CVPR 2026 Papers with AI Summaries
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
→ CVPR 2026 papers with code (1047)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 4071 CVPR 2026 papers by keyword, author or institution —
free trial on arXivSub.
$\alpha$Matte4K & $\mu$Matting: Dataset and Model for Ultra-Micro Precision Alpha Video Matting
Xinyi Chen (East China Normal University), Haichuan Song (Kuaishou Technology)
SegmentationRecurrent Neural NetworkAuto EncoderVideoBenchmark
🎯 What it does: Proposed a two-stage high-resolution video matting framework called µ Matting and constructed a high-quality 4K portrait video matting dataset named α Matte4K.
$\oslash$ 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)
Domain AdaptationImageBiomedical Data
🎯 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.
$\phi$-DPO: Fairness Direct Preference Optimization Approach to Continual Learning in Large Multimodal Models
Thanh-Dat Truong (University of Arkansas), Khoa Luu (University of Arkansas)
OptimizationMeta LearningReinforcement Learning from Human FeedbackTransformerMultimodalityBenchmark
🎯 What it does: Proposes a continual learning framework ϕ-DPO for large-scale multimodal models, combining Fairness Direct Preference Optimization (Fairness DPO) to simultaneously alleviate catastrophic forgetting and fairness issues caused by data imbalance;
$L^{2}DGS$: Low-Light Dynamic Gaussian Splatting
Ashish Kumar (Indian Institute of Technology Madras), Rajagopalan N Ambasamduram (Indian Institute of Technology Madras)
RestorationGenerationGaussian SplattingVideoPoint Cloud
🎯 What it does: Proposed a fully self-supervised 4D Gaussian point cloud framework, L DGS 2, which can directly reconstruct complete bright dynamic scenes from low-light videos and synthesize arbitrary spatiotemporal views.
2-Shots in the Dark: Low-Light Denoising with Minimal Data Acquisition
Liying Lu (École Polytechnique Fédérale de Lausanne), Sabine Süsstrunk
RestorationData SynthesisImage
🎯 What it does: Proposes a method to synthesize low-light denoising training data using only one noisy image and one dark frame, generating diverse signal-agnostic noise through frequency domain spectral sampling;
240FPS Stereo Vision from Monocular Mixed Spikes
Yeliduosi Xiaokaiti (Peking University), Boxin Shi (Peking University)
Data SynthesisDepth EstimationImageVideo
🎯 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).
2D-LFM: Lifting Foundation Model without 3D Supervision
Mosam Dabhi (Carnegie Mellon University), Simon Lucey (Adelaide University)
Pose EstimationDepth EstimationTransformerImage
🎯 What it does: Investigated single-frame 3D structure recovery methods using only 2D keypoints without 3D annotations, and proposed the 2D-LFM model;
2ndMatch: Finetuning Pruned Diffusion Models via Second-Order Jacobian Matching
Caleb Zheng (University of Washington), Eli Shlizerman (University of Washington)
GenerationComputational EfficiencyKnowledge DistillationDiffusion modelScore-based ModelImage
🎯 What it does: Efficiently fine-tune diffusion models after sparse pruning to reduce quality degradation.
3D Gaussian Splatting at Arbitrary Resolutions with Compact Proxy Anchors
Mingyun Jeong (Hanyang University), Donghyeon Cho (Hanyang University)
GenerationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Propose a 3D Gaussian scattering framework capable of high-quality view synthesis at any resolution;
3D Gaussian Splatting from Unposed Spike Stream
Yijia Guo (Peking University), Tiejun Huang (Peking University)
Pose EstimationSpiking Neural NetworkGaussian SplattingSimultaneous Localization and MappingOptical FlowTime SeriesSequential
🎯 What it does: Propose a framework named Nope-SGS, capable of reconstructing high-quality 3D scenes and estimating camera trajectories from high-frequency spike image streams without pose priors.
3D Gaussian Splatting with Self-Constrained Priors for High Fidelity Surface Reconstruction
Takeshi Noda (Tsinghua University), Zhizhong Han (Wayne State University)
GenerationNeural Radiance FieldGaussian SplattingImage
🎯 What it does: Propose a self-constrained prior to improve high-precision surface reconstruction in 3D Gaussian Splatting (3DGS)
3D sans 3D Scans: Scalable Pre-training from Video-Generated Point Clouds
Ryousuke Yamada (AIST), Yuki M. Asano (University of Technology Nuremberg)
SegmentationRepresentation LearningTransformerVideoPoint Cloud
🎯 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;
3D Space as a Scratchpad for Editable Text-to-Image Generation
Oindrila Saha (University of Massachusetts Amherst), Kevin Blackburn-Matzen (Adobe Research)
GenerationData SynthesisTransformerLarge Language ModelAgentic AIPrompt EngineeringDiffusion modelImageText
🎯 What it does: Propose using 3D space as a scratchpad for editable text-to-image generation, leveraging LLM multi-agent systems to accomplish scene construction, arrangement, rotation, and camera selection, while achieving identity and depth-controlled image generation through SIGMA-Gen.
