guo (Hong Kong Polytechnic University), Wangmeng Zuo (Harbin Institute of Technology)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: A decoupled visual language reasoning framework is proposed, where image parsing is handled by a vision-specific model and logical reasoning is managed by a language model, utilizing joint rewards for collaborative optimization.
INTER: Mitigating Hallucination in Large Vision-Language Models by Interaction Guidance Sampling
Xin Dong (University of Chinese Academy of Sciences), Bo Zheng (Taobao & Tmall Group of Alibaba)
GenerationData SynthesisTransformerVision Language ModelImageMultimodality
🎯 What it does: Proposes INTER, a training-agnostic interactive guided sampling method that significantly reduces hallucination outputs in large visual language models.
Inter2Former: Dynamic Hybrid Attention for Efficient High-Precision Interactive Segmentation
You Huang (Xiamen University), Rongrong Ji (Xiamen University)
SegmentationComputational EfficiencyTransformerMixture of ExpertsImage
🎯 What it does: Proposes Inter2Former, which achieves efficient inference for high-precision interactive segmentation through dynamic computation allocation.
InteractAvatar: Modeling Hand-Face Interaction in Photorealistic Avatars with Deformable Gaussians
Kefan Chen (Brown University), Aayush Prakash (Meta Reality Labs)
GenerationPose EstimationGaussian SplattingVideo
🎯 What it does: This paper presents InteractAvatar, a hybrid mesh-Gaussian representation based on 3D Gaussian Splatting, capable of accurately reconstructing dynamic interactions between hands and faces in multi-view videos, supporting pose-driven hand details, shadows, and geometric and appearance deformations of hand-face interactions.
Interaction-Merged Motion Planning: Effectively Leveraging Diverse Motion Datasets for Robust Planning
Giwon Lee (KAIST), Kuk-Jin Yoon (DGIST)
Domain AdaptationRobotic IntelligenceMultimodality
🎯 What it does: This paper proposes an interactive motion planning framework IMMP based on model merging, designed to achieve robust planning using multi-source trajectory data in the target domain.
InterGSEdit: Interactive 3D Gaussian Splatting Editing with 3D Geometry-Consistent Attention Prior
Minghao Wen (Nanjing University of Aeronautics and Astronautics), Dong Liang (Nanjing University of Aeronautics and Astronautics)
GenerationData SynthesisDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: An interactive 3D Gaussian Splatting editing framework called InterGSEdit is proposed, which constructs a 3D geometric consistency attention prior (GAP 3D) through user-selected key viewpoints and CLIP semantic consistency selection (CSCS), and achieves multi-view consistent and detail-rich 3D editing using an Adaptive Fusion Network (AFN) in a diffusion model.
Intermediate Connectors and Geometric Priors for Language-Guided Affordance Segmentation on Unseen Object Categories
Yicong Li (National University of Singapore), Angela Yao (National University of Singapore)
Object DetectionSegmentationTransformerVision Language ModelTextPoint Cloud
🎯 What it does: Proposes the GLANCE framework, which enhances language-guided 3D usability segmentation through a cross-modal connector and a geometry-aware query generator, particularly improving generalization on unseen object categories.
Interpretable point cloud classification using multiple instance learning
Matt De Vries (Sentinal4D), Chris Bakal (Institute of Cancer Research)
ClassificationExplainability and InterpretabilityTransformerPoint CloudBiomedical Data
🎯 What it does: This paper proposes a multi-instance learning-based explainable 3D point cloud classification framework called POINTMIL, which provides explanations at the point level and improves classification performance.
Interpretable Zero-Shot Learning with Locally-Aligned Vision-Language Model
Shiming Chen (Mohamed bin Zayed University of Artificial Intelligence), Fahad Shahbaz Khan (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationDomain AdaptationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: A local alignment method based on a pre-trained vision-language model, LaZSL, is proposed, which matches local features of images with attributes using optimal transport to achieve interpretable zero-shot classification.
InterSyn: Interleaved Learning for Dynamic Motion Synthesis in the Wild
Yiyi Ma (Tsinghua University), Xuelong Li (Institute of Artificial Intelligence China Telecom)
GenerationData SynthesisTransformerDiffusion modelMultimodality
🎯 What it does: The InterSyn framework is proposed, which can learn from integrated single and multi-person actions to generate natural and smooth interactive movements.
