CVPR 2026 Papers — Page 2
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Adaptive Anisotropic Gaussian Splatting for Multi-contrast MRI Arbitrary-Scale Super-Resolution with Anatomy Guidance
Qiuhai Yan (East China Normal University), Guixu Zhang (East China Normal University)
Super ResolutionGaussian SplattingBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes GaussM2 ASR, a multi-contrast MRI arbitrary-scale super-resolution framework based on 2D Gaussian splatting;
Adaptive Auxiliary Prompt Blending for Target-Faithful Diffusion Generation
Kwanyoung Lee (Hanyang University), Dong-Jin Kim (Hanyang University)
GenerationPrompt EngineeringDiffusion modelFlow-based ModelImage
🎯 What it does: Propose the Adaptive Auxiliary Prompt Blending (AAPB) framework, which utilizes auxiliary anchors to adaptively blend the diffusion process under low-density distributions, thereby enhancing target alignment and structural fidelity in rare concept generation and image editing.
Adaptive Bayesian Early-Exit Networks for Efficient Non-Transferable Learning
Siyu Luan (University of Copenhagen), Zhenyi Wang (University of Central Florida)
Domain AdaptationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes an adaptive Bayesian early exit network named ENL-DEE for non-transferable learning (NTL), achieving model ownership verification and usage authorization while significantly improving training and inference efficiency.
Adaptive Capacity Autoregressive Visual Tracking
Tong Lin (Xi'an Jiaotong University), Xing Wei (Xi'an Jiaotong University)
Object TrackingDiffusion modelVideo
🎯 What it does: Propose ARTrack-AC, an adaptive capacity autoregressive visual tracking framework that dynamically adjusts inference capacity to balance accuracy and speed;
Adaptive Confidence Regularization for Multimodal Failure Detection
Moru Liu (Technical University of Munich), Mario Trapp (Technical University of Munich)
Anomaly DetectionConvolutional Neural NetworkVideoMultimodalityPoint CloudAudio
🎯 What it does: This paper proposes a failure detection framework called ACR for multi-modal systems, which can proactively identify and reject unreliable results when prediction errors occur.
Adaptive Data Augmentation with Multi-armed Bandit: Sample-Efficient Embedding Calibration for Implicit Pattern Recognition
Minxue Tang (Accenture), Yujia Bao (Accenture)
RecognitionData SynthesisRepresentation LearningTransformerLarge Language ModelImageText
🎯 What it does: Proposes the ADAMAB framework, which achieves few-shot embedding calibration by training a lightweight calibration network on a fixed embedding model and employing multi-armed bandit-driven adaptive data augmentation.
Adaptive Depth Lightweight RGB-T Tracking with Holistic Token Routing
Tian Ding (Nanjing University), Ying Tai (Nanjing University)
Object TrackingComputational EfficiencyTransformerContrastive LearningMultimodality
🎯 What it does: This paper proposes ADTrack, a lightweight RGB-T tracker that integrates an adaptive early stopping mechanism and global token-guided cross-modal interaction.
Adaptive Learned Image Compression with Graph Neural Networks
Yunuo Chen (Shanghai Jiao Tong University), Guo Lu (Shanghai Jiao Tong University)
CompressionConvolutional Neural NetworkGraph Neural NetworkTransformerAuto EncoderImage
🎯 What it does: Proposes GLIC, an adaptive learning image compression framework based on graph neural networks (GNNs), which achieves content-adaptive feature aggregation using dual-scale graphs and adaptive adjacency.
Adaptive Spatial-Temporal Window: Unlocking the Potential of Event Cameras in Heterogeneous Velocity Scenarios
Zhipeng Sui (Tsinghua University), Wenhui Wang (Tsinghua University)
Object DetectionObject TrackingAutonomous DrivingMultimodality
🎯 What it does: Propose the Adaptive Spatial-Temporal Window (ASTW) strategy to perform adaptive spatiotemporal partitioning of event streams, adapting to multi-speed targets in the scene;
Adaptive Spectral Feature Forecasting for Diffusion Sampling Acceleration
Jiaqi Han (Stanford University), Stefano Ermon (Stanford University)
GenerationComputational EfficiencyDiffusion modelImageVideoBenchmark
🎯 What it does: Proposes a training-free feature caching and prediction method (Spectrum), which accelerates text-to-image/video diffusion sampling by globally fitting latent features along the time axis using Chebyshev polynomials during the diffusion process, while maintaining generation quality.
Adaptive Video Distillation: Mitigating Oversaturation and Temporal Collapse in Few-Step Generation
Yuyang You (Peking University), Peng Jiang (Kuaishou Technology)
GenerationKnowledge DistillationDiffusion modelAuto EncoderVideoBenchmark
🎯 What it does: Propose a distillation framework for video diffusion models to address color over-saturation and temporal collapse issues that occur during multi-step generation;
AdapTok: Learning Adaptive and Temporally Causal Video Tokenization in a 1D Latent Space
Yan Li (Shanghai Jiao Tong University), Xue Yang (Shanghai Jiao Tong University)
CompressionRepresentation LearningTransformerAuto EncoderVideo
🎯 What it does: Designed and implemented an adaptive temporal causal video tokenizer, AdapTok, which dynamically allocates token numbers based on video content while maintaining a 1-D latent space and temporal causality.
AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition
Zichuan Lin (Tencent Hunyuan), Deheng Ye (Tencent Hunyuan)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Propose the AdaptVision framework, which utilizes tool calls to achieve coarse-to-fine adaptive visual token acquisition, significantly reducing token consumption;
AdaRadar: Rate Adaptive Spectral Compression for Radar-based Perception
Jinho Park (Columbia University), Mingoo Seok (Columbia University)
Compression
🎯 What it does: Propose AdaRadar, an adaptive spectral-domain compression framework that can compress high-dimensional radar Range-Doppler data to extremely low bitrates, while dynamically adjusting the compression rate through task-oriented feedback control.
AdaSFormer: Adaptive Serialized Transformers for Monocular Semantic Scene Completion from Indoor Environments
Xuzhi Wang (Tianjin Normal University), Ziping Zhao (Tianjin Normal University)
SegmentationDepth EstimationConvolutional Neural NetworkTransformerImagePoint Cloud
🎯 What it does: Proposed an indoor monocular semantic scene completion model named AdaSFormer based on an adaptive serialized Transformer.
AdaSpark: Adaptive Sparsity for Efficient Long-Video Understanding
Handong Li (University of Chinese Academy of Sciences), Jing Liu (University of Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: This paper proposes AdaSpark, an adaptive sparse framework for efficient long video understanding;
AdaSpot: Spend Resolution Where It Matters for Precise Event Spotting
Artur Xarles (Universitat de Barcelona), Albert Clapés (Universitat de Barcelona)
Object DetectionConvolutional Neural NetworkRecurrent Neural NetworkVideo
🎯 What it does: Propose the AdaSpot framework, which uses low-resolution global features to guide adaptive selection of the most informative regions of interest (ROI) in each frame for high-resolution processing, achieving high-precision event localization;
AdaSVD: Singular Value Decomposition with Adaptive Mechanisms for Large Multimodal Models
Zhiteng Li (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
CompressionComputational EfficiencyImageTextMultimodality
🎯 What it does: Propose AdaSVD, an adaptive SVD compression method tailored for large-scale multimodal models.
Addressing Exacerbated Attention Sink for Source-Free Cross-Domain Few-Shot Learning
Shuai Yi (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelImageBiomedical Data
🎯 What it does: Under the cross-domain few-shot learning (CDFSL) scenario, researchers find that standard target domain few-shot fine-tuning significantly amplifies the visual attention sink phenomenon, and propose a dynamic weighting method based on Token Importance Recalibration (TIR) to suppress 'sink' tokens and strengthen tokens with stronger class discriminativeness, thereby improving cross-domain few-shot classification performance.
ADSeeker: A Knowledge-Grounded Reasoning Framework for Industry Anomaly Detection and Reasoning
Kai Zhang (Shandong University), Jinglin Zhang (Shandong University)
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose ADSeeker, an industrial anomaly detection and reasoning framework based on knowledge retrieval and multimodal large language models;
Advancing Cancer Prognosis with Hierarchical Fusion of Genomic, Proteomic and Pathology Imaging Data from a Systems Biology Perspective
Junjie Zhou (Nanjing University of Aeronautics and Astronautics), Daoqiang Zhang (Nanjing University of Aeronautics and Astronautics)
ClassificationGraph Neural NetworkTransformerLarge Language ModelVision Language ModelImageMultimodalityTabular
🎯 What it does: This paper proposes the HFGPI framework, which utilizes multi-modal features from genomics, proteomics, and tissue slice images to achieve cancer prognosis prediction.
Advancing Image Classification with Discrete Diffusion Classification Modeling
Omer Belhasin (Technion Israel Institute of Technology), Michael Elad (Technion Israel Institute of Technology)
ClassificationConvolutional Neural NetworkDiffusion modelScore-based ModelImage
🎯 What it does: Propose a discrete diffusion classification model, DiDiCM, which learns the posterior distribution for image classification by performing diffusion in the discrete label space, thereby enhancing classification performance in high-uncertainty scenarios.
Adversarial Style Optimization: Enhancing VLM Jailbreaks by GRPO-based Stylistic Triggers Optimization
Bingjun Luo (Tsinghua University), Xinpeng Ding (Xidian University)
Safty and PrivacyAdversarial AttackReinforcement LearningVision Language ModelFlow-based ModelImageTextMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes the Adversarial Style Optimization (ASO) framework, which enhances the success rate of jailbreak attacks on vision-language models (VLMs) through image style optimization;
AdvFM: Lookahead Flow-Matching Velocity-Field Attacks for Imperceptible and Transferable Adversarial Examples
Runze Liu (Harbin Institute of Technology), Zhaoyang Zhang (Harbin Institute of Technology)
Adversarial AttackFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: Propose an attack framework called AdvFM based on flow matching, which injects perturbations into the velocity field over continuous time instead of directly modifying images, generating unconstrained adversarial examples
AE2VID: Event-based Video Reconstruction via Aperture Modulation
Chenxu Bai (Peking University), Boxin Shi (Peking University)
RestorationData SynthesisConvolutional Neural NetworkRecurrent Neural NetworkAuto EncoderVideo
🎯 What it does: By periodically adjusting the aperture on an event camera, and combining events generated by aperture modulation with traditional motion-triggered events as inputs, the AE2VID framework is constructed to achieve high-frame-rate, high-dynamic-range video reconstruction.
AERGS-SLAM: Auto-Exposure-Robust Stereo 3D Gaussian Splatting SLAM
Zhiyu Zhou (South China University of Technology), Yu Liu (South China University of Technology)
OptimizationGaussian SplattingSimultaneous Localization and MappingImageVideo
🎯 What it does: Proposed an adaptive exposure-robust 3D Gaussian Splatting SLAM framework called AERGS-SLAM, achieving realistic mapping while simultaneously performing illumination-robust localization and exposure control.
