These 1047 CVPR 2026 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every CVPR 2026 paper, free trial on arXivSub.
$\oslash$ Source Models Leak What They Shouldn't $\nrightarrow$: Unlearning Zero-Shot Transfer in Domain Adaptation Through Adversarial Optimization
Arnav Devalapally (Indian Institute of Technology, Hyderabad), Vineeth N. Balasubramanian (Indian Institute of Technology, Hyderabad)
CodeDomain AdaptationImageBiomedical Data
π― What it does: Proposes a method called SCADA-UL that simultaneously achieves machine forgetting for source-exclusive categories in source-free domain adaptation, unifying forgetting and adaptation through adversarial optimization and remarking strategies.
π― What it does: Proposed a fully self-supervised 4D Gaussian point cloud framework, Lβ―DGSβ―2, which can directly reconstruct complete bright dynamic scenes from low-light videos and synthesize arbitrary spatiotemporal views.
Yeliduosi Xiaokaiti (Peking University), Boxin Shi (Peking University)
CodeData SynthesisDepth EstimationImageVideo
π― What it does: Design and implement a monocular hybrid spiking camera system by adding an LCD temporal domain modulation to a single eye, capturing hybrid light signals using a high-frame-rate spiking camera, and then obtaining 240FPS stereo video through baseline least squares decoupling and subsequent deep learning reconstruction (SMS-Net).
π― What it does: 3D self-supervised pretraining using unlabeled indoor video-generated point clouds (VGPC), constructing the RoomTours dataset with 49,000 scenes, and training the LAM3C model based on this dataset;
π― What it does: Propose 3M-TI, a calibration-free multi-camera cross-modal diffusion framework designed to enhance the resolution and texture quality of mobile thermal imaging.
A Causal Marriage between VLM and IRM from Understanding to Reasoning
Ziliang Chen (Research Institute of Multiple Agents and Embodied Intelligence, Peng Cheng Laboratory), Liang Lin (Research Institute of Multiple Agents and Embodied Intelligence, Peng Cheng Laboratory)
CodeDomain AdaptationRepresentation LearningTransformerReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper applies token-level causal representation theory to equate CLIP's contrastive learning objective with IRM, proposing a mid-training scheme that injects invariant learning signals into pre-trained CLIP to obtain the CLIP-IRM model; subsequently, the invariant alignment score from CLIP-IRM is used as a process-level reward to guide reinforcement learning inference in multi-modal large language models (MLLM), achieving cross-domain robustness for out-of-distribution (OOD) understanding and reasoning.
A Closed-Form Solution for Debiasing Vision-Language Models with Utility Guarantees Across Modalities and Tasks
Tangzheng Lian (King's College London), Oya Celiktutan (King's College London)
CodeSafty and PrivacyExplainability and InterpretabilityLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Proposes a training- and data-agnostic visual language model debiasing method with a closed-form solution, achieving fairness while maintaining efficacy in cross-modal tasks.
π― What it does: For AI-generated image detection, the Differential-Differential (DID) method is proposed: first, use a pre-trained diffusion model to reconstruct the input image and calculate the first-order reconstruction error; then use this reconstructed image to reconstruct again, obtaining the second-order error; finally, use two classifiers to jointly discriminate the authenticity of the image.
π― What it does: Add a few Adapters to the frozen DINOv2 encoder, leveraging contrastive learning and knowledge distillation to achieve unified alignment of multi-modal features such as RGB, Depth, and Segmentation, forming a cross-modal 'omnivorous' visual encoder.
A Multi-Agent Perception-Action Alliance for Efficient Long Video Reasoning
Yichang Xu (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
CodeComputational EfficiencyTransformerLarge Language ModelAgentic AIVision-Language-Action ModelVideoTextMultimodality
π― What it does: Propose a multi-agent perception-action alliance (A4VL) framework that employs multi-round perception exploration and action exploration iterations to achieve efficient long video question answering.
