CVPR 2026 Papers — Page 35
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
Subspace Alignment for CLIP-based Continual Learning via Canonical Correlation Analysis
Huan Zhang (Wuhan University), Fan Lyu (Computer Vision Center Universitat Autonoma De Barcelona)
Representation LearningSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose a subspace alignment framework called CCA-CL based on Canonical Correlation Analysis to address the visual-textual asymmetric drift problem that occurs in CLIP during continuous learning.
SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling
Camile Lendering (Eindhoven University of Technology), Egor Bondarau
Anomaly DetectionTransformerImage
🎯 What it does: Propose a training-free few-shot anomaly detection method called SubspaceAD, which leverages local features extracted from a frozen DINOv2 vision foundation model. These features are used to fit a low-dimensional subspace via PCA, with reconstruction residuals serving as anomaly scores;
SunFaded: Illumination-Aware Gaussian Splatting for Dark Scenes with Camera-Mounted Active Lighting
Wenjie Chang (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
RestorationGenerationDepth EstimationGaussian SplattingImage
🎯 What it does: Propose a method that uses Gaussian Splatting to construct a lighting-agnostic scene representation when capturing images with a moving light source in dark environments, achieving unlit scene reconstruction and view synthesis by separating illumination from the intrinsic color of the scene.
SuP: Sub-cloud Driven Point Cloud Registration
Sheldon Fung (University of Western Australia), Xuequan Lu (University of Western Australia)
Pose EstimationConvolutional Neural NetworkTransformerPoint Cloud
🎯 What it does: Propose a sub-cloud based point cloud registration framework SuP, transforming low-overlap registration into a problem of mining high-overlap sub-cloud pairs;
Superman: Unifying Skeleton and Vision for Human Motion Perception and Generation
Xinshun Wang (Peking University), Mengyuan Liu (Peking University)
GenerationPose EstimationTransformerLarge Language ModelMultimodalitySequential
🎯 What it does: Propose a unified generative framework called Superman, which can simultaneously accomplish 3D pose estimation, motion prediction, and motion interpolation;
Suppressing Non-Semantic Noise in Masked Image Modeling Representations
Martine Hjelkrem-Tan (University of Oslo), Adín Ramírez Rivera (University of Oslo)
ClassificationSegmentationRepresentation LearningImage
🎯 What it does: This paper analyzes the non-semantic noise that occurs in masked image modeling (MIM) self-supervised learning and proposes a post-processing method called SOAP (Semantically Orthogonal Artifact Projection) to suppress this noise, thereby improving the performance of downstream tasks under zero-shot scenarios.
SURF: Signature-Retained Fast Video Generation
Kaixin Ding (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
GenerationTransformerDiffusion modelFlow-based ModelVideo
🎯 What it does: Propose the SURF framework to accelerate high-resolution video generation while preserving signature features such as layout, semantics, and motion from pre-trained models to the maximum extent.
SurgCoT: Advancing Spatiotemporal Reasoning in Surgical Videos through a Chain-of-Thought Benchmark
Gui Wang (Shenzhen University), Linlin Shen (Shenzhen University)
Object DetectionObject TrackingTransformerLarge Language ModelVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes SurgCoT, a fine-grained spatiotemporal reasoning benchmark spanning seven surgical categories and 35 surgical procedures, constructed through three-stage chain reasoning and quintuple annotations (Question, Options, Knowledge, Clues, Answer) to build structured reasoning chains.
SV-GS: Sparse View 4D Reconstruction with Skeleton-Driven Gaussian Splatting
Jun-Jee Chao (University of Minnesota), Volkan Isler (University of Texas at Austin)
GenerationData SynthesisDiffusion modelGaussian SplattingVideoMesh
🎯 What it does: SV-GS proposes a method for four-dimensional reconstruction using skeletal-driven Gaussian splatting under sparse temporal observations.