Intervening in Black Box: Concept Bottleneck Model for Enhancing Human Neural Network Mutual Understanding
Nuoye Xiong (Xidian University), Liang Zhang (Xidian University)
ClassificationExplainability and InterpretabilityKnowledge DistillationConvolutional Neural NetworkTransformerLarge Language ModelImage
🎯 What it does: A black-box model intervention framework based on the concept bottleneck, CBM-HNMU, is proposed. It automatically extracts visual and natural language concepts, determines gradients, and prunes harmful concepts, then distills the improved knowledge back to the original model to enhance interpretability and classification accuracy.
Intra-modal and Cross-modal Synchronization for Audio-visual Deepfake Detection and Temporal Localization
Ashutosh Anshul (Nanyang Technological University), Eng Siong Chng (Nanyang Technological University)
RecognitionAnomaly DetectionTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: This paper proposes a two-stage multimodal deepfake detection and temporal localization framework, which first captures the temporal consistency in videos through self-supervised audio-visual synchronization learning, and then utilizes pretrained features for deepfake detection and localization of local forged segments.
Intra-view and Inter-view Correlation Guided Multi-view Novel Class Discovery
Xinhang Wan (National University of Defense Technology), Kunlun He
MultimodalityBiomedical Data
🎯 What it does: A new multi-view novel class discovery framework IICMVNCD based on the correlation within and between views is proposed, addressing the issues of single view and pseudo-labels.
IntrinsicControlNet: Cross-distribution Image Generation with Real and Unreal
Jiayuan Lu (Zhejiang University), Yuchi Huo (Zhejiang Lab)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes IntrinsicControlNet, a cross-distribution control generation framework that utilizes intrinsic image properties such as material, geometry, and lighting to generate highly realistic images and supports 3D asset editing.
IntroStyle: Training-Free Introspective Style Attribution using Diffusion Features
Anand Kumar (University of California San Diego), Nuno Vasconcelos (University of California San Diego)
Data SynthesisRetrievalDiffusion modelImage
🎯 What it does: A completely training-free and external model-free style attribution method called IntroStyle is proposed, which utilizes the features of the diffusion model itself for style similarity assessment, and a synthetic dataset specifically for evaluating style attribution, ArtSplit, is released.
Inverse 3D Microscopy Rendering for Cell Shape Inference with Active Mesh
Sacha Ichbiah, Hervé Turlier (National Center for Scientific Research)
SegmentationOptimizationComputational EfficiencyImageMeshBiomedical Data
🎯 What it does: A differentiable 3D fluorescence microscope renderer, deltaMic, has been designed and implemented, which infers cell shapes directly from real 3D microscopic images using triangular meshes and a differentiable point spread function (PSF) through an inverse rendering approach.
Inverse Image-Based Rendering for Light Field Generation from Single Images
Hyunjun Jung (Gwangju Institute of Science and Technology), Hae-Gon Jeon (Yonsei University)
GenerationDepth EstimationTransformerDiffusion modelOptical FlowImage
🎯 What it does: A neural rendering network called iIBRnet is designed to reverse-engineer the optical flow space from a single image, achieving the generation of a complete 4D light field from a single image and supporting viewpoint synthesis and post-photographic effects.
Invisible Watermarks, Visible Gains: Steering Machine Unlearning with Bi-Level Watermarking Design
Yuhao Sun (University of Science and Technology of China), Sijia Liu (Michigan State University)
ClassificationGenerationOptimizationData-Centric LearningDiffusion modelImage
🎯 What it does: This paper proposes a watermark-assisted machine forgetting method called WATER4MU, which enhances the model's forgetting effect on specified data by embedding controllable watermarks into the training data.
InvRGB+L: Inverse Rendering of Complex Scenes with Unified Color and LiDAR Reflectance Modeling
Xiaoxue Chen (Tsinghua University), Shenlong Wang (Tsinghua University)
Image TranslationRestorationAutonomous DrivingNeural Radiance FieldImagePoint Cloud
🎯 What it does: This paper proposes the InvRGB+L inverse rendering framework, which utilizes single-frame RGB+LiDAR sequences to reconstruct re-lightable large-scale dynamic scenes, supporting nighttime simulation, lighting changes, and object insertion.