A Self-Conditioned Representation Guided Diffusion Model for Realistic Text-to-LiDAR Scene Generation
Wentao Qu, Liang Xiao (Njust)
CodeData SynthesisAutonomous DrivingConvolutional Neural NetworkVision Language ModelDiffusion modelTextPoint CloudBenchmark
π― What it does: Proposed T2LDM, a self-conditioned representation-guided text-to-LiDAR scene generation diffusion model, addressing the issues of generated smoothness and directional confusion caused by the scarcity of text-LiDAR pairs.
A Stitch in Time: Learning Procedural Workflow via Self-Supervised Plackett-Luce Ranking
Chengan Che (King's College London), Luis C. Garcia-Peraza-Herrera (King's College London)
CodeRepresentation LearningTransformerVideo
π― What it does: Propose PL-Stitch, a self-supervised learning framework that learns procedural representations from surgery and cooking videos by leveraging Plackett-Luce ranking tasks and spatiotemporal jigsaw puzzles.
A Temporal and Content Co-Awareness Latent Diffusion for Controllable Hand Image Generation
Shuang Hao (Beijing University of Posts and Telecommunications), Jingyu Wang (Beijing University of Posts and Telecommunications)
CodeGenerationData SynthesisTransformerVision Language ModelDiffusion modelImageMultimodality
π― What it does: Propose a Temporal and Content CoβAwareness (TCCA) framework based on latent diffusion models, which generates geometrically accurate and visually consistent hand images under given target hand poses and reference appearances.
π― What it does: This paper proposes and systematizes adversarial exploitation against membership inference attacks in visual models, defining Membership Fabrication Attack (MFA), Membership Fabrication Detection (MFD), and Adversarial Robust Membership Inference (AR-MIA), and demonstrates their effectiveness on multiple standard datasets.
A2GC: Asymmetric Aggregation with Geometric Constraints for Locally Aggregated Descriptors
Zhenyu Li (Qilu University of Technology), Tianyi Shang (Fuzhou University)
CodeRetrievalTransformerContrastive LearningImage
π― What it does: This paper proposes a new visual place recognition (VPR) method called A2GCβVPR, which integrates asymmetric optimal transport aggregation and geometric constraints;
Kaiyuan Ji (Shanghai Artificial Intelligence Laboratory), Guangtao Zhai
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Built an advertising aesthetics evaluation framework A3, including the A-Law three-stage assessment method, A3 Dataset with 120K aligned data, the A-Align multimodal large language model, and the A-Bench benchmark;
π― What it does: Proposes a two-stage diffusion model dataset condensation framework named D2C, aiming to significantly accelerate diffusion model training under extremely low data budget conditions.
π― What it does: Proposes a hybrid parallel framework that integrates conditional base partitioning with adaptive pipeline switching to accelerate diffusion model inference.
CodeCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoBenchmark
π― What it does: To address the real-time inference bottleneck in streaming video large language models (VideoLLMs), the STC (Streaming Token Compression) framework is proposed. It reduces visual encoding costs through caching and dynamic recomputation, and achieves efficient inference by compressing the LLM's preceding sequence via dual anchor pruning.
Act2See: Emergent Active Visual Perception for Video Reasoning
Martin Q. Ma (Carnegie Mellon University), Louis-Philippe Morency (Carnegie Mellon University)
CodeLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelDiffusion modelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the ACT2SEE framework, enabling Vision-Language Models (VLMs) to actively retrieve or generate frames in video reasoning and enhance inference quality through supervised fine-tuning.
π― What it does: Propose a dual-arm robot control framework based on a pre-trained 3D geometric foundation model, which extracts 3D latent representations from RGB images and fuses them with 2D semantic features and robot pose. A conditional diffusion network is used to simultaneously predict future action segments and 3D point clouds, enhancing spatial perception and coordination in dual-arm manipulation.
ActiveAD: Planning-Oriented Active Learning for End-to-End Autonomous Driving
Han Lu (Shanghai Jiao Tong University), Junchi Yan (COWAROBOT Co. Ltd.)