SVAgent: Storyline-guided Long Video Understanding via Cross-Modal Multi-Agent Collaboration
Zhongyu Yang (Heriot Watt University), Yingfang Yuan (Heriot Watt University)
TransformerLarge Language ModelAgentic AIVision Language ModelVideoTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed SVAgent, a multi-agent framework, achieving long-video question answering through storyline-driven cross-modal collaborative reasoning.
SVBench: Evaluation of Video Generation Models on Social Reasoning
Wenshuo Peng (Tsinghua University), Kaipeng Zhang (Shanda AI Research)
GenerationData SynthesisExplainability and InterpretabilityAgentic AIPrompt EngineeringVision Language ModelVideoTextBenchmark
🎯 What it does: Proposed SVBench, the first social reasoning benchmark for video generation models, designed with multi-dimensional evaluations based on 30 classic psychology experiments.
SVHalluc: Benchmarking Speech-Vision Hallucination in Audio-Visual Large Language Models
Chenshuang Zhang (KAIST), Tae-Hyun Oh (KAIST)
Anomaly DetectionLarge Language ModelVideoMultimodalityBenchmarkAudio
🎯 What it does: Proposed the SVHalluc benchmark to assess the hallucination behavior of audio-visual large language models during speech and vision alignment, and evaluate existing models.
SWIFT: Sliding Window Reconstruction for Few-Shot Training-Free Generated Video Attribution
Chao Wang (University Of Science And Technology Of China), Kejiang Chen (University Of Science And Technology Of China)
ClassificationAnomaly DetectionAuto EncoderVideo
🎯 What it does: This paper proposes SWIFT, a few-shot untrained video attribution method based on sliding window reconstruction, for identifying whether a video originates from a specified generative model;
SwiftTailor: Efficient 3D Garment Generation with Geometry Image Representation
Phuc Pham (Qualcomm), Phong Nguyen (Qualcomm)
GenerationTransformerImageMesh
🎯 What it does: Propose a two-stage framework SwiftTailor, which first uses a lightweight Vision-Language model called PatternMaker to generate 2D sewing patterns, and then employs GarmentSewer to map the patterns into a unified UV space Garment Geometry Image, directly decoding it into a 3D garment mesh;
SwiftVLA: Unlocking Spatiotemporal Dynamics for Lightweight VLA Models at Minimal Overhead
Chaojun Ni (Peking University), Wenjun Mei (Peking University)
Computational EfficiencyKnowledge DistillationRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelImageMultimodality
🎯 What it does: Proposes SwiftVLA, a lightweight vision-language-action (VLA) model, which leverages a pre-trained 4D visual geometry transformer, time caching, and Fusion Tokens to achieve 4D spatiotemporal modeling of 2D images. The model incorporates 4D knowledge through a mask-and-reconstruct strategy, and during inference, the 4D branch can be removed to maintain extremely low computational overhead.
SwitchCraft: Training-Free Multi-Event Video Generation with Attention Controls
Qianxun Xu (Westlake University), Chi Zhang (Westlake University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelVideoText
🎯 What it does: Introduce SwitchCraft, a training-agnostic multi-event video generation framework that achieves temporal control over multiple events by precisely modulating cross-attention queries during the inference stage.
SymphoMotion: Joint Control of Camera Motion and Object Dynamics for Coherent Video Generation
Guiyu Zhang (Chinese University of Hong Kong), Li Jiang (Chinese University of Hong Kong)
GenerationDepth EstimationTransformerVision Language ModelDiffusion modelVideoPoint Cloud
🎯 What it does: Propose the SymphoMotion framework to achieve joint control of camera motion and object dynamics, generating spatially consistent videos that align with camera and object paths.
Symphony: A Cognitively-Inspired Multi-Agent System for Long-Video Understanding
Haiyang Yan (Institute of Automation, Chinese Academy of Sciences), Mengyi Liu (Kuaishou Technology)
RecognitionRetrievalTransformerLarge Language ModelAgentic AIVision Language ModelVideoRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a multi-agent system called Symphony for long video understanding, simulating human cognition for task decomposition and collaboration
SyncDreamer: Controllable and Expressive Avatar Generation Beyond the Talking Head
Fatemeh Nazarieh (University of Surrey), Muhammad Awais (University of Surrey)
GenerationData SynthesisTransformerReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageVideoTextMultimodalityAudio
🎯 What it does: Generate identity-preserving, emotionally expressive, and controllable talking-head or full-body animations in an end-to-end manner using a single reference image, audio, and text prompts.