IQA-Adapter: Exploring Knowledge Transfer from Image Quality Assessment to Diffusion-based Generative Models
Khaled Abud (Lomonosov Moscow State University), Dmitriy Vatolin (Lomonosov Moscow State University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: By incorporating the knowledge of an IQA (Image Quality Assessment) model into a diffusion-based generative model, controllable adjustment of the quality of generated images is achieved. An IQA-Adapter is proposed to learn the implicit relationship between images and quality scores, and it supports quality transfer from reference images.
IRASim: A Fine-Grained World Model for Robot Manipulation
Fangqi Zhu (Hong Kong University of Science and Technology), Tao Kong (Hong Kong University of Science and Technology)
GenerationRobotic IntelligenceTransformerDiffusion modelWorld ModelVideo
🎯 What it does: This paper presents IRASim, a fine-grained world model for robotic operations that can generate high-fidelity videos based on historical observations and the robot's action trajectories, accurately simulating robot-object interactions.
IRGPT: Understanding Real-world Infrared Image with Bi-cross-modal Curriculum on Large-scale Benchmark
Zhe Cao (Beijing Institute of Technology), Ruiheng Zhang (Beijing Institute of Technology)
RecognitionImage TranslationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Developed IRGPT, a multimodal large language model for real infrared images, and constructed a dataset of over 260K infrared-text pairs, IR-TD, while evaluating on 9 infrared tasks.
Iris: Breaking GUI Complexity with Adaptive Focus and Self-Refining
Zhiqi Ge (Zhejiang University), Yueting Zhuang (Zhejiang University)
RecognitionObject DetectionOptimizationTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A visual GUI agent named Iris is proposed, utilizing Information-Sensitive Cropping (ISC) and Self-Improving Bidirectional Learning (SRDL) to achieve efficient and accurate GUI understanding and localization.
Is CLIP ideal? No. Can we fix it? Yes!
Raphi Kang (California Institute of Technology), Pietro Perona (California Institute of Technology)
RetrievalRepresentation LearningConvolutional Neural NetworkVision Language ModelContrastive LearningImageText
🎯 What it does: Analyzed the geometric constraints of the CLIP latent space, proving that it cannot simultaneously satisfy the demands of attribute binding, spatial relationships, and negation. Subsequently proposed Dense Cosine Similarity Maps (DCSM) and Functional Rows mechanism, which calculates cosine similarity at the token and patch levels in CLIP and trains a lightweight CNN to score it, achieving more accurate image-text matching.
Is Less More? Exploring Token Condensation as Training-free Test-time Adaptation
Zixin Wang (University of Queensland), Yadan Luo (University of Queensland)
CompressionDomain AdaptationTransformerVision Language ModelImage
🎯 What it does: A training-independent test-time adaptive method TCA is proposed, which enhances the zero-shot inference performance of VLMs like CLIP under distribution shift by dynamically compressing tokens in Vision Transformer and using domain-aware token pooling for logits self-correction.
Is Meta-Learning Out? Rethinking Unsupervised Few-Shot Classification with Limited Entropy
Yunchuan Guan (Huazhong University of Science and Technology), Lei Li (Nanyang Technological University)
ClassificationMeta LearningImage
🎯 What it does: The paper proposes a meta-learning based unsupervised few-shot classification framework called MINO, and fairly compares meta-learning with full-class training by introducing an entropy-constrained supervision setting.
Is Tracking Really More Challenging in First Person Egocentric Vision?
Matteo Dunnhofer (University of Udine), Christian Micheloni (University of Udine)
Object TrackingTransformerVideoBenchmark
🎯 What it does: This paper proposes the VISTA benchmark, which uses synchronized first-person and third-person videos to evaluate visual object tracking and explores the impact of perspective on performance.
Is Visual in-Context Learning for Compositional Medical Tasks within Reach?
Simon Reiß (Karlsruhe Institute of Technology), Rainer Stiefelhagen (Karlsruhe Institute of Technology)
SegmentationGenerationTransformerGenerative Adversarial NetworkImageBiomedical DataComputed TomographyUltrasound
🎯 What it does: This paper proposes a visual context learning framework that can handle multi-step medical task sequences on a single model and supports task adaptation during testing.
ISP2HRNet: Learning to Reconstruct High Resolution Image from Irregularly Sampled Pixels via Hierarchical Gradient Learning
Yuanlin Wang (Peking University), Tiejun Huang (Peking University)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: Proposes the ISP2HRNet network, which reconstructs high-resolution images from irregularly sampled pixels.