3D-Aware Implicit Motion Control for View-Adaptive Human Video Generation
Zhixue Fang (Kuaishou Technology), Kun Gai (Kuaishou Technology)
GenerationPose EstimationTransformerDiffusion modelFlow-based ModelVideoText
🎯 What it does: Proposed the 3DiMo framework, which combines an end-to-end implicit 3D motion encoder with a pre-trained video generator to achieve 3D motion control based on 2D-driven videos, and supports text-driven camera perspective switching.
3D-Aware Multi-Task Learning with Cross-View Correlations for Dense Scene Understanding
Xiaoye Wang (University of Cambridge), Wei-Hong Li (University of Bristol)
SegmentationDepth EstimationConvolutional Neural NetworkTransformerImageVideo
🎯 What it does: Propose a 3D-aware multi-task learning framework that enhances geometric consistency in multi-task dense prediction by leveraging cross-view association.
3D-Fixer: Coarse-to-Fine In-place Completion for 3D Scenes from a Single Image
Ze-Xin Yin (Nankai University), Jin Xie (Nanjing University)
GenerationDepth EstimationTransformerDiffusion modelImagePoint Cloud
🎯 What it does: Propose the 3D-Fixer framework, which utilizes partial geometric information from a single image to generate scene-level 3D assets, achieving high-quality single-view scene reconstruction.
3D-IDE: 3D Implicit Depth Emergent
Chushan Zhang (Australian National University), Hongdong Li (Australian National University)
RecognitionDepth EstimationTransformerLarge Language ModelContrastive LearningVideoMultimodalityPoint Cloud
🎯 What it does: By using implicit depth self-supervised training on RGB videos to train a single visual encoder, multi-modal large language models can achieve 3D perception and reasoning without relying on 3D inputs.
3D-LATTE: Latent Space 3D Editing from Textual Instructions
Maria Parelli (University of Tbingen), Andreas Geiger (University of Tbingen)
GenerationVision Language ModelDiffusion modelGaussian SplattingTextMesh
🎯 What it does: Proposed 3D-LATTE, a method for instruction-based 3D asset editing in the latent space of a 3D diffusion model, enabling high-quality, view-consistent editing of geometry and appearance.
3D-Object Perception Transformer (3PT)
Agastya Kalra (Intrinsic Innovation LLC), Huaijin Chen (Intrinsic Innovation LLC)
Object DetectionSegmentationPose EstimationRepresentation LearningTransformerContrastive LearningImageBenchmark
🎯 What it does: Proposes 3PT, a unified depth-free input 3D object-aware Transformer for zero-shot detection, segmentation, and 6DoF pose estimation.
3D-VCD: Hallucination Mitigation in 3D-LLM Embodied Agents through Visual Contrastive Decoding
Makanjuola Adekunmi Ogunleye (Virginia Tech), Ismini Lourentzou (University of Illinois Urbana Champaign)
Explainability and InterpretabilityComputational EfficiencyRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelContrastive LearningMultimodalityGraphBenchmark
🎯 What it does: Propose a model-free, inference-time 3D Visual Contrastive Decoding (3D-VCD) framework to eliminate hallucinations in 3D multi-modal large language models when performing tasks based on 3D scenes.
3DrawAgent: Teaching LLM to Draw in 3D with Early Contrastive Experience
Hongcan Xiao (Beijing University of Posts and Telecommunications), Yonggang Qi (Beijing University of Posts and Telecommunications)
GenerationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMesh
🎯 What it does: Proposes 3DrawAgent, a training-agnostic, large language model (LLM)-based 3D sketch generation framework that sequentially draws 3D Bézier curves through natural language instructions and achieves self-improvement via geometric feedback.
3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects
Zhicheng Liang (Chinese University of Hong Kong Shenzhen), Fangxin Wang (Chinese University of Hong Kong Shenzhen)
GenerationData SynthesisLarge Language ModelDiffusion modelNeural Radiance FieldGaussian SplattingImagePoint CloudBenchmark
🎯 What it does: Proposed 3DReflecNet, a large object-centric dataset containing over 120,000 synthetic instances, over 1,000 real-world scans, and 7 million views, along with a multi-task benchmark (image matching, SfM, NVS, reflection removal, relighting).
3M-TI: High-Quality Mobile Thermal Imaging via Calibration-free Multi-Camera Cross-Modal Diffusion
Minchong Chen (Shanghai Jiao Tong University), Jun Zhang (Tsinghua University)
RestorationObject DetectionSegmentationSuper ResolutionSupervised Fine-TuningPrompt EngineeringDiffusion modelAuto EncoderImageMultimodality
🎯 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.