SynCLIP: Synonym-Coherent Language-Image Pretraining for Robust Open-Vocabulary Dense Perception
Mingjie Xie (Beihang University), Yue Deng (Beihang University)
Object DetectionSegmentationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the SynCLIP framework to address the issue of inconsistent localization caused by synonyms in open-vocabulary dense perception based on CLIP, employing two modules: SSA (Semantic Consistent Spatial Attention Alignment) and SAR (Spatial Attention Refinement) for synonym consistency learning and fine-grained processing;
SyncMos: Scalable Motion Synchronisation for Multi-Agent Scene Interaction
Lingxiao Li (University of Melbourne), Taehyun Rhee (University of Melbourne)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelTextMeshGraphBenchmark
🎯 What it does: Propose the SyncMos framework, leveraging a single-agent diffusion model and LLM planning to achieve time-synchronized motion generation for multi-agent 3D scene interactions.
Synergistic Bleeding Region and Point Detection in Laparoscopic Surgical Videos
Jialun Pei (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
Object DetectionSegmentationTransformerOptical FlowVideoBiomedical Data
🎯 What it does: Proposes a dual-task online detection framework for simultaneous detection of blood cavity regions and blood spots in laparoscopic surgery videos, and constructs the first public blood cavity detection dataset SurgBlood.
SynMotion: Semantic-Visual Adaptation for Motion Customized Video Generation
Shuai Tan (University of Hong Kong), Hengshuang Zhao (Huazhong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Proposes a motion-customized video generation framework, SynMotion, integrating semantic and visual dual embeddings, capable of learning actions from a few example videos and transferring them to diverse subjects.
Synthesizing Visual Concepts as Vision-Language Programs
Antonia Wüst (TU Darmstadt), Kristian Kersting (TU Darmstadt)
GenerationData SynthesisExplainability and InterpretabilityVision Language ModelMultimodality
🎯 What it does: Propose the Vision-Language Programs (VLP) framework, which compiles structured visual descriptions generated by vision-language models (VLMs) into neural symbolic programs, inducing visual rules from a few image examples and performing reasoning.
Synthetic Curriculum Reinforces Compositional Text-to-Image Generation
Shijian Wang (Southeast University), Cunjian Chen (Monash University)
GenerationData SynthesisLarge Language ModelReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose the CompGen framework, which utilizes synthetic data with controllable difficulty generated through scene graph generation for reinforcement learning, thereby enhancing the combinatorial synthesis capability of text-to-image (T2I) models in handling multiple objects, attributes, and relationships.
Synthetic Object Compositions for Scalable and Accurate Learning in Detection, Segmentation, and Grounding
Weikai Huang (University of Washington), Ranjay Krishna (University of Washington)
Image HarmonizationObject DetectionSegmentationData SynthesisLarge Language ModelImageText
🎯 What it does: This paper proposes the SOC (Synthetic Object Compositions) data synthesis pipeline, which randomly generates 3D positions and camera configurations, and uses generative harmonization and area-weighted blending to compose 20 million high-quality segmented synthetic objects into 2 million scene images, while automatically generating precise masks, bounding boxes, and referential expressions.
SynthRGB-T: Language-Vision Guided Image Translation for Diversity Synthesis
Jiangang Ding (Chang'an University), Wei Li (Chang'an University)
Image TranslationTransformerVision Language ModelDiffusion modelMultimodality
🎯 What it does: Designed and implemented the SynthRGB-T framework to achieve bidirectional, diverse translation between infrared and visible light images, capable of generating high-fidelity, controllable cross-modal images based on user text prompts.