JailbreakDiffBench: A Comprehensive Benchmark for Jailbreaking Diffusion Models
Xiaolong Jin (Pennsylvania State University), Xiangyu Zhang (Purdue University)
GenerationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageVideoTextMultimodalityBenchmark
🎯 What it does: This paper presents JailbreakDiffBench, a security evaluation benchmark for text-to-image/video diffusion models, which includes human-annotated adversarial prompts and generated images, evaluation protocols, adversarial attack modules, and system security analysis tools.
Jailbreaking Multimodal Large Language Models via Shuffle Inconsistency
Shiji Zhao (Institute of Artificial Intelligence, State Key Laboratory of Virtual Reality Technology and Systems, Beihang University), Xingxing Wei (Institute of Artificial Intelligence, State Key Laboratory of Virtual Reality Technology and Systems, Beihang University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This study investigates and utilizes multimodal large language models to understand the inconsistency between the comprehension and security mechanisms of shuffled (shuffle) malicious text and image instructions, proposing a query-driven black-box text-image SI-Attack method.
Jigsaw++: Imagining Complete Shape Priors for Object Reassembly
Jiaxin Lu (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
GenerationData SynthesisRectified FlowPoint Cloud
🎯 What it does: This paper proposes Jigsaw++, a method that utilizes generative models to generate complete shape priors from partially assembled point clouds, assisting in object reassembly tasks.
Joint Asymmetric Loss for Learning with Noisy Labels
Jialiang Wang (Harbin Institute of Technology), Xiangyang Ji (Tsinghua University)
ClassificationImage
🎯 What it does: A joint asymmetric loss framework JAL is proposed, which combines a novel Asymmetric Mean Squared Error (AMSE) with traditional symmetric loss for robust training in noisy label environments.
Joint Diffusion Models in Continual Learning
Paweł Skierś (Warsaw University of Technology), Kamil Deja (Warsaw University of Technology)
ClassificationGenerationKnowledge DistillationDiffusion modelImage
🎯 What it does: A continuous learning framework based on the Joint Diffusion Model (JDCL) is proposed, utilizing the same network to simultaneously perform generative replay and classification tasks, reducing model parameters and training time.
Joint Learning of Pose Regression and Denoising Diffusion with Score Scaling Sampling for Category-level 6D Pose Estimation
Seunghyun Lee (KAIST), Tae-Kyun Kim (Imperial College London)
Pose EstimationDiffusion modelScore-based ModelPoint Cloud
🎯 What it does: This paper proposes a method that combines learning for pose regression and a denoising diffusion model, achieving single-pose inference through time-dependent score scaling sampling.
Joint Self-Supervised Video Alignment and Action Segmentation
Ali Shah Ali (Retrocausal), Quoc-Huy Tran (Retrocausal)
RecognitionSegmentationVideo
🎯 What it does: This paper proposes a unified framework for self-supervised video alignment and action segmentation based on fused Gromov-Wasserstein optimal transport (FGW) and structural priors (VAOT and VASOT).
Joint Semantic and Rendering Enhancements in 3D Gaussian Modeling with Anisotropic Local Encoding
Jingming He (City University of Hong Kong), Sam Kwong (Lingnan University)
SegmentationGenerationTransformerNeural Radiance FieldGaussian SplattingPoint Cloud
🎯 What it does: A 3D Gaussian modeling framework with joint semantic and rendering enhancement is proposed;
JointDiT: Enhancing RGB-Depth Joint Modeling with Diffusion Transformers
Kwon Byung-Ki (POSTECH), Tae-Hyun Oh (KAIST)
GenerationDepth EstimationTransformerDiffusion modelFlow-based ModelImageMultimodality
🎯 What it does: This paper proposes JointDiT, a model for RGB-Depth joint distribution modeling based on a diffusion Transformer, capable of performing joint generation, depth estimation, and depth-conditioned image generation at any noise level.
JPEG Processing Neural Operator for Backward-Compatible Coding
Woo Kyoung Han (Korea University), Kyong Hwan Jin (Korea University)
RestorationCompressionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a learning-based codec-decoder JPNeO that is compatible with the existing JPEG standard, significantly improving image compression and reconstruction quality while maintaining compatibility with traditional JPEG.