4C4D: 4 Camera 4D Gaussian Splatting
Junsheng Zhou (Tsinghua University), Yu-Shen Liu (Tsinghua University)
GenerationData SynthesisGaussian SplattingVideo
🎯 What it does: Reconstruct high-quality 4D scenes from videos captured using only four cameras, enabling perspective synthesis at any arbitrary time point.
4D Local Modeling Toward Dynamic Global Perception for Ambiguity-free Rotation-Invariant Point Cloud Analysis
Jiaxun Guo (Concordia University), Nizar Bouguila (Concordia University)
ClassificationRecognitionSegmentationGraph Neural NetworkPoint Cloud
🎯 What it does: This paper proposes a novel framework called Ga4DPF, based on learnable rotatable and deformable transformations in 4D space, to achieve unambiguous rotation-invariant features in point cloud analysis.
4D Primitive-Mache: Glueing Primitives for Persistent 4D Scene Reconstruction
Kirill Mazur (Imperial College London), Andrew J. Davison (Imperial College London)
Object TrackingSegmentationPose EstimationOptical FlowVideo
🎯 What it does: Propose a 4D Primitive-Match (4DPM) framework that generates complete and sustainable 4D scene reconstruction by decomposing monocular video into rigid 3D primitives and estimating their SE(3) motion;
4D-RGPT: Toward Region-level 4D Understanding via Perceptual Distillation
Chiao-An Yang (NVIDIA), Min-Hung Chen (NVIDIA)
Knowledge DistillationTransformerLarge Language ModelVision Language ModelOptical FlowVideoTextMultimodalityBenchmark
🎯 What it does: Proposes 4D-RGPT, a multimodal large language model (LLM) specifically designed for regional 4D visual question answering, and enhances 4D perception capabilities through the P4D training framework.
4DEquine: Disentangling Motion and Appearance for 4D Equine Reconstruction from Monocular Video
Jin Lyu, Xiaoying Tang (Southern University Of Science And Technology)
GenerationData SynthesisTransformerDiffusion modelGaussian SplattingImageVideo
🎯 What it does: Propose a framework named 4DEquine for recovering 4D horses (including shape, motion, and appearance) from monocular videos, decomposing it into two submodules: motion recovery (AniMoFormer) and static appearance recovery (EquineGS).
4DP-QA: Scalable QA for 4D Perception in Vision Language Models
Seokju Cho (NVIDIA), Orazio Gallo (NVIDIA)
Object TrackingLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes an expandable 4D cognitive data generation pipeline, constructs a large-scale 4DP-QA dataset, and uses it to train vision-language models (VLMs) to enhance understanding of 4D (spatial + motion) scenes in videos.
4DSurf: High-Fidelity Dynamic Scene Surface Reconstruction
Renjie Wu (Australian National University), Miaomiao Liu (Australian National University)
GenerationGaussian SplattingVideo
🎯 What it does: Propose a dynamic scene surface reconstruction method using Gaussian Splatting, capable of generating temporally consistent and detail-rich 3D surfaces from sparse view videos.
4DWorldBench: A Comprehensive Evaluation Framework for 3D/4D World Generation Models
Yiting Lu (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
GenerationLarge Language ModelVision Language ModelOptical FlowImageVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes 4DWorldBench, a unified multimodal, physics-aware evaluation framework for systematic assessment of 3D/4D world generation models under image, video, and text conditions.
A Bit is All You Need! Efficient Video Capture via Single Bit Imaging
Kanchana Vaishnavi Gandikota (University of Siegen), Paramanand Chandramouli (Independent Researcher)
RestorationCompressionComputational EfficiencyTransformerVideo
🎯 What it does: This paper proposes a single-bit video capture framework that collects only 1-bit data at the sensor level using time-varying threshold quantization, and reconstructs it into high-bit depth video through a deep learning network.
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)
Domain AdaptationRepresentation LearningTransformerReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 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)
Safty and PrivacyExplainability and InterpretabilityLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 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.
A Closer Look at Cross-Domain Few-Shot Object Detection: Fine-Tuning Matters and Parallel Decoder Helps
Xuanlong Yu (Intellindust AI Lab), Di Yang (Suzhou Institute for Advanced Research, USTC)
Object DetectionTransformerSupervised Fine-TuningImageBenchmark
🎯 What it does: This paper proposes a hybrid ensemble decoder that integrates a shared layer with parallel decoding branches, combined with a staged fine-tuning framework, for cross-domain few-shot object detection.