AeroAgent: A Vision-Physics-Decision Framework for Aerodynamic Vehicle Design
Ye Liu (Eastern Institute of Technology), Yuntian Chen (Eastern Institute of Technology)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelMultimodalityMeshBenchmarkPhysics Related
🎯 What it does: Proposes AeroAgent, a unified vision-physical-decision framework for achieving early vehicle wind tunnel design under strict CFD budgets; and trains AeroFormer, a physics surrogate based on this framework, to rapidly predict three-dimensional pressure fields, velocity fields, and drag coefficients.
AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D Reconstruction
Hanyang Liu (Ohio State University), Rongjun Qin (Ohio State University)
GenerationOptimizationGaussian SplattingVideoPhysics Related
🎯 What it does: Proposes a physically consistent dynamic Gaussian splatting framework called AeroDGS for single-sequence aerial 4D reconstruction, enabling the reconstruction of physically consistent 4D models from single-view dynamic urban scenes.
AeroGS: Scale-Aware Gaussian Splatting for Pose-Free Dynamic UAV Scene Reconstruction
Tingyun Li (Wuhan University), Jiahao Liu (Wuhan University)
Autonomous DrivingGaussian SplattingSimultaneous Localization and MappingOptical FlowVideo
🎯 What it does: Propose AeroGS, which utilizes a scale-aware Gaussian scattering framework to simultaneously recover camera trajectory and dynamic scene geometry from monocular UAV videos without pose information.
Aesthetic Camera Viewpoint Suggestion with 3D Aesthetic Field
Sheyang Tang (University of Waterloo), Zhou Wang (University of Waterloo)
Recommendation SystemOptimizationTransformerGaussian SplattingImageVideo
🎯 What it does: Propose a 3D aesthetic scene model based on 3D Gaussian Splatting, combined with a two-stage (coarse sampling + gradient upsampling) perspective search to automatically suggest high-aesthetic viewpoints from sparse camera captures.
Affine Perspective-Three-Point Problem
Gaku Nakano (NEC Corporation)
Pose EstimationImage
🎯 What it does: Proposes direct closed-form solvers for the Perspective-Three-Point (P3P) problem under both weak perspective and parametric perspective affine camera models, and provides a lightweight iterative refinement method based on affine solutions, ultimately achieving accuracy comparable to existing precise P3P solvers.
Affordance Field Intervention: Enabling VLAs to Escape Memory Traps in Robotic Manipulation
Siyu Xu (University of Sydney), Chang Xu (University of Sydney)
SegmentationDepth EstimationRobotic IntelligenceLarge Language ModelVision-Language-Action ModelImageTextMultimodalityPoint Cloud
🎯 What it does: Propose the Affordance Field Intervention (AFI) framework, utilizing a 3D position-aware Affordance Field (SAF) as a plug-in module to help Vision-Language-Action (VLA) models escape 'memory traps' in out-of-distribution environments through historical rollback, SAF-guided waypoint sampling, and SAF-scored reselection of optimal trajectories.
Affordance-First Decomposition for Continual Learning in Video-Language Understanding
Mengzhu xu, Canran Xiao (Nanjing Normal University)
Representation LearningMeta LearningLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose Affordance-First Decomposition (AFD) to decompose video-language continual learning, mapping videos to slowly changing affordance tokens and enabling targeted adaptation through query-routing adapters.
AffordGen: Generating Diverse Demonstrations for Generalizable Object Manipulation with Affordance Correspondence
Jiawei Zhang (Shanghai Qi Zhi Institute), Huazhe Xu (Tsinghua University)
Data SynthesisRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelPoint CloudMesh
🎯 What it does: By leveraging semantic affordance key points in 3D meshes, a small number of real demonstrations are transformed into thousands of diverse, full 6D robotic manipulation trajectories. These synthetic data are used to train closed-loop visual motion control policies, achieving zero-shot generalization to novel objects.
AffordGrasp: Cross-Modal Diffusion for Affordance-Aware Grasp Synthesis
Xiaofei Wu (ShanghaiTech University), Xuming He (ShanghaiTech University)
GenerationRobotic IntelligenceLarge Language ModelDiffusion modelTextPoint Cloud
🎯 What it does: Propose a cross-modal diffusion framework named AffordGrasp, which can generate human grasping poses that are both semantically meaningful and physically feasible based on text instructions and 3D object geometry.
AffordMatcher: Affordance Learning in 3D Scenes from Visual Signifiers
Nghia Vu (AIOZ Ltd), Anh Nguyen (Indian Institute of Science)
SegmentationTransformerVision Language ModelContrastive LearningImageTextMultimodalityPoint CloudBenchmark
🎯 What it does: Construct a large-scale multimodal dataset named AffordBridge, and propose the AffordMatcher method, which uses visual indicators (RGB images + text) to locate and segment interactive regions in 3D scenes.
Affostruction: 3D Affordance Grounding with Generative Reconstruction
Chunghyun Park (Pohang University of Science and Technology), Minsu Cho (Pohang University of Science and Technology)
GenerationTransformerVision-Language-Action ModelFlow-based ModelImageTextMultimodalityPoint CloudMesh
🎯 What it does: This paper proposes a framework called Affostruction, which can reconstruct complete 3D object geometry from local RGBD images and localize affordance regions on the complete shape for arbitrary natural language action queries.