Kaleidoscopic Background Attack: Disrupting Pose Estimation with Multi-Fold Radial Symmetry Textures
Xinlong Ding (University of Science and Technology Beijing), Jiansheng Chen (University of Science and Technology Beijing)
Pose EstimationAdversarial AttackImage
🎯 What it does: A background attack method based on multi-fold symmetric textures (Kaleidoscopic Background Attack, KBA) is proposed, which disrupts the pose estimation model of sparse-view cameras by constructing similar disc-shaped backgrounds from multiple perspectives.
Kaputt: A Large-Scale Dataset for Visual Defect Detection
Sebastian Höfer (Amazon), Anton Milan (Amazon)
Object DetectionAnomaly DetectionTransformerSupervised Fine-TuningVision Language ModelImageBenchmark
🎯 What it does: A large-scale and diverse retail logistics visual defect detection dataset, Kaputt, has been constructed, and various existing defect detection and anomaly detection methods have been benchmarked on this dataset.
KDA: Knowledge Diffusion Alignment with Enhanced Context for Video Temporal Grounding
Ran Ran (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
TransformerDiffusion modelVideo
🎯 What it does: Proposes the KDA (Knowledge Diffusion Alignment) framework, which utilizes a background residual diffusion model to remove background from global video features to generate query-relevant temporal knowledge, and combines multi-layer video knowledge extraction, knowledge prompting reasoning, and span-enhanced decoding to achieve video temporal localization.
Keep Your Friends Close, and Your Enemies Farther: Distance-aware Voxel-wise Contrastive Learning for Semi-supervised Multi-organ Segmentation
Haochen Zhao (Beihang University), Yong Wang (Chinese Academy of Medical Sciences)
SegmentationContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A distance-aware voxel contrastive learning method (DVCL) is proposed, which utilizes unreliable voxels with pseudo-labels to learn more discriminative features.
Kestrel: 3D Multimodal LLM for Part-Aware Grounded Description
Mahmoud Ahmed (King Abdullah University of Science and Technology), Mohamed Elhoseiny (King Abdullah University of Science and Technology)
SegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningMultimodalityPoint Cloud
🎯 What it does: The Part-Aware Point Grounded Description (PaPGD) task is proposed, combining text generation and point cloud segmentation to achieve fine-grained descriptions and mask segmentation of 3D object components.
Keyframe-oriented Vision Token Pruning: Enhancing Efficiency of Large Vision Language Models on Long-Form Video Processing
Yudong Liu (Duke University), Yiran Chen (Duke University)
Computational EfficiencyTransformerSupervised Fine-TuningVision Language ModelVideoBenchmark
🎯 What it does: This paper proposes the Keyframe-Oriented Vision Token Pruning (KVTP) framework, which utilizes a query-frame correlation predictor to achieve adaptive pruning rates for each frame in long videos, significantly reducing the number of visual tokens while maintaining contextual continuity.
kh: Symmetry Understanding of 3D Shapes via Chirality Disentanglement
Weikang Wang (University of Bonn), Florian Bernard (University of Bonn)
Diffusion modelMesh
🎯 What it does: An unsupervised process is proposed to extract chiral features of 3D shape vertices from 2D base models, achieving decoupling of left and right directions.
KinMo: Kinematic-aware Human Motion Understanding and Generation
Pengfei Zhang (University of California), Bindita Chaudhuri (Flawless AI)
GenerationRetrievalTransformerLarge Language ModelAuto EncoderContrastive LearningVideoTextMultimodality
🎯 What it does: The KinMo framework is proposed, which uses six joint-based kinematic groups (torso, head, left arm, right arm, left leg, right leg) to provide a fine-grained description of human actions and constructs the corresponding KinMo dataset. This framework enables tasks such as text-action retrieval, text-driven fine-grained action generation, editing, and trajectory control.
Know "No" Better: A Data-Driven Approach for Enhancing Negation Awareness in CLIP
Junsung Park (Seoul National University), Sungroh Yoon (Seoul National University)
Data SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper employs a data-driven approach, utilizing large language models and multimodal language models to generate visual-text pairs containing negation words, and fine-tunes the text encoder of CLIP with this data to obtain NegationCLIP, which has a stronger ability to perceive negation.