CodeAutonomous DrivingImagePoint Cloud
π― What it does: Propose a planning-oriented active learning framework, ActiveAD, to select the most valuable samples in end-to-end autonomous driving (E2E-AD), significantly reducing the need for expensive annotations (3D boxes, semantic segmentation).
π― What it does: Propose AdaBet, a gradient-agnostic hierarchical and channel selection method for efficiently performing transfer learning on pre-trained deep networks in resource-constrained devices;
π― What it does: This paper proposes a training-agnostic adaptive query-key clustering framework called AdaCluster, for sparse attention acceleration in Diffusion Transformer video generation models while maintaining generation quality.
AdaIAT: Adaptively Increasing Attention to Generated Text to Alleviate Hallucinations in LVLM
Li'an Zhong (Sun Yat-Sen University), Xiangui Kang (Foshan University)
CodeLarge Language ModelVision Language ModelImageMultimodality
π― What it does: Propose two adaptive attention enhancement methods, IAT and AdaIAT, which significantly reduce hallucinations in large-scale vision-language models by leveraging visual information from generated text.
Adapting a Pre-trained Single-Cell Foundation Model to Spatial Gene Expression Generation from Histology Images
Donghai Fang (Sun Yat-sen University), Wenwen Min (Yunnan University)
CodeData SynthesisTransformerDiffusion modelImageBiomedical Data
π― What it does: Adapt the pre-trained single-cell foundation model (sc-FM) into a generative model that produces spatial gene expression from H&E tissue images.
Adapting In-context Generation for Enhanced Composed Image Retrieval
Haiwen Li (Beijing University of Posts and Telecommunications), Fei Su (Beijing University of Posts and Telecommunications)
CodeRetrievalDomain AdaptationTransformerSupervised Fine-TuningMixture of ExpertsVision Language ModelDiffusion modelContrastive LearningImageText
π― What it does: Adapt text-to-image models using a small amount of labeled data to generate unbiased synthetic query-target image triplets, and employ two-stage training to improve CIR performance.
Tong Lin (Xi'an Jiaotong University), Xing Wei (Xi'an Jiaotong University)
CodeObject TrackingDiffusion modelVideo
π― What it does: Propose ARTrack-AC, an adaptive capacity autoregressive visual tracking framework that dynamically adjusts inference capacity to balance accuracy and speed;
π― 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.
π― 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.
π― 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)
CodeCompressionOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Propose the AdaptVision framework, which utilizes tool calls to achieve coarse-to-fine adaptive visual token acquisition, significantly reducing token consumption;
π― 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;
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)
CodeClassificationDomain AdaptationTransformerSupervised Fine-TuningVision Language ModelImageBiomedical Data
π― 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.
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)
CodeSegmentationAnomaly DetectionTransformerLarge Language ModelVision Language ModelImageMultimodalityBiomedical Data
π― 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.
Air-Know: Arbiter-Calibrated Knowledge-Internalizing Robust Network for Composed Image Retrieval
Zhiheng Fu (Shandong University), Zixu Li (Shandong University)
CodeRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
π― 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).
π― 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)
CodeClassificationDepth EstimationExplainability and InterpretabilityKnowledge DistillationTransformerContrastive LearningImage
π― 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)
CodeRobotic IntelligenceLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringVision-Language-Action ModelImageTextMultimodalityBenchmark
π― 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;
π― 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.
An Efficient Token Compression Framework for Visual Object Tracking
Weijing Wu (Guangxi Normal University), Shuxiang Song (Guangxi Normal University)
CodeObject TrackingTransformerVideo
π― 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);
Anatomica: Localized Control over Geometric and Topological Properties for Anatomical Diffusion Models
Karim Kadry (MIT), Elazer R. Edelman (MIT)
CodeGenerationDiffusion modelImageBiomedical Data
π― 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.
π― 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.