AG-VAS: Anchor-Guided Zero-Shot Visual Anomaly Segmentation with Large Multimodal Models
Zhen Qu (Institute of Automation Chinese Academy of Sciences), Xingang Wang (Institute of Automation Chinese Academy of Sciences)
SegmentationAnomaly DetectionTransformerLarge Language ModelVision Language ModelImageMultimodalityBiomedical Data
🎯 What it does: Propose a zero-shot visual anomaly segmentation framework AG-VAS based on a multimodal model, which can directly output a binary anomaly mask.
Agent4FaceForgery: Multi-Agent LLM Framework for Realistic Face Forgery Detection
Yingxin Lai (Xiamen University), Xiaochun Cao (Sun Yat-sen University)
Anomaly DetectionTransformerLarge Language ModelAgentic AIGenerative Adversarial NetworkImageTextMultimodality
🎯 What it does: Propose Agent4FaceForgery, a multi-agent LLM framework that simulates the human face forgery lifecycle to generate eco-effective multimodal data;
AgentDet: A Shared-Blackboard Multi-Agent Framework for Zero-/Few-Shot Object Detection
Haolin Li (Tsinghua University), Biqing Huang (Tsinghua University)
Object DetectionTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelImageRetrieval-Augmented Generation
🎯 What it does: Proposes a multi-agent shared blackboard framework AgentDet, which unifies zero-shot and few-shot object detection through the integration of large language models (LLMs) and visual knowledge bases.
Agentic Retoucher for Text-To-Image Generation
Shaocheng Shen (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
RestorationGenerationTransformerAgentic AIVision Language ModelImageMultimodality
🎯 What it does: Proposes Agentic Retoucher, a three-agent closed-loop (perception-reasoning-execution) framework that automatically detects and repairs fine-grained distortions in text-to-image generation.
Agentic Video Summarization via Self-Reflecting Multimodal Understanding
Miaotian Guo (Beijing University of Posts and Telecommunications), Dongsheng Jiang (Huawei Technologies Co., Ltd)
GenerationTransformerLarge Language ModelAgentic AIVision Language ModelVideoMultimodality
🎯 What it does: Designed and implemented a closed-loop workflow called AgenticVS based on agents to automatically generate, verify, and self-correct video summaries.
AGENTSAFE: Benchmarking the Safety of Embodied Agents on Hazardous Instructions
Zonghao Ying (Beihang University), Xianglong Liu (Beihang University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelTextMultimodalityBenchmark
🎯 What it does: Propose the AGENTSAFE benchmark for systematically evaluating the safety of VLM-driven embodied agents when facing hazardous instructions, and reveal the root causes of safety failures through multi-stage diagnosis.
AGFT: Alignment-Guided Fine-Tuning for Zero-Shot Adversarial Robustness of Vision-Language Models
Yubo Cui, Zheng Zhang
ClassificationRepresentation LearningAdversarial AttackTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Propose Alignment-Guided Fine-Tuning (AGFT), which enhances zero-shot adversarial robustness while preserving cross-modal semantic structure through adversarial fine-tuning of a pre-trained VLM (CLIP).
Agile Deliberation: Concept Deliberation for Subjective Visual Classification
Leijie Wang (University of Washington), Ariel Fuxman
ClassificationLarge Language ModelPrompt EngineeringVision Language ModelImageText
🎯 What it does: Proposes the Agile Deliberation framework, helping users construct high-quality classifiers for subjective visual concepts through iterative processes and conceptual scope refinement.
AGiLe: Learning Robust Long-Horizon Manipulation via Affordance-Grounded Bidirectional Latent Planning
Zixuan Chen (Nanjing University), Yang Gao (Nanjing University)
Robotic IntelligenceTransformerDiffusion modelSequential
🎯 What it does: Proposed the AGiLe framework for robust planning and execution in long-horizon manipulation tasks, integrating bidirectional latent planning with empowerment-guided execution modules.
AHS: Adaptive Head Synthesis via Synthetic Data Augmentations
Taewoong Kang (KAIST), Jaegul Choo (KAIST)
Image TranslationImage HarmonizationData SynthesisPose EstimationConvolutional Neural NetworkTransformerDiffusion modelImageMultimodality
🎯 What it does: Propose a zero-shot head swapping method AHS, which utilizes synthetic data augmentation and multimodal attention mechanisms to achieve high-quality head synthesis with expression and pose synchronization in full upper-body images.
AIMDepth: Asymmetric Image-Event Mamba for Monocular Depth Estimation
Luoxi Jing (Peking University), Mengzhu Wang (Advanced Institute of Big Data)
Depth EstimationMultimodality
🎯 What it does: Proposes AIMDepth, a heterogeneous event-image depth estimation framework based on Mamba (state-space model), achieving high-precision monocular depth prediction with low computational cost through input-layer frequency-domain prior alignment, feature-layer asymmetric encoding, and multi-level local fusion modules for cross-modal alignment and fusion.
Air-Know: Arbiter-Calibrated Knowledge-Internalizing Robust Network for Composed Image Retrieval
Zhiheng Fu (Shandong University), Zixu Li (Shandong University)
RetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes the Air-Know framework to address the Noisy Triplet Correspondence (NTC) problem in Composed Image Retrieval, achieving robust learning through a three-stage method: External Prior Arbitration (EPA), Expert Knowledge Internalization (EKI), and Dual-Stream Harmonization (DSR).