T2SGrid: Temporal-to-Spatial Gridification for Video Temporal Grounding
Chaohong Guo (South China University of Technology), Chengjiang Long (Bytedance Inc)
RetrievalTransformerSupervised Fine-TuningVision Language ModelVideoText
🎯 What it does: This paper proposes the Temporal-to-Spatial Gridification (T2SGrid) method, which concatenates short temporal frame windows of videos into 2D grid images, leveraging the spatial attention of Vision-LLM to achieve video temporal grounding.
TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective
Chaohu Liu (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)
Data-Centric LearningTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodalityTabular
🎯 What it does: Propose the TableMix framework, which significantly improves the performance of MLLMs on table reasoning tasks by mixing three types of data (multimodal table reasoning, pure text mathematical reasoning, table perception) and employing difficulty-aware reward shaping (DRS).
Tackling Alignment Ambiguity in Person Retrieval through Conversational Attribute Mining
Hao Zou (University Of Electronic Science And Technology Of China), Mingzhu Cai (University Of Electronic Science And Technology Of China)
RetrievalTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes the CECA framework, which utilizes multimodal large language models for conversational attribute mining to enhance the alignment between text and images with fine-grained attribute information, thereby improving the performance of text-to-image person retrieval.
Tackling Model Bias via Game-theoretic Multi-agent Collaboration Framework for Hateful Meme Classification
Yiwei Wei (Tianjin University), Longbiao Wang (Tianjin University)
ClassificationTransformerLarge Language ModelReinforcement LearningAgentic AIVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes GECO, a game theory-based multi-agent collaboration framework for addressing model bias in hate meme classification.
TACO: Task-Aware Contrastive Learning for Joint LiDAR Localization and 3D Object Detection
Leyuan Xing (Xiamen University), Cheng Wang (Xiamen University)
Object DetectionAutonomous DrivingConvolutional Neural NetworkContrastive LearningPoint Cloud
🎯 What it does: Propose the TACO framework to achieve joint learning of LiDAR localization and 3D object detection.
TacSIm: A Dataset and Benchmark for Football Tactical Style Imitation
Peng Wen (Capital University of Physical Education And Sports), Qiurui Wang (Capital University of Physical Education And Sports)
Object DetectionObject TrackingConvolutional Neural NetworkReinforcement LearningAuto EncoderVideoTime SeriesBenchmark
🎯 What it does: Constructed the TacSIm dataset and benchmark by reconstructing complete 11 vs. 11 player trajectories from Premier League broadcast videos, providing a unified tactical replication evaluation framework.
TAG-MoE: Task-Aware Gating for Unified Generative Mixture-of-Experts
Yu Xu (University of Chinese Academy of Sciences), Fan Tang (University of Chinese Academy of Sciences)
GenerationTransformerMixture of ExpertsVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: Proposed a task-aware sparse expert network, TAG-MoE, for unifying image generation and editing tasks.
TagSplat: Topology-Aware Gaussian Splatting for Dynamic Mesh Modeling and Tracking
Hanzhi Guo (Beijing Institute of Technology), Chenyu Xu (Soul Shell Technology Co Ltd)
GenerationPose EstimationGaussian SplattingVideoMesh
🎯 What it does: Propose a topology-aware dynamic reconstruction framework based on Gaussian scattering, which can generate topologically consistent mesh sequences and achieve precise 3D keypoint tracking.
TAlignDiff: Automatic Tooth Alignment assisted by Diffusion-based Transformation Learning
Yunbi Liu (Nanjing University of Posts and Telecommunications), Qingshan Liu (Nanjing University of Posts and Telecommunications)
Diffusion modelContrastive LearningPoint Cloud
🎯 What it does: Developed TAlignDiff, an automatic tooth alignment system that combines a point cloud regression network and a diffusion probability model to predict tooth transformation matrices, achieving precise tooth alignment.
Talk2Move: Reinforcement Learning for Text-Instructed Object-Level Geometric Transformation in Scenes
Jing Tan (AWS Agentic AI), Stefano Soatto (AWS Agentic AI)
Reinforcement LearningVision-Language-Action ModelDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes Talk2Move, a text-guided object geometry transformation framework based on reinforcement learning, which can achieve spatial editing of objects in images, such as translation, rotation, and scaling.