π― 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.
π― 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)
CodeAnomaly DetectionTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
π― 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);
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)
CodeCompressionComputational EfficiencyVision Language ModelImageVideo
π― 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.
APPO: Attention-guided Perception Policy Optimization for Video Reasoning
Henghui Du (Renmin University of China), Di Hu (Renmin University of China)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodality
π― What it does: This paper proposes an attention-guided perception policy optimization algorithm called APPO, aimed at enhancing the fine-grained perception capabilities of multimodal large language models in video reasoning tasks.
Are We Ready for RL in Text-to-3D Generation? A Progressive Investigation
Yiwen Tang (Shanghai AI Laboratory), Bin Zhao (Shanghai AI Laboratory)
CodeGenerationLarge Language ModelReinforcement LearningTextMeshBenchmark
π― What it does: This paper proposes and systematically evaluates the feasibility of applying reinforcement learning (RL) to text-to-3D autoregressive generation, and based on this, develops a novel hierarchical RL framework Hi-GRPO and an RL-enhanced model AR3D-R1; meanwhile, it introduces the MME-3DR benchmark focusing on implicit reasoning capabilities.
π― What it does: Propose a heterogeneous RGB-event stereo fusion framework named ARES, which jointly estimates optical flow, disparity, and scene flow.
ARGUS: Defending Against Multimodal Indirect Prompt Injection via Steering Instruction-Following Behavior
Weikai Lu (South China University Of Technology), Hao Peng (Beihang University)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
π― What it does: This study proposes ARGUS, a defense framework against multi-modal indirect injection attacks (IPI) based on activation space control, which includes the construction of a cross-image, video, and audio IPI benchmark, controllability analysis of instruction-following behavior, activation steering, injection detection, and post-processing mechanisms.
π― What it does: This paper proposes an autoregressive framework called ARMFlow based on MeanFlow for online 3D human reaction generation, and designs an offline version named ReMFlow;
Attack for Defense: Adversarial Agents for Point Prompt Optimization Empowering Segment Anything Model
Xueyu Liu (Taiyuan University of Technology), Yongle Chen (Taiyuan University of Technology)
CodeSegmentationAdversarial AttackTransformerReinforcement LearningPrompt EngineeringImageBiomedical Data
π― What it does: Proposed Point Prompt Defender (PPD), which automatically optimizes the point prompts of the Segment Anything Model through an attack-defense dual-agent reinforcement learning framework.
π― What it does: Designed and implemented a retrieval-based AI-generated image attribution framework called LIDA, which uses 'generated fingerprints' produced from low-bit planes as input. After unsupervised pre-training and few-shot adaptation, the framework can efficiently attribute and detect deepfakes for unknown generators.
AudioStory: Generating Long-Form Narrative Audio with Large Language Models
Yuxin Guo (University of Chinese Academy of Sciences), Wei Zou (University of Chinese Academy of Sciences)
CodeGenerationTransformerLarge Language ModelDiffusion modelFlow-based ModelAudio
π― What it does: This paper proposes AudioStory, an end-to-end framework that integrates large language models (LLM) with audio diffusion models (DiT) to generate long-form narrative audio, achieving instruction following, temporal coherence, and emotional consistency;
Authorize-on-Demand: Dynamic Authorization with Legality-Aware Intellectual Property Protection for VLMs
Lianyu Wang (Key Laboratory of Brain Machine Intelligence Technology Ministry of Education), Daoqiang Zhang (Key Laboratory of Brain Machine Intelligence Technology Ministry of Education)
CodeClassificationDomain AdaptationSafty and PrivacyVision Language ModelContrastive LearningImage
π― What it does: Propose a dynamic authorization VLM (Vision-Language Model) intellectual property protection framework named AoD-IP, which supports on-demand switching of authorization domains and enables legitimacy judgment.
π― What it does: Proposed the BrepARG framework, which unifies the geometric and topological information of B-rep into an overall token sequence, and completes end-to-end generation through autoregressive Transformer.