AirSim360: A Panoramic Simulation Platform within Drone View
Xian Ge, Lu Qi (Insta360 Research)
SegmentationPose EstimationDepth EstimationAutonomous DrivingSupervised Fine-TuningVision-Language-Action ModelImageTextMultimodalityPoint Cloud
🎯 What it does: Developed the AirSim360 platform to achieve panoramic rendering from UAV perspectives, synchronized data collection and annotation, and provide closed-loop simulation and interactive human-machine systems.
AKCMamba-YOLO: Selective State Space Models For Real-Time Object Detection
Long Chen (East China Jiaotong University), Zizhu Fan (Shanghai University of Electric Power)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposes AKCMamba-YOLO, integrating the selective state space model with YOLO to achieve real-time high-precision detection.
AlcheMinT: Fine-grained Temporal Control for Multi-Reference Consistent Video Generation
Sharath Girish (Snap Inc), Sergey Tulyakov (Snap Inc)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelAuto EncoderImageVideoTextMultimodalityBenchmark
🎯 What it does: Developed a video generation framework called AlcheMinT, which can precisely control the appearance and disappearance of multiple reference subjects within specified time intervals during video generation.
Alert-CLIP: Abnormality-aware Latent-Enhanced Representation Tuning of CLIP for Video Anomaly Detection
Yiyan Zhu (Beijing University of Posts and Telecommunications), Jingyu Wang (Pengcheng Laboratory)
Anomaly DetectionTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: For video anomaly detection (VAD), this paper proposes Alert-CLIP, a framework that enhances CLIP's anomaly perception capability in anomaly recognition through multi-level vision-language alignment learning.
Align Images Before You Generate
Shihua Zhang (National University of Singapore), Xinchao Wang (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelImageVideo
🎯 What it does: This paper proposes a plug-and-play adapter called CorrAdapter for multi-image diffusion models, which enhances spatiotemporal consistency by leveraging the inherent correspondences in the model's internal intermediate features before image generation.
Align Once to Explain: Feature Alignment for Scalable B-cosification of Foundational Vision Transformers
Raphael Maser (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
ClassificationDepth EstimationExplainability and InterpretabilityKnowledge DistillationTransformerContrastive LearningImage
🎯 What it does: Proposes the ALign Once to Explain (ALOE) method, which converts large vision foundation models into interpretable B-cos networks through a single unlabeled feature alignment, achieving one-time, unsupervised, and directly reusable performance on downstream tasks.
Align While Search: Belief-Guided Exploratory Inference for World-Grounded Embodied Agents
Seohui Bae (LG AI Research), Woohyung Lim (LG AI Research)
Robotic IntelligenceLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringVision-Language-Action ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose an agent called Align While Search (AWS) that performs adaptive exploratory reasoning during testing, controlling actions through posterior-guided belief updates;
Aligning Multi-Character Narrative Image Generation with Multi-Aspect Human Preferences
Ziyi Gao (Fudan University), Yi-Ping Phoebe Chen (La Trobe University)
GenerationExplainability and InterpretabilityReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a fine-grained human preference dataset named NI-RLHF for multi-character narrative image generation. Based on this dataset, an interpretable NIReward reward model and ADPO optimization algorithm were trained, achieving better consistency between generated images and human aesthetic preferences.
Aligning Text, Images and 3D Structure Token-by-Token
Aadarsh Sahoo (California Institute of Technology), Georgia Gkioxari (California Institute of Technology)
GenerationTransformerLarge Language ModelVision Language ModelAuto EncoderImageTextPoint CloudMesh
🎯 What it does: Propose Kyvo, a unified autoregressive LLM framework that aligns structured 3D scenes with text and images, enabling 3D tasks such as rendering, reconstruction, instruction following, and question answering.
Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow
Chengxin Liu (Korea Advanced Institute Of Science And Technology), Tae-Hyun Oh (Korea Advanced Institute Of Science And Technology)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Propose a training-agnostic method (Adaptive Information Flow, AIF) that dynamically regulates the flow of visual-textual information during inference by blocking the attention between text tokens and irrelevant visual tokens to enhance the perceptual capabilities of vision-language models (VLM).
AlignPose: Generalizable 6D Pose Estimation via Multi-view Feature-metric Alignment
Anna Šárová Mikeštíková (Czech Technical University in Prague), Vladimir Petrik (Czech Technical University in Prague)
Pose EstimationOptimizationTransformerImage
🎯 What it does: Proposes a multi-view general 6D pose estimation method called AlignPose, which aggregates multi-camera RGB data and infers the pose of unseen objects through feature metric optimization.
All in One: Unifying Deepfake Detection, Tampering Localization, and Source Tracing with a Robust Landmark-Identity Watermark
Junjiang Wu (Xinjiang University), Zhiqing Guo (Xinjiang University)
Object DetectionRetrievalAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposed a unified deepfake detection, tampering localization, and source tracking framework called LIDMark, which employs a 152-dimensional landmark-identity watermark combined with a Factorized-Head Decoder to achieve multi-task recovery.
All Roads Lead to Rome: Incentivizing Divergent Thinking in Vision-Language Models
Xinyu Tian (Australian National University), Jing Zhang (Australian National University)
Reinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper investigates the performance of RL-based Vision-Language models (particularly GRPO) on reasoning tasks, finding that they are prone to diversity collapse, and proposes Multi-Group Policy Optimization (MUPO) to maintain diversity across multiple reasoning paths, achieving significant improvements on various mathematical and general reasoning benchmarks.