A Combination of Noise and Bilateral Filters Achieve Supralinear and Scalable Adversarial Robustness in CNNs
Nicolas Stalder (Institute of Neuroinformatics ETH Zurich University of Zürich), Pau Vilimelis Aceituno (Institute of Neuroinformatics ETH Zurich University of Zürich)
Adversarial AttackConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes a simple preprocessor that combines Gaussian noise with multiple bilateral filters. Theoretically, it is proven that both components enhance robustness against adversarial attacks through complementary mechanisms, and they are uniformly applied during both training and inference phases;
A Cross-view Fusion Framework for Robust 6-DoF Grasp Pose Estimation
Kangjian Zhu (Nanjing University of Science and Technology), Jin Xie (Nanjing University of Science and Technology)
Pose EstimationRobotic IntelligenceConvolutional Neural NetworkTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposed a cross-view fusion framework based on auxiliary perspectives, which employs self-supervised contrastive learning to enforce spatial consistency and directional discriminability on point cloud features, and enhances the robustness of 6-DoF grasp pose estimation through a cross-view alignment cylindrical fusion module.
A Debiased Reconstruction-based Framework for Training-Free Detection of AI-Generated Images
Sungik Choi, Moontae Lee
GenerationAnomaly DetectionDiffusion modelAuto EncoderImageBenchmark
🎯 What it does: This paper proposes a training-free AI-generated image detection method that enhances discriminative ability by removing background and latent space biases.
A Difference-in-Difference Approach to Detecting AI-Generated Images
Xinyi Qi (Tsinghua University), Hongyi Zhou (Tsinghua University)
ClassificationAnomaly DetectionConvolutional Neural NetworkDiffusion modelGenerative Adversarial NetworkImage
🎯 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.
A Faster Path to Continual Learning
Wei Li (Sichuan University), Tao Feng (Tsinghua University)
OptimizationComputational EfficiencyImage
🎯 What it does: Proposed a C-Flat Turbo optimizer that accelerates training in continual learning tasks by reusing gradient components.
A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens
Tommie Kerssies (Amazon), Liang-Chieh Chen (Amazon)
GenerationComputational EfficiencyTransformerWorld ModelVideo
🎯 What it does: Propose a method that compresses the difference between consecutive frames into a single continuous token via DeltaToken, constructing the DeltaWorld generative world model, which can efficiently generate multiple feasible futures in a single forward pass.
A Geometric Algebra-Informed 3DGS Framework for Wireless Channel Prediction
Jingzhou Shen (Florida International University), Xuyu Wang (Florida International University)
TransformerGaussian SplattingPhysics Related
🎯 What it does: Developed a 3D Gaussian splatting framework based on geometric algebra (GAI-GS) for accurate wireless channel prediction.
A Mixed Diet Makes DINO An Omnivorous Vision Encoder
Rishabh Kabra (Google DeepMind), Niloy J. Mitra (University College London)
ClassificationSegmentationDepth EstimationRetrievalKnowledge DistillationRepresentation LearningTransformerContrastive LearningMultimodality
🎯 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 More Word-like Image Tokenization for MLLMs
Hyun Lee (Seoul National University), Joonseok Lee (Seoul National University)
Computational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposed a visual tokenizer called DiVT, which clusters Vision Transformer's patch embeddings into semantically coherent visual tokens, dynamically determines the number of tokens through similarity thresholds, and adjusts the granularity via threshold regulation.
A Multi-Agent Perception-Action Alliance for Efficient Long Video Reasoning
Yichang Xu (Georgia Institute of Technology), Ling Liu (Georgia Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelAgentic AIVision-Language-Action ModelVideoTextMultimodality
🎯 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 Polynomial Chaos Framework for Causal Discovery in Nonlinear Uncertain Systems
Liang Cao (University of British Columbia)
Explainability and InterpretabilityTabularPhysics Related
🎯 What it does: Proposed a causal discovery framework called PCELiNGAM that uses polynomial chaos expansion (PCE) to represent random noise, enabling the identification of causal structures and quantification of uncertainty in nonlinear, uncertain industrial systems.
A Provable Energy-Guided Test-Time Defense Boosting Adversarial Robustness of Large Vision-Language Models
Mujtaba Hussain Mirza (Sapienza University of Rome), Iacopo Masi (Sapienza University of Rome)
Safty and PrivacyComputational EfficiencyAdversarial AttackVision Language ModelScore-based ModelImageMultimodality
🎯 What it does: Proposed an energy-guided test-time defense method called ET3 to enhance the robustness of large vision-language models (LVLM) against adversarial attacks.
A Sanity Check for Multi-In-Domain Face Forgery Detection in the Real World
Jikang Cheng (Peking University), Ling Liang (Peking University)
Anomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImageBenchmark
🎯 What it does: Proposed a multi-domain face forgery detection framework called DevDet, addressing the issue where domain differences overwhelm authenticity differences in real-time frame detection when the domain is unspecified.
A Self-Conditioned Representation Guided Diffusion Model for Realistic Text-to-LiDAR Scene Generation
Wentao Qu, Liang Xiao (Njust)
Data SynthesisAutonomous DrivingConvolutional Neural NetworkVision Language ModelDiffusion modelTextPoint CloudBenchmark
🎯 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 Semantically Disentangled Unified Model for Multi-category 3D Anomaly Detection
SuYeon Kim (Kyung Hee University), MyeongAh Cho (Kyung Hee University)
Anomaly DetectionTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposes a unified 3D anomaly detection framework called SeDiR, which first obtains category-aware global representations through coarse-to-fine global tokenization, then achieves semantic disentanglement via category-conditional contrastive learning, and finally realizes semantically consistent geometric reconstruction through a geometry-guided decoder, thereby enabling anomaly detection and localization for multi-class objects within a unified model.