AVION: Aerial Vision-Language Instruction from Offline Teacher to Prompt-Tuned Network
Yu Hu (University of British Columbia Okanagan), Mohsen Zardadi (TerraSense Analytics)
CodeClassificationRetrievalKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This study proposes a knowledge distillation framework called AVION for efficiently transferring large vision-language models to remote sensing image tasks;
AXG-Reasoner: Error Detection and Explanation in Long Task Videos with Vision-Language Models
Shih-Po Lee (Northeastern University), Ehsan Elhamifar (Northeastern University)
CodeAnomaly DetectionExplainability and InterpretabilityGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelVision-Language-Action ModelVideo
π― What it does: Propose the AXG-Reasoner framework, combining a frozen Vision-Language Model (VLM) with an automatically constructed Action Execution Graph (AXG) and a Time Action Segmentation (TAS) model, to detect and explain errors in long-task videos.
b-CLIP: Text-Conditioned Contrastive Learning for Multi-Granular Vision-Language Alignment
Fatimah Zohra (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
CodeRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Propose a multi-grained text conditional contrastive learning framework Ξ²-CLIP, which aligns hierarchically decomposed text (complete sentences, clauses, phrases) with visual image regions and introduces Ξ²-Contextualized Contrastive Alignment Loss (Ξ²-CAL) to regulate semantic overlap.
π― What it does: Propose the BTP framework, which utilizes pre-trained point language models to directly perform zero-shot 3D anomaly detection and localization on raw point clouds.
π― What it does: In fine-grained object detection of remote sensing images, a hierarchical label structure is utilized to propose balanced hierarchical contrastive learning and query decoupling strategies to enhance detection performance.
BAMI: Training-Free Bias Mitigation in GUI Grounding
Borui Zhang (Tsinghua University), Jiwen Lu (Tsinghua University)
CodeObject DetectionLarge Language ModelPrompt EngineeringImage
π― What it does: Propose a training-agnostic reasoning method called BAMI to improve the accuracy of GUI localization models in high-resolution, element-dense interfaces;
π― What it does: This paper proposes Batch Loss Score (BLS), which utilizes the mean batch loss available at each step to infer sample importance through sample-specific exponential moving average (EMA), achieving lossless dynamic data pruning;
Batman: Benign Knowledge Alignment Through Malicious Null Space in Federated Backdoor Attack
Wenwen He (Wuhan University), Mang Ye (Nanyang Technological University)
CodeFederated LearningAdversarial AttackImage
π― What it does: Propose a new backdoor attack method in federated learning called Batman, which can maintain a high attack success rate while improving stealthiness.
π― What it does: Proposed and made public SurgCleanβthe first real-world multi-type endoscopic surgery image restoration benchmark datasetβand constructed a unified evaluation framework based on it. Systematic evaluation of 22 common restoration algorithms was conducted, further analyzing structural differences between surgical and natural images and their impacts on depth estimation and semantic segmentation.
Better, Stronger, Faster: Tackling the Trilemma in MLLM-based Segmentation with Simultaneous Textual Mask Prediction
Jiazhen Liu (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
CodeSegmentationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: Proposed an All-mask prediction paradigm, where STAMP generates text during the dialogue phase and subsequently predicts the entire image's segmentation mask in one parallel step, achieving unification of semantic dialogue and pixel-level segmentation.
π― What it does: Propose the BEV-SLD method, which learns scene-specific keypoints (scene landmarks) in LiDAR BEV images through self-supervised learning, achieving global localization without dense maps;
Beyond Duality: A Hybrid Framework of Leveraging Shared and Private Features for RGB-Event Object Detection
Keyao Wang (Hebei University Of Technology), Haiyong Chen (Hebei University Of Technology)
CodeObject DetectionTransformerMultimodality
π― What it does: Propose the SPFD network, which utilizes frequency domain consistency to separate shared and private features, and adaptively fuses them in the TriAdapt Encoder and TriInject Decoder to achieve RGB-Event object detection;
Beyond Euclidean Gossip: KL-Barycentric Consensus on Heterogeneous and Imbalanced Images
Lu Xu (University of Hong Kong), Guosheng Yin (University of Hong Kong)
CodeClassificationSegmentationFederated LearningImageBiomedical Data
π― What it does: Propose a decentralized deep learning framework based on KL barycenter natural gradient variational inference, which can achieve model aggregation in non-i.i.d. and imbalanced sample environments.