All Vehicles Can Lie: Efficient Adversarial Defense in Fully Untrusted-Vehicle Collaborative Perception via Pseudo-Random Bayesian Inference
Yi Yu (Wuhan University), Jiaqi Feng (Wuhan University)
Anomaly DetectionAutonomous DrivingComputational EfficiencyAdversarial AttackPoint Cloud
🎯 What it does: Proposes the PRBI (Pseudo-Random Bayesian Inference) defense framework, leveraging perceptual consistency between consecutive frames as a self-supervised reference, and achieving detection and isolation of malicious vehicles in fully collaborative perception without trusted vehicles through pseudo-random grouping and Bayesian inference;
All-in-One Slider for Attribute Manipulation in Diffusion Models
Weixin Ye (Beijing Jiaotong University), Xuecheng Nie (Meitu)
GenerationVision Language ModelDiffusion modelAuto EncoderImageText
🎯 What it does: Proposes All-in-One Slider, which utilizes a sparse autoencoder to decompose text embeddings into decoupled attribute vectors, enabling fine-grained, continuous, and conflict-free control over multiple facial attributes within diffusion models.
ALLNet: Multi-task Dense Prediction for Degraded Images
Weiran Wang (Xidian University), Yunsong Li (Xidian University)
RestorationSegmentationDepth EstimationConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: A unified ALLNet framework is proposed for degraded images, achieving both image restoration and multi-task pixel-level prediction, enabling multi-task dense prediction within a single network.
AMap: Distilling Future Priors for Ahead-Aware Online HD Map Construction
Ruikai Li (Beihang University), Zhiyong Cui (Durham University)
Autonomous DrivingKnowledge DistillationVideoPoint Cloud
🎯 What it does: Designed the AMap framework, which improves the accuracy of forward region HD map construction by distilling knowledge from a teacher model with future temporal context into a student model that only uses the current frame.
AMB3R: Accurate Feed-forward Metric-scale 3D Reconstruction with Backend
Hengyi Wang (University College London), Lourdes Agapito (University College London)
Pose EstimationDepth EstimationOptimizationTransformerSimultaneous Localization and MappingImage
🎯 What it does: Developed a multi-view feedforward model AMB3R, utilizing a sparse voxel backend for metric-scale 3D reconstruction, and proposed a training-free, model-agnostic end-to-end pipeline (AMB3R-VO and AMB3R-SfM) supporting online visual odometry and structure from motion with any number of images.
AMusE: Audio-Visual Benchmark and Alignment Framework for Agentic Multi-Speaker Understanding
Sanjoy Chowdhury (University of Maryland), Raviteja Vemulapalli (Apple)
Large Language ModelReinforcement LearningAgentic AIVision Language ModelVideoMultimodalityBenchmarkAudio
🎯 What it does: Constructed the AMUSE benchmark, which includes six multi-speaker audio-visual understanding tasks, and evaluated multimodal large language models under three assessment modes: zero-shot, guided, and agent; simultaneously proposed the RAFT alignment framework (RRO + SRA), significantly improving the model's agent reasoning and cross-modal consistency.
An Efficient Token Compression Framework for Visual Object Tracking
Weijing Wu (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
Object TrackingTransformerVideo
🎯 What it does: Propose a visual tracking framework named ETCTrack, which first compresses and then interacts, utilizing an adaptive token compressor (ATC) to remove redundant visual tokens from historical templates, and achieving deep interaction between search features and compressed templates through hierarchical interaction blocks (HIBlock);
An Empirical Study on How Video-LLMs Answer Video Questions
Chenhui Gou, Hamid Rezatofighi (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper systematically analyzes the internal processing of models through elimination experiments on the attention mechanisms of various Video-LLMs in the VideoQA task.
An Instance-Centric Panoptic Occupancy Prediction Benchmark for Autonomous Driving
Yi Feng (Tongji University), Rui Fan (Tongji University)
Data SynthesisAutonomous DrivingMultimodalityMeshBenchmark
🎯 What it does: Constructed a unified high-quality 3D mesh library ADMesh (containing 15k+ vehicle models, buildings, traffic facilities, and other multi-level semantic meshes), and generated a high-resolution, instance-level, physically consistent panoptic occupancy dataset CarlaOcc (100k frames, 0.05m voxel resolution, complete semantic and instance annotations) based on ADMesh and the CARLA simulation platform.
An Optimal Transport-driven Approach for Cultivating Latent Space in Online Incremental Learning
Quyen Tran (Rutgers University), Trung Le (Monash University)
OptimizationRepresentation LearningImage
🎯 What it does: Designed an online hybrid model based on optimal transport (MMOT) to dynamically learn a multi-center latent space and improve online incremental learning performance with a dynamic preservation strategy.
Anatomica: Localized Control over Geometric and Topological Properties for Anatomical Diffusion Models
Karim Kadry (MIT), Elazer R. Edelman (MIT)
GenerationDiffusion modelImageBiomedical Data
🎯 What it does: Proposed a three-dimensional multi-class anatomical segmentation generation framework that achieves controllable geometric and topological properties through local cube control domains during inference.
Anatomical Domain Shifts: Test-time Heterogeneous Adaptation for 3D Human Pose Prediction
Qiongjie Cui (Xi'an Jiaotong University), Na Zhao (Singapore University of Technology and Design)
Pose EstimationDomain AdaptationGraphTime Series
🎯 What it does: Propose a 3D human pose prediction framework based on test-time heterogeneous adaptation (TT-HA), which can adapt to distribution drift of different body parts in dynamic environments.