Know Your Attention Maps: Class-specific Token Masking for Weakly Supervised Semantic Segmentation
Joëlle Hanna (University of St.Gallen), Damian Borth (University of St.Gallen)
SegmentationTransformerImage
🎯 What it does: This paper proposes an end-to-end weakly supervised semantic segmentation method based on multi-class [CLS] tokens in Vision Transformer, generating pseudo-masks directly from self-attention maps through random masking and attention head pruning.
Knowledge Distillation for Learned Image Compression
Yunuo Chen (Shanghai Jiao Tong University), Wenjun Zhang (Shanghai Jiao Tong University)
CompressionKnowledge DistillationImage
🎯 What it does: A framework called SMoDi based on staged modular distillation is proposed to compress large-scale learning-based image compression models into lightweight KDIC student models.
Knowledge Distillation with Refined Logits
Wujie Sun (Zhejiang University), Can Wang (Zhejiang University)
Knowledge DistillationImage
🎯 What it does: A new knowledge distillation method called Refined Logic Distillation (RLD) is proposed to address the limitations of current logic distillation methods, particularly the impact of incorrect predictions from the teacher model on the learning of the student model.
Knowledge Transfer from Interaction Learning
Yilin Gao (Shanghai University), Shugong Xu (Xi'an Jiaotong-Liverpool University)
ClassificationObject DetectionSegmentationDomain AdaptationKnowledge DistillationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes to achieve knowledge transfer by explicitly modeling interaction patterns and utilizing the cross-modal attention information from Visual Language Models (VLM) to guide the learning of Visual Foundation Models (VFM).
Knowledge-Guided Part Segmentation
Xuejian Gou (Xidian University), Wenping Ma (Xidian University)
SegmentationGraph Neural NetworkTransformerLarge Language ModelImageMultimodality
🎯 What it does: This paper proposes a knowledge-guided fine-grained part segmentation framework KPS, which utilizes large language models to construct a part structure knowledge graph and combines it with coarse-grained object guidance to achieve high-precision pixel-level segmentation of fine-grained parts in images.
KOEnsAttack: Towards Efficient Data-Free Black-Box Adversarial Attacks via Knowledge-Orthogonalized Substitute Ensembles
Chaoyong Yang (Fuzhou University), Wei Lin (Fujian University of Technology)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A data-independent black-box adversarial attack framework called KOEnsAttack is proposed, which generates synthetic samples through a generator, then transfers them to the controversial regions of two sub-networks using sample hardness enhancement, and finally trains a twin network ensemble model with a knowledge orthogonalization module. White-box attacks are then performed on this ensemble model to generate transferable adversarial samples.
KV-Edit: Training-Free Image Editing for Precise Background Preservation
Tianrui Zhu (Tsinghua University), Yansong Tang (Tsinghua University)
Image TranslationImage HarmonizationRestorationTransformerDiffusion modelFlow-based ModelImageBenchmark
🎯 What it does: A training-free image editing method named KV-Edit is proposed, which utilizes the KV cache of the DiT model to retain only the key-value pairs of the background, and then regenerates only the foreground to achieve precise background consistency.
LA-MOTR: End-to-End Multi-Object Tracking by Learnable Association
Peng Wang (Renmin University of China), Deying Li (Renmin University of China)
Object TrackingTransformerVideo
🎯 What it does: This paper proposes LA-MOTR, an end-to-end multi-object tracking framework that separates detection and association tasks.
Laboring on less labors: RPCA Paradigm for Pan-sharpening
Honghui Xu (Zhejiang University of Technology), Jianwei Zheng (Zhejiang University of Technology)
RestorationSuper ResolutionConvolutional Neural NetworkTransformerImage
🎯 What it does: A disassembly network (RUN) based on Robust Principal Component Analysis (RPCA) is proposed, which models the spatial downsampling error as a sparse Spatial Offset Component (SOC), thereby eliminating the need for approximating complex spatial downsampling matrices.
LACONIC: A 3D Layout Adapter for Controllable Image Creation
Léopold Maillard (Dassault Systèmes), Maks Ovsjanikov (L'École Polytechnique)
GenerationData SynthesisTransformerDiffusion modelImagePoint Cloud
🎯 What it does: We propose LACONIC, an adapter that can inject 3D semantic layout conditions into a pre-trained text-to-image diffusion model, supporting the generation of multi-view consistent realistic images from a single-view layout, and allowing for position, size, and semantic editing of individual objects.