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)
Representation LearningTransformerVideo
🎯 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 Style is Worth One Code: Unlocking Code-to-Style Image Generation with Discrete Style Space
Huijie Liu (Beihang University), Guoliang Kang (Beihang University)
GenerationTransformerVision Language ModelDiffusion modelContrastive LearningImageText
🎯 What it does: Developed the CoTyle framework to achieve image style generation based on numerical style codes, supporting three modes: code-to-style, image-conditioned, and style interpolation.
A Supervised Multi-task Framework for Joint cryo-ET Restoration Enabled by Generative Physical Simulation
Xinsheng Wang (Beijing Institute of Technology), Bin Hu (Beijing Institute of Technology)
RestorationData SynthesisConvolutional Neural NetworkSupervised Fine-TuningGenerative Adversarial NetworkImageBiomedical DataComputed Tomography
🎯 What it does: Propose a fully supervised multi-task framework cryoDeRec that jointly performs denoising and missing wedge completion, directly recovering high-quality 3D reconstructions from raw low SNR cryo-ET volume data.
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)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Propose a Temporal and Content Co‑Awareness (TCCA) framework based on latent diffusion models, which generates geometrically accurate and visually consistent hand images under given target hand poses and reference appearances.
A Training-Free Style-Personalization via SVD-Based Feature Decomposition
Kyoungmin Lee (DGIST), Sunghoon Im (DGIST)
GenerationTransformerImageTextBenchmark
🎯 What it does: This paper proposes a training-free method for personalized image generation, which directly controls the style of generated images during inference using a single style reference image while maintaining semantic consistency.
A Unified Framework for Knowledge Transfer in Bidirectional Model Scaling
Jianlu Shen (Southeast University), Xin Geng (Southeast University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerImageText
🎯 What it does: This paper proposes a unified bidirectional knowledge transfer framework called BoT, which enables parameter-free and training-cost-free weight initialization between source models and target models of different scales.
A Unified Perspective on Adversarial Membership Manipulation in Vision Models
Ruize Gao (National University of Singapore), Feng Liu (Chinese University of Hong Kong)
Adversarial AttackConvolutional Neural NetworkImage
🎯 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)
RetrievalTransformerContrastive LearningImage
🎯 What it does: This paper proposes a new visual place recognition (VPR) method called A2GC‑VPR, which integrates asymmetric optimal transport aggregation and geometric constraints;
A3: Towards Advertising Aesthetic Assessment
Kaiyuan Ji (Shanghai Artificial Intelligence Laboratory), Guangtao Zhai
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 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;
Abstract 3D Perception for Spatial Intelligence in Vision-Language Models
Yifan Liu (Tsinghua University), Hanspeter Pfister (Massachusetts Institute of Technology)
Object DetectionSegmentationDepth EstimationVision Language ModelDiffusion modelImageVideo
🎯 What it does: Propose SandboxVLM, a no-training framework that injects abstract 3D bounding boxes into existing vision-language models (VLMs), leveraging video diffusion to generate multi-view perspectives, depth estimation, 2D segmentation and projection, clustering, and other steps to construct a 3D Sandbox, thereby enhancing the spatial reasoning capabilities of VLMs.
Accelerating Autoregressive Video Diffusion via History-Guided Cache and Residual Correction
Kepan Nan (Nanjing University), Ying Tai (Nanjing University)
GenerationComputational EfficiencyDiffusion modelVideoText
🎯 What it does: Propose a training-agnostic cache acceleration framework named ARCache for autoregressive video diffusion models (ARDM), consisting of two modules: history-guided cache (HGC) and enhanced residual correction (ERC).
Accelerating Diffusion Model Training under Minimal Budgets: A Condensation-Based Perspective
Rui Huang (Hong Kong University of Science and Technology), Zeke Xie (Hong Kong University of Science and Technology)
GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelImage
🎯 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.
Accelerating Diffusion via Hybrid Data-Pipeline Parallelism Based on Conditional Guidance Scheduling
Euisoo Jung, Jae-Gil Lee (Korea Advanced Institute Of Science And Technology)
GenerationComputational EfficiencyTransformerDiffusion modelImageText
🎯 What it does: Proposes a hybrid parallel framework that integrates conditional base partitioning with adaptive pipeline switching to accelerate diffusion model inference.