Anchor-Guided Gradient Alignment for Incomplete Multimodal Learning
Zhi-Hao Guan (Nanjing University of Science and Technology), Yang Yang (Nanjing University of Science and Technology)
Representation LearningData-Centric LearningAuto EncoderMultimodalityRetrieval-Augmented Generation
🎯 What it does: Studied the problem of learning imbalance caused by missing modalities in multi-modal learning, and proposed the Anchor-guided Gradient Alignment (ANGA) framework.
AnchorFlow: Training-Free 3D Editing via Latent Anchor-Aligned Flows
Zhenglin Zhou (Zhejiang University), Tat-Seng Chua (National University of Singapore)
GenerationFlow-based ModelMeshBenchmark
🎯 What it does: Propose AnchorFlow, a no-training, no-mask 3D editing framework that achieves stable editing processes through global latent anchor consistency
Anchoring and Rescaling Attention for Semantically Coherent Inbetweening
Tae Eun Choi (Yonsei University), Seong Jae Hwang (Yonsei University)
GenerationData SynthesisTransformerDiffusion modelVideoTextBenchmark
🎯 What it does: Propose an untrained Attention mechanism that utilizes self-attention information from keyframes to guide the generation of intermediate frames, achieving semantically consistent and rhythm-stable frame interpolation.
Anchoring the Mind of Multimodal Reasoners: Cognitive Bias as a Vector for Jailbreak Attacks
Linhua Cong (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
Adversarial AttackPrompt EngineeringMultimodalityChain-of-Thought
🎯 What it does: Proposed a cross-modal jailbreak framework RA-Attack based on the cognitive anchoring effect, using structured visual mind maps and text security anchors to induce multi-modal reasoning models to generate malicious outputs.
AnchorSplat: Feed-Forward 3D Gaussian Splatting With 3D Geometric Priors
Xiaoxue Zhang (Huawei Inc), Dave Zhenyu Chen (Huawei Inc)
GenerationDepth EstimationTransformerGaussian SplattingPoint Cloud
🎯 What it does: AnchorSplat achieves scene-level high-quality novel view synthesis by performing anchor-aligned feed-forward operations on Gaussians in 3D space.
Ani3DHuman: Photorealistic 3D Human Animation with Self-guided Stochastic Sampling
Qi Sun (City University of Hong Kong), Jing Liao (City University of Hong Kong)
GenerationPose EstimationDiffusion modelGaussian SplattingVideoPoint CloudStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed a 3D human animation framework named ANI3DHUMAN based on hierarchical motion representation and self-guided stochastic sampling, capable of generating high-fidelity, identity-preserving non-rigid dynamic videos.
Animator-Centric Skeleton Generation on Objects with Fine-Grained Details
Mingze Sun (Tsinghua Shenzhen International Graduate School), Ruqi Huang (Tsinghua Shenzhen International Graduate School)
GenerationTransformerPoint CloudMesh
🎯 What it does: Propose an autoregressive skeleton generation framework tailored for animators, capable of automatically generating fine-grained, semantically consistent skeletons on complex 3D meshes;
AniMimic: Imitating 3D Animation from Video Priors
Tianyi Xie (University of California, Los Angeles), Chenfanfu Jiang (University of California, Los Angeles)
GenerationData SynthesisDiffusion modelVideoMesh
🎯 What it does: Developed a 3D mesh animation framework AniMimic based on motion priors from video diffusion models, capable of generating editable, physically simulable, and realistic animations from static meshes;
Annotation-Efficient Coreset Selection for Context-dependent Segmentation
Jin Zhang (Beijing Institute of Technology), Ruiheng Zhang (Beijing Institute of Technology)
SegmentationData-Centric LearningDiffusion modelImage
🎯 What it does: A method for core sample selection using weak annotations is proposed for context-aware segmentation tasks.
Anomaly as Non-Conformity via Training-Free Graph Laplacian Energy Minimization
Jungwook Seo (Hanyang University), Sungyong Baik (Hanyang University)
Anomaly DetectionOptimizationTransformerImage
🎯 What it does: Propose a training-free anomaly detection framework ANoCo, which solves anchored Laplacian energy minimization on a bipartite graph between queries and positive examples, using feature drift magnitude to measure whether queries are inconsistent with the normal manifold.
Anomaly-Related Residual Fields for Cross-domain Anomaly Detection
Kewei Gao (Zhejiang University), Yijun Bei (Zhejiang University)
Anomaly DetectionTransformerDiffusion modelImage
🎯 What it does: This paper proposes a cross-domain unlabeled anomaly detection framework that constructs a residual evolution field (REF) using the residual evolution of diffusion models to separate abnormal-related non-stationary signals, and achieves source domain model transfer to the target domain through cross-domain field alignment (CFA).
AnomalyVFM -- Transforming Vision Foundation Models into Zero-Shot Anomaly Detectors
Matic Fučka (University of Ljubljana), Danijel Skočaj (University of Ljubljana)
Data SynthesisAnomaly DetectionTransformerVision Language ModelFlow-based ModelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Propose a general framework AnomalyVFM, which can convert any pre-trained visual foundation model (VFM) into a powerful zero-shot anomaly detector.