Talking Together: Synthesizing Co-Located 3D Conversations from Audio
Mengyi Shan (University of Washington), Zeng Huang (University of Rochester)
GenerationData SynthesisPrompt EngineeringDiffusion modelTextMeshAudio
🎯 What it does: Generate complete 3D facial animations for two co-located characters from a mixed voice stream, including precise lip synchronization, head pose, and relative spatial relationships.
TALO: Pushing 3D Vision Foundation Models Towards Globally Consistent Online Reconstruction
Fengyi Zhang (University of Queensland), Yadan Luo (University of Queensland)
Autonomous DrivingOptimizationSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: To maintain global geometric consistency across submaps in online 3D vision foundation models (e.g., VGGT, π3, MapAnything), this paper proposes TALO, an alignment framework based on Thin Plate Spline (TPS) and global control points, and introduces a point-agnostic submap registration method based on camera poses.
TALON: Test-time Adaptive Learning for On-the-Fly Category Discovery
Yanan Wu (China Agricultural University), Zhenbo Li (Concordia University)
ClassificationRecognitionDomain AdaptationTransformerSupervised Fine-TuningContrastive LearningImageBenchmark
🎯 What it does: To address the challenge of identifying known classes and discovering unknown classes in online unlabeled streams, this paper proposes the TALON framework for test-time adaptive learning, which continuously updates the feature extractor and class prototypes during inference to enable immediate learning of newly emerging classes.
TAMER: A Tri-Modal Contrastive Alignment and Multi-Scale Embedding Refinement Framework for Zero-Shot ECG Diagnosis
Xuewei Zhou (Wuhan University of Science and Technology), Junlin Xu (Wuhan University of Science and Technology)
Anomaly DetectionRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImageTextMultimodalityTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Proposes a tri-modal self-supervised learning framework named TAMER for zero-shot ECG diagnosis, jointly processing ECG signals, spectrograms, and clinical reports.
Taming Generative Diffusion Model for Task-Oriented Infrared Imaging
Tengyu Ma (Dalian University of Technology), Risheng Liu (Dalian University of Technology)
RestorationObject DetectionSegmentationDiffusion modelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Proposed a task-oriented infrared imaging framework that models infrared image recovery as a single-step diffusion process, combining dynamic clock estimation with spectral regularization to achieve high-quality visual recovery and semantic consistency;
Taming Noise-Induced Prototype Degradation for Privacy-Preserving Personalized Federated Fine-Tuning
Yuhua Wang (Beihang University), Zhiming Zheng (Beihang University)
Federated LearningSafty and PrivacyKnowledge DistillationTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a client privacy plugin VPDR for ProtoPFL, which includes variance-based adaptive prototype perturbation (VPP) and knowledge distillation-guided soft clipping regularization (DCR), significantly improving the accuracy and robustness of personalized federated fine-tuning without compromising LDP security.
Taming Preference Mode Collapse via Directional Decoupling Alignment in Diffusion Reinforcement Learning
Chubin Chen (Tsinghua University), Xiu Li (Tsinghua University)
GenerationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelImageTextBenchmark
🎯 What it does: Proposed a directional decoupled alignment framework named D²-Align to correct reward signals during the reinforcement learning alignment process in text-to-image diffusion models, effectively suppressing preference mode collapse (PMC) while maintaining generation diversity; simultaneously constructed a new multi-dimensional diversity evaluation benchmark called DivGenBench.
Taming Sampling Perturbations with Variance Expansion Loss for Latent Diffusion Models
Qifan Li (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
GenerationDiffusion modelAuto EncoderImage
🎯 What it does: Propose a 'Variance Expansion Loss' (VE Loss) that constructs a more robust latent space against sampling perturbations without relying on KL regularization, thereby improving the generation quality of latent diffusion models.