π― What it does: Proposes a learnable linear prediction (LPΒ²) framework for feature caching and prediction in Diffusion Transformers (DiT), enabling training-free inference acceleration.
Beyond Heuristic Prompting: A Concept-Guided Bayesian Framework for Zero-Shot Image Recognition
Hui Liu (City University of Hong Kong), Haoliang Li (City University of Hong Kong)
CodeClassificationRecognitionLarge Language ModelVision Language ModelImage
π― What it does: Propose a concept-based Bayesian framework (CGBC) that leverages large language models (LLMs) to generate class-specific concepts and adjusts their weights through an adaptive soft clipping likelihood function, achieving zero-shot image classification.
CodeOptimizationExplainability and InterpretabilityComputational EfficiencySupervised Fine-TuningReinforcement LearningVision Language ModelVideoMultimodality
π― What it does: Proposed the VideoThinker framework for lightweight multimodal language model (MLLM) video reasoning, addressing the issue of models easily falling into perceptual shortcuts and limited reasoning capabilities during reinforcement learning fine-tuning.
Beyond Single Images: A Comprehensive Benchmark for Album-Level Vision-Language Understanding
Shawn Huang (Brigham Young University), Bryan Morse (Brigham Young University)
CodePrompt EngineeringVision Language ModelImageMultimodalityBenchmark
π― What it does: Proposes AlbumBench as a multi-image understanding benchmark for personal photo albums, defining four tasks: intent selection, intent scoring, group labeling, and group clustering.
Beyond Static Frames: Temporal Aggregate-and-Restore Vision Transformer for Human Pose Estimation
Hongwei Fang (Zhejiang Gongshang University), Wenwu Yang (Zhejiang Gongshang University)
CodePose EstimationTransformerVideo
π― What it does: Proposed a video-based 2D human pose estimation framework named TAR-ViTPose, which can incorporate temporal modeling while maintaining the pure Transformer architecture of ViTPose.
π― What it does: Propose a pairing training method for infrared and visible images without relying on strict alignment, aiming for high-performance fusion.
π― What it does: This paper proposes a satellite image block adjustment method that does not rely on sparse matching points, directly optimizing global correction parameters through dense feature consistency.
π― What it does: In traditionally trained deep networks, structured pruning and learnable masks are used to directly extract bias-invariant subnetworks from within the model without requiring retraining or additional bias-invariant data.
π― What it does: Propose the Bi-CMPStereo bidirectional cross-modal prompting framework, which uses event-frame heterogeneous stereo vision to achieve high-precision disparity estimation
BiGain: Unified Token Compression for Joint Generation and Classification
Jiacheng Liu (VILA Lab, Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (VILA Lab, Mohamed bin Zayed University of Artificial Intelligence)
π― What it does: To address the acceleration problem of diffusion models, the BiGain framework is proposed, which can significantly improve the classification performance of the same model while maintaining generation quality.
Bilevel Layer-Positioning LoRA for Real Image Dehazing
Yan Zhang (Sun Yat-sen University), Zhuo Su (Sun Yat-sen University)
CodeRestorationDomain AdaptationVision Language ModelImage
π― What it does: Propose an unsupervised semantic guidance loss based on CLIP (H2C) and a two-level optimization strategy for LoRA layer localization (BiLaLoRA), for efficient image dehazing from synthetic domains to real fog scenes;
π― What it does: In the underwater instance segmentation task, this paper proposes the BiPA method, which utilizes dual-layer optimization to learn specialized dense prompts and effectively transfers SAM to the underwater domain through dual-modal prompts (dense + sparse) and a foreground attention injection module (FAI).