Accelerating Diffusion-based Video Editing via Heterogeneous Caching: Beyond Full Computing at Sampled Denoising Timestep
Tianyi Liu (Nanyang Technological University), Lap-Pui Chau (Hong Kong Polytechnic University)
GenerationComputational EfficiencyTransformerDiffusion modelVideoBenchmark
🎯 What it does: Propose a training-agnostic heterogeneous cache framework HetCache for video editing tasks with diffusion Transformer (DiT), achieving significant acceleration through selective caching at both time-step and token levels.
Accelerating Streaming Video Large Language Models via Hierarchical Token Compression
Yiyu Wang (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoBenchmark
🎯 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.
ACE-Merging: Data-Free Model Merging with Adaptive Covariance Estimation
Bo Xu (Hong Kong University of Science and Technology Guangzhou), Chengwei Qin (Hong Kong University of Science and Technology Guangzhou)
Knowledge DistillationRepresentation LearningTransformerMixture of ExpertsImageText
🎯 What it does: Propose a data-agnostic model merging framework ACE-Merging, which merges multi-task expert models in the weight space by leveraging covariance information from fine-tuning updates.
AceTone: Bridging Words and Colors for Conditional Image Grading
Tianren Ma (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
Image TranslationGenerationLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: Propose AceTone, a generative multi-modal color grading system based on 3D LUT;
ACoT-VLA: Action Chain-of-Thought for Vision-Language-Action Models
Linqing Zhong (Beihang University), Guanghui Ren (AgiBot)
Robotic IntelligenceTransformerVision-Language-Action ModelFlow-based ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the Vision-Language-Action model ACoT-VLA based on action space reasoning, which generates coarse-grained reference trajectories using the Explicit Action Reasoner (EAR) and extracts implicit action priors through the Implicit Action Reasoner (IAR), thereby guiding the final action prediction.
ACPV-Net: All-Class Polygonal Vectorization for Seamless Vector Map Generation from Aerial Imagery
Weiqin Jiao (University of Twente), Claudio Persello (University of Twente)
GenerationDiffusion modelImage
🎯 What it does: This paper proposes an end-to-end ACPV-Net capable of automatically generating panoramic, seamless, and topologically consistent multi-class vector maps in a single inference.
Act Like a Pathologist: Tissue-Aware Whole Slide Image Reasoning
Wentao Huang (Stony Brook University), Chen Wang (Mayo Clinic)
SegmentationTransformerLarge Language ModelVision Language ModelMultimodalityBiomedical DataBenchmark
🎯 What it does: For the path vision question answering task, HistoSelect is proposed, which utilizes problem-guided tissue segmentation and patch selection to extract only the tissue and patches relevant to the question for input into a multimodal LLM, significantly reducing the number of visual tokens.
Act2See: Emergent Active Visual Perception for Video Reasoning
Martin Q. Ma (Carnegie Mellon University), Louis-Philippe Morency (Carnegie Mellon University)
Large Language ModelSupervised Fine-TuningAgentic AIVision Language ModelDiffusion modelVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 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.
ActAvatar: Temporally-Aware Precise Action Control for Talking Avatars
Ziqiao Peng (Renmin University of China), Jun He (Renmin University of China)
GenerationTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelDiffusion modelFlow-based ModelVideoTextMultimodalityAudio
🎯 What it does: Achieve temporal action control over conversational avatars through structured text prompts while maintaining audio synchronization.
Action Motifs: Self-Supervised Hierarchical Representation of Human Body Movements
Genki Kinoshita (Kyoto University), Ko Nishino (Kyoto University)
Pose EstimationRepresentation LearningTransformerVideoSequential
🎯 What it does: Propose A4Mer, a self-supervised nested Transformer that learns hierarchical representations of human motion: Action Atoms (atomic actions) and Action Motifs (action phrases), achieving hierarchical learning through masked prediction of joint embeddings.
Action-Geometry Prediction with 3D Geometric Prior for Bimanual Manipulation
Chongyang Xu (Sichuan University), Shuaicheng Liu (University of Electronic Science and Technology of China)
Robotic IntelligenceTransformerDiffusion modelImagePoint CloudBenchmark
🎯 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.
Action-Sketcher: From Reasoning to Action via Visual Sketches for Robotic Manipulation
Huajie Tan (Peking University), Shanghang Zhang (Beijing Academy of Artificial Intelligence)
Robotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelFlow-based ModelImageVideoTextMultimodality
🎯 What it does: Proposed a framework called Action-Sketcher that uses visual sketches (Visual Sketch) to bridge high-level reasoning and low-level actions in robot manipulation.
ActionMesh: Animated 3D Mesh Generation with Temporal 3D Diffusion
Remy Sabathier (Meta Reality Labs), Tom Monnier (Meta Reality Labs)
GenerationData SynthesisDiffusion modelAuto EncoderImageVideoTextMultimodalityMeshBenchmark
🎯 What it does: Propose ActionMesh, an end-to-end, fast, skeleton-free, and topologically consistent generative model that can generate animated 3D meshes from multimodal inputs such as text, images, videos, or 3D meshes + text;
Activation Matters: Test-time Activated Negative Labels for OOD Detection with Vision-Language Models
Yabin Zhang (Harbin Institute of Technology), Curtis Langlotz (Stanford University)
Anomaly DetectionContrastive LearningMultimodality
🎯 What it does: Proposes a Test-time Activated Negative Labels (TANL) method to dynamically select negative labels with high activation for out-of-distribution (OOD) samples during inference, thereby improving OOD detection performance.