AnthroTAP: Learning Point Tracking with Real-World Motion
Inès Hyeonsu Kim (KAIST AI), Seungryong Kim (KAIST AI)
Object TrackingData SynthesisPose EstimationOptical FlowVideoMesh
🎯 What it does: By fitting the SMPL model to human videos, using ray casting to determine visibility and optical flow consistency filtering, automatically generate high-quality point tracking pseudo annotations and use them to train a point tracking model.
Anti-Degradation Lifelong Multi-View Clustering
Xingfeng Li (Southwest University of Science and Technology), Zhenwen Ren (Southwest University of Science and Technology)
Representation LearningMeta LearningAuto EncoderMultimodality
🎯 What it does: Proposes an Anti-Decline Lifelong Multi-View Clustering (ALMC) framework that can continuously learn in real-world scenarios where views are collected in a streaming manner and avoid knowledge loss.
Anti-I2V: Safeguarding your Photos from Malicious Image-to-video Generation
Duc Vu (Qualcomm AI Research), Anh Tran (Qualcomm AI Research)
Safty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelContrastive LearningVideo
🎯 What it does: Propose Anti-I2V, an adversarial defense method for image-to-video diffusion models, which can suppress identity consistency and temporal coherence of video generation while preserving image visual quality.
AntiStyler: Defending Object Detection Models Against Adversarial Patch Attacks Using Style Removal
Idan Yankelev (Ben Gurion University of Negev), Asaf Shabtai (Ben Gurion University of Negev)
Image TranslationObject DetectionAdversarial AttackImage
🎯 What it does: Proposed a fast zero-shot defense method called AntiStyler based on style removal to defend against adversarial patch attacks on object detection models.
ANTS: Adaptive Negative Textual Space Shaping for OOD Detection via Test-Time MLLM Understanding and Reasoning
Wenjie Zhu (Hong Kong Polytechnic University), Lei Zhang (Hong Kong Polytechnic University)
Anomaly DetectionTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposes a zero-shot, no-training OOD detection framework ANTS, which dynamically constructs an adaptive negative text space by generating expressive negative sentences and visual similarity negative labels during testing using a multimodal large language model (MLLM);
Any Resolution Any Geometry: From Multi-View To Multi-Patch
Wenqing Cui, Peter Wonka (Kaust)
Depth EstimationTransformerImage
🎯 What it does: Propose a unified multi-patch Transformer model, URGT, which jointly estimates high-quality depth maps and normal maps from a single high-resolution image.
Any2Any 3D Diffusion Models with Knowledge Transfer: A Radiotherapy Planning Study
Yuhan Wang (UC Santa Cruz), Riqiang Gao (Siemens Healthineers)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelAuto EncoderBiomedical DataComputed Tomography
🎯 What it does: Proposed DiffKT3D, an Any2Any (any-to-any) 3D diffusion framework for radiotherapy dose prediction in medical imaging.
Any4D: Unified Feed-Forward Metric 4D Reconstruction
Jay Karhade (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
Convolutional Neural NetworkTransformerSupervised Fine-TuningImageMultimodality
🎯 What it does: Proposed a unified feed-forward multi-view Transformer that can directly predict metric-scale dense 4D reconstruction (geometry + motion) from multi-frame images (optional additional modalities).
AnyDoc: Enhancing Document Generation via Large-Scale HTML/CSS Data Synthesis and Height-Aware Reinforcement Optimization
Jiawei Lin (Xi'an Jiaotong University), Christopher Tensmeyer (Adobe Research)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Propose the AnyDoc framework to achieve three document generation tasks (intent→document, document de-rendering, element→document), uniformly represented using HTML/CSS.
AnyID: Ultra-Fidelity Universal Identity-Preserving Video Generation from Any Visual References
Jiahao Wang (Xi'an Jiaotong University), Jieping Ye (Alibaba Cloud Computing)
GenerationTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderImageVideoText
🎯 What it does: Propose the AnyID framework to achieve high-fidelity identity-preserving video generation under multi-free-form references (portraits, likenesses, videos)
AnyLift: Scaling Motion Reconstruction from Internet Videos via 2D Diffusion
Hongjie Li (Stanford University), Jiajun Wu (Stanford University)
Data SynthesisPose EstimationTransformerDiffusion modelScore-based ModelVideo
🎯 What it does: Propose a two-stage AnyLift framework that utilizes a 2D diffusion model to recover 3D human motion and person-object interactions in world coordinates from internet videos under dynamic cameras.
AnyPcc: Compressing Any Point Cloud with a Single Universal Model
Kangli Wang (Peking University), Wei Gao (Peking University)
CompressionConvolutional Neural NetworkPoint Cloud
🎯 What it does: Proposes the AnyPcc framework, achieving geometric compression of arbitrary point clouds with a single unified model;
ApET: Approximation-Error Guided Token Compression for Efficient VLMs
Qiankun Ma (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Hairong Zheng (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
CompressionComputational EfficiencyVision Language ModelImageVideo
🎯 What it does: Proposed a visual token compression framework called ApET based on approximate error, which compresses visual tokens of VLMs without relying on attention mechanisms.
APEX: A Decoupled Memory-based Explorer for Asynchronous Aerial Object Goal Navigation
Daoxuan Zhang (Harbin Institute of Technology), Shuo Yang (Harbin Institute of Technology)
Object DetectionSegmentationRobotic IntelligenceGraph Neural NetworkReinforcement LearningVision Language ModelImage
🎯 What it does: Proposed APEX — an asynchronous parallel three-module architecture for UAV goal-oriented navigation, capable of efficiently memorizing, decision-making, and localizing targets in complex 3D environments