π― What it does: For real-world restoration tasks, we propose BiProLoRA: an adaptation paradigm that combines self-supervised distribution consistency learning (DFL) with a two-layer Prompt-LoRA to achieve high-fidelity restoration from synthetic data to real data.
π― What it does: Propose a matching-based Bidirectional Interaction Transformation Network (BIT), addressing the modality gap and data imbalance in visible-infrared person re-identification through bidirectional interaction and adaptive local matching between visible and infrared image pairs.
Black-box Membership Inference Attacks on the Pre-training Data of Image-generation Models
Tao Qi (Beijing University of Posts and Telecommunications), Yongfeng Huang (Tsinghua University)
CodeSafty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageText
π― What it does: Propose the SD-MIA framework for black-box membership inference attacks targeting pre-training data of diffusion-based image generation models.
BlackMirror: Black-Box Backdoor Detection for Text-to-Image Models via Instruction-Response Deviation
Feiran Li (Institute of Information Engineering, CAS), Qingming Huang (School of Computer Science and Technology, University of Chinese Academy of Sciences)
CodeAnomaly DetectionSafty and PrivacyLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Proposed a black-box backdoor detection framework called BlackMirror for text-to-image models, which can identify various backdoor attacks without accessing the model's internal structure.
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelVideoTextMultimodality
π― What it does: Analyze the internal mechanisms of large multimodal models (LMMs) post-trained with RLVR (Reinforcement Learning with Verifiable Rewards), and propose the Activation Replay method that does not require additional training. This method enhances reasoning capabilities by adding learnable noise to visual tokens during testing, making low-entropy activations of the RLVR model closer to those of the original LMM.
π― What it does: Proposed the cross-domain incremental object detection (CDIOD) benchmark and the Dynamic Group Subspace (DGS) framework to achieve cross-domain incremental learning.
π― What it does: Propose a Dual Granularity Alignment (DGA) framework for visual reprogramming in the CLIP model, combining visual hierarchy (multi-scale images) and semantic hierarchy (label hierarchy) to enhance classification performance in downstream tasks.
π― What it does: This paper proposes a Bootstrapping-based multi-view learning framework, BML, to address the test-time view correspondence noise (TNC) problem. It directly generates noisy samples on the training set through adaptive masking and view shuffling, and employs a lightweight estimator to perform supervised learning on the reliability of each view, thereby achieving robust multi-view fusion.
π― What it does: Pre-trained image semantic segmentation models are converted into temporal semantic segmentation models that can adapt using only a few video frames through test-time distillation and lightweight attention fusion.
π― What it does: Proposes DynUAVβa benchmark for multi-object tracking under complex high-speed maneuvers from a UAV perspective, containing 42 video clips, 1.7 million annotated frames, covering eight categories including vehicles, pedestrians, and industrial machinery.
CodeExplainability and InterpretabilityVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
π― What it does: This work proposes a no-training, plug-and-play dual-path decoding framework called PND, which dynamically balances visual likelihood and language prior during inference through positive and negative decoding paths, suppressing hallucinations in multimodal models.
π― What it does: Propose CODSR, a controllable one-step diffusion network for recovering high-quality images from low-quality images, balancing structural fidelity and perceptual quality.
Bridging RGB and Hematoxylin Components: An Interleaved Guidance and Fusion Framework for Point Supervised Nuclei Segmentation
Zihan Huan (Guilin University of Electronic Technology), Zhenbing Liu (Guilin University of Electronic Technology)
CodeSegmentationConvolutional Neural NetworkBiomedical Data
π― What it does: This paper proposes a point-supervised nucleus segmentation framework called DFGNet based on dual representations (RGB and hemocyanin components), which can achieve high-precision instance segmentation using sparse point annotations.