LaCoOT: Layer Collapse through Optimal Transport
Victor Quétu (Telecom Paris), Enzo Tartaglione
GenerationCompressionConvolutional Neural NetworkImage
🎯 What it does: This paper reduces the distribution difference of intermediate features by incorporating optimal transport-based Max-Sliced Wasserstein distance regularization during the training process, allowing for the direct conversion of multiple layers into equivalent linear blocks and their simultaneous pruning in the post-training phase.
LaneDiffusion: Improving Centerline Graph Learning via Prior Injected BEV Feature Generation
Zijie Wang (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
SegmentationAutonomous DrivingTransformerDiffusion modelImage
🎯 What it does: This paper proposes LaneDiffusion, which utilizes a diffusion model to generate lane centerline priors in the Bird's Eye View (BEV) feature layer, and then obtains high-quality lane centerline maps and topological structures through a decoder.
LangBridge: Interpreting Image as a Combination of Language Embeddings
Jiaqi Liao (Shanghai AI Laboratory), Yu Cheng (Shanghai AI Laboratory)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper studies and proposes the LangBridge adapter, which maps visual features to the text embedding space of large language models (LLMs) through a linear combination of language word vectors, thus achieving cross-LLM transfer without pre-training.
LangScene-X: Reconstruct Generalizable 3D Language-Embedded Scenes with TriMap Video Diffusion
Fangfu Liu (Tsinghua University), Yueqi Duan (Tsinghua University)
SegmentationGenerationData SynthesisTransformerDiffusion modelGaussian SplattingImageVideo
🎯 What it does: Generate a generalizable 3D language embedded scene from only two images with sparse perspectives.
LANGTRAJ: Diffusion Model and Dataset for Language-Conditioned Trajectory Simulation
Wei-Jer Chang (University of California Berkeley), Francesco Pittaluga
GenerationData SynthesisAutonomous DrivingGraph Neural NetworkDiffusion modelMultimodality
🎯 What it does: This paper presents LANGTRAJ, a diffusion model for multi-agent traffic trajectory generation conditioned on natural language, achieving controllable and realistic scene simulation within a closed-loop training framework.
Language Decoupling with Fine-grained Knowledge Guidance for Referring Multi-object Tracking
Guangyao Li (Xiamen University), Hanzi Wang (Xiamen University)
Object TrackingAutonomous DrivingTransformerVideoText
🎯 What it does: An end-to-end reference multi-object tracking (RMOT) framework named DKGTrack is proposed, which disassembles natural language expressions into location descriptions and motion states, and achieves precise target localization and trajectory prediction through fine language guidance.
Language Driven Occupancy Prediction
Zhu Yu (Zhejiang University), Hui-Liang Shen (Zhejiang University)
SegmentationAutonomous DrivingTransformerVision Language ModelAuto EncoderImagePoint Cloud
🎯 What it does: Designed and implemented the LOcc framework, which generates dense fine-grained 3D language occupancy ground truth using a semantic transfer label pipeline, and trains an open vocabulary occupancy prediction network; at the same time, it modifies the prediction head of existing supervised occupancy models to make it compatible with both geometric and language heads.
Language-Driven Multi-Label Zero-Shot Learning with Semantic Granularity
Shouwen Wang (Huazhong University of Science and Technology), Zhigang Zeng (Medical University of Vienna)
ClassificationTransformerLarge Language ModelVision Language ModelContrastive LearningImageText
🎯 What it does: Utilizing large language models to generate text descriptions that include visual attributes, hierarchical relationships, and co-occurring scenes, and learning distinguishable reconstructed category names in the CLIP shared space to achieve multi-label zero-shot classification without image data;
LaRender: Training-Free Occlusion Control in Image Generation via Latent Rendering
Xiaohang Zhan (Tencent), Dingming Liu (Tencent)
GenerationDiffusion modelImage
🎯 What it does: A training-free and fine-tuning-free image generation method called LaRender is proposed, which can accurately control the occlusion relationships of multiple objects in an image and supports advanced effects such as transparency, fog, and light intensity.
Large Learning Rates Simultaneously Achieve Robustness to Spurious Correlations and Compressibility
Melih Barsbey (Imperial College London), Tolga Birdal (Imperial College London)
ClassificationCompressionOptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: The paper studies how using a larger learning rate (LR) in deep learning training can simultaneously enhance the model's robustness to spurious correlations and the network's compressibility.