Taming the Long Tail: Rebalancing Adversarial Training via Adaptive Perturbation
Lilin Zhang (Sichuan University), Xianggen Liu (Sichuan University)
ClassificationAdversarial AttackImage
🎯 What it does: For adversarial training under long-tailed data, adaptive perturbation rebalancing is used to adjust the training distribution.
Taming Video Models for 3D and 4D Generation via Zero-Shot Camera Control
Chenxi Song (Westlake University), Chi Zhang (Westlake University)
GenerationData SynthesisDepth EstimationDiffusion modelAuto EncoderOptical FlowImageVideo
🎯 What it does: Proposes WorldForge, a framework that generates three-dimensional/four-dimensional scenes and controls camera trajectories without requiring retraining, achieving all tasks during the inference phase.
TANGO: Learning Distribution-wise Foundation Prior Consistency and Instance-wise Style Calibration for Medical Image Generalization
Chuang Liu (Beihang University), Haogang Zhu (Beihang University)
SegmentationDomain AdaptationConvolutional Neural NetworkTransformerContrastive LearningImageBiomedical Data
🎯 What it does: Proposed the TanGo framework, combining distribution consistency learning during training and instance style adaptive calibration during testing, achieving continuous test-time adaptation (CTTA) and single test-time adaptation (TTA) for medical image segmentation tasks.
TANGO: Text-Anchored Guided Optimization for Robust Fine-tuning Vision-Language Models under Label Noise
Tengfei Ma (Huazhong University of Science and Technology), Wei Wei (Huazhong University of Science and Technology)
ClassificationOptimizationRepresentation LearningData-Centric LearningSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Leveraging pure semantic anchors generated from text, using the text encoder of VLM as an independent supervisory source to replace the traditional linear classifier, the TANGO framework is proposed to achieve robust refined learning;
TAP: A Token-Adaptive Predictor Framework for Training-Free Diffusion Acceleration
Haowei Zhu (Tsinghua University), Bin Wang (Tsinghua University)
GenerationComputational EfficiencyDiffusion modelImageVideo
🎯 What it does: Proposes the Token-Adaptive Predictor (TAP) framework, which utilizes a lightweight probe in the first layer of the full model to dynamically select the most suitable predictor for each token at each sampling step, thereby accelerating diffusion model inference.
TAPE: Task-Adaptive Prototype Evolution in Audio-Language Models for Fully Few-shot Class-incremental Audio Classification
Yunlong Gao (Dalian University of Technology), Xinyue Liu (Dalian University of Technology)
ClassificationRecognitionTransformerVision Language ModelAudio
🎯 What it does: The study proposes a task-adaptive prototype evolution framework based on an audio-language model to address catastrophic forgetting and overfitting in fully few-shot incremental audio classification tasks.
TAR: Token-Aware Refinement for Fine-grained Generalized Category Discovery
Xingyu Yang (Southeast University), Xiu-Shen Wei (Southeast University)
ClassificationRecognitionTransformerContrastive LearningImage
🎯 What it does: Proposes the Token-Aware Refinement (TAR) framework, which suppresses attention artifacts in Vision Transformers (ViT) using three modules: TRM, CATS, and GRM. It enhances fine-grained features by leveraging entire token sequences rather than relying solely on [CLS] tokens, thereby improving performance in Generalized Category Discovery (GCD) tasks.
Target-Aware Invertible Encoder with Reconstruction Guidance for Infrared Small Target Detection
Shule Yan (Nanjing University of Science and Technology), Zexuan Ji (Nanjing University of Science and Technology)
Object DetectionFlow-based ModelAuto EncoderImageBenchmark
🎯 What it does: Propose an infrared small target detection framework named InvDet, which utilizes a reversible encoder to preserve information and further enhances detection performance through reconstruction-guided training.
TAS-LoRA: Transformer Architecture Search with Mixture-of-LoRA Experts
Jeimin Jeon (Yonsei University), Bumsub Ham (Yonsei University)
ClassificationNeural Architecture SearchTransformerMixture of ExpertsImage
🎯 What it does: Propose a Mixture-of-Experts (MoLE) framework