Active Inference for Micro-Gesture Recognition: EFE-Guided Temporal Sampling and Adaptive Learning
Weijia Feng (Tianjin Normal University), Minglai Shao (Tianjin University)
RecognitionConvolutional Neural NetworkReinforcement LearningVideo
🎯 What it does: This paper proposes a micro-gesture recognition framework called UAAI based on active inference, which utilizes Expected Free Energy (EFE)-guided temporal and spatial sampling, as well as uncertainty-aware gain (UMIX), to achieve active observation and adaptive learning of micro-gestures.
Active Intelligence in Video Avatars via Closed-loop World Modeling
Xuanhua He (Hong Kong University of Science and Technology), Qifeng Chen (Hong Kong University of Science and Technology)
Robotic IntelligenceLarge Language ModelPrompt EngineeringVision-Language-Action ModelWorld ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes the L-IVA benchmark for multi-step tasks and designs the ORCA framework, enabling video avatars to autonomously achieve long-term goals in a generated environment.
Active Perceptual Inference: A Corticothalamic-Inspired Dynamic Nested Recurrent Network for Multimodal Sentiment Analysis with Incomplete Data
Yujuan Zhang (Beijing Institute of Technology), Xia Wu (Beijing Institute of Technology)
ClassificationRecurrent Neural NetworkTransformerMultimodality
🎯 What it does: Propose a brain-inspired dynamic nested recurrent network (DNRNet), which achieves active perception reasoning and completion of frame-level missing data in multimodal sentiment analysis through local and global loops.
ActiveAD: Planning-Oriented Active Learning for End-to-End Autonomous Driving
Han Lu (Shanghai Jiao Tong University), Junchi Yan (COWAROBOT Co. Ltd.)
Autonomous DrivingImagePoint Cloud
🎯 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).
ActiveGrasp: Information-Guided Active Grasping with Calibrated Energy-based Model
Boshu Lei (University of Pennsylvania), Kostas Daniilidis (University of Pennsylvania)
Robotic IntelligenceGaussian SplattingImagePoint Cloud
🎯 What it does: Developed an active grasping framework that generates grasping distributions in SE(3) space using a calibrated energy model, and selects the next optimal viewpoint based on the entropy reduction of this distribution.
ActivePolicy: Active Gaussian Reconstruction and Optimization Strategy Based on Global-Local Information Gain
Yingzhao Li (Harbin Institute of Technology), Lijun Zhao (Harbin Institute of Technology)
GenerationDepth EstimationOptimizationRobotic IntelligenceGraph Neural NetworkGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Proposes a rendering-driven active 3D Gaussian reconstruction framework named ActivePolicy, which can automatically select the optimal camera pose (NBV) when exploring unknown environments and simultaneously improve geometric completeness and rendering quality at both global and local levels.
ActiveVLA: Injecting Active Perception into Vision-Language-Action Models for Precise 3D Robotic Manipulation
Zhenyang Liu (Fudan University), Yanwei Fu (Fudan University)
Robotic IntelligenceVision-Language-Action ModelImageBenchmark
🎯 What it does: Designed a framework called ActiveVLA that integrates active perception into vision-language-action models, enabling the localization of key 3D regions under multi-view settings and actively selecting optimal viewpoints and zooming perspectives to achieve fine manipulation.
ActivityForensics: A Comprehensive Benchmark for Localizing Manipulated Activity in Videos
Peijun Bao (Zhejiang University), Xudong Jiang (Nanyang Technological University)
Anomaly DetectionTransformerDiffusion modelVideoTextMultimodalityBenchmark
🎯 What it does: This paper proposes the first task and dataset for video activity-level forgery localization called ActivityForensics, and presents a diffusion-based feature regularization baseline named Temporal Artifact Diffuser (TADiff).
AcTTA: Rethinking Test-Time Adaptation via Dynamic Activation
Hyeongyu Kim (Yonsei University), Dosik Hwang (Yonsei University)
Domain AdaptationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes AcTTA, a framework that achieves model adaptation during testing through learnable activation functions.
AD-GBC: Anisotropic Granular-Ball Skip-Connection Refiner for UNet-Based Medical Image Segmentation
Xiya Shen (Macau University of Science and Technology), Li Feng (Macau University of Science and Technology)
SegmentationConvolutional Neural NetworkBiomedical DataUltrasound
🎯 What it does: Introduce the AD-GBC module into the UNet structure, utilizing learnable anisotropic granular-ball (spherical) for reconstructing semantic skip connections. Pixel features are first aggregated to the regional level and then broadcasted back to pixels, enhancing semantic consistency and boundary precision in medical image segmentation.
AdaBet: Gradient-free Layer Selection for Efficient Training of Deep Neural Networks
Irene Tenison (Massachusetts Institute of Technology), Mohammad Malekzadeh (Nokia Bell Labs)
ClassificationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 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;
AdaCluster: Adaptive Query-Key Clustering for Sparse Attention in Video Generation
Haoyue Tan (University of Science and Technology of China), Cheng Li (University of Science and Technology of China)
GenerationComputational EfficiencyTransformerDiffusion modelVideo
🎯 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.
AdaDexTrack: Dynamic Modulation for Adaptive and Generalizable Dexterous Manipulation Tracking
Jianibieke Adalibieke (Shanghai Qi Zhi Institute), Li Yi (Tsinghua University)
Object TrackingDomain AdaptationRobotic IntelligenceLarge Language ModelReinforcement LearningDiffusion modelTextSequential
🎯 What it does: Propose AdaDexTrack, which employs a closed-loop modulator to achieve adaptive tracking of language-guided hand-object interaction trajectories
AdaIAT: Adaptively Increasing Attention to Generated Text to Alleviate Hallucinations in LVLM
Li'an Zhong (Sun Yat-Sen University), Xiangui Kang (Foshan University)
Large Language ModelVision Language ModelImageMultimodality
🎯 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.
AdapAction: Adaptive Target Action Backdoor Attack against GUI Agents
Baicheng Chen (Chinese University of Hong Kong), Baoyuan Wu (Chinese University of Hong Kong)
Knowledge DistillationAdversarial AttackLarge Language ModelSupervised Fine-TuningMultimodality
🎯 What it does: This paper proposes AdapAction, an adaptive backdoor attack mechanism targeting GUI agents based on multimodal large language models, which can dynamically select malicious behaviors after activation based on the current screen state and user instructions.
AdaPrior: Bayesian-Inspired Adaptive Prior Correction for Long-Tailed Continual Learning
S Divakar Bhat (Honda R&D Japan), Bhuvan Aggarwal (Honda R&D Japan)
ClassificationImage
🎯 What it does: Propose AdaPrior, a Bayesian-inspired adaptive prior correction method, to address the prior drift problem in long-tail class incremental learning (LTCIL), combining the AdaPrior Loss during training with a lightweight posterior correction during inference, forming a complete single-stage solution without additional training steps.
Adapter Shield: A Unified Framework with Built-in Authentication for Preventing Unauthorized Zero-Shot Image-to-Image Generation
Jun Jia (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
GenerationSafty and PrivacyAdversarial AttackTransformerDiffusion modelImage
🎯 What it does: Propose Adapter Shield, a pluggable, password-authentication-supported unified protection framework for preventing unauthorized zero-copy image-to-image (I2I) diffusion model generation;
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)
Data SynthesisTransformerDiffusion modelImageBiomedical Data
🎯 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)
RetrievalDomain AdaptationTransformerSupervised Fine-TuningMixture of ExpertsVision Language ModelDiffusion modelContrastive LearningImageText
🎯 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.
Adapting Lightweight Image-based Counting Models for Video Crowd Counting
Weibo Shu (City University of Hong Kong), Antoni B. Chan (City University of Hong Kong)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkVideo
🎯 What it does: Studies how to migrate lightweight image-based crowd counting models to video scenarios, proposing the use of statistical regularization to leverage spatiotemporal information without adding extra modules, achieving real-time inference.
Adapting Point Cloud Analysis via Multimodal Bayesian Distribution Learning
Xingyu Zhu (University of Science and Technology of China), Hanwang Zhang (Nanyang Technological University)
ClassificationRecognitionDomain AdaptationComputational EfficiencyRepresentation LearningLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodalityPoint Cloud
🎯 What it does: Propose a training-free, real-time point cloud analysis method called BayesMM, which can adaptively recognize point clouds without requiring retraining.
Adaptive 3D Perception for Small Aerial Targets Under Sparse Sampling via Reinforcement Learning
Shenghai Yuan (Nanyang Technological University), Enwen Hu (Beijing University of Posts and Telecommunications)
Object DetectionObject TrackingReinforcement LearningPoint Cloud
🎯 What it does: Proposes an adaptive 3D perception framework based on reinforcement learning, A3PRL, for detecting and tracking small aerial targets (SAT) under sparse long-range LiDAR
Adaptive Action Chunking at Inference-time for Vision-Language-Action Models
Yuanchang Liang (National University of Singapore), Prahlad Vadakkepat (National University of Singapore)
Computational EfficiencyRobotic IntelligenceVision-Language-Action ModelDiffusion modelFlow-based ModelMultimodality
🎯 What it does: Propose an AAC strategy that adaptively determines action block size during inference based on action entropy, applied to Vision-Language-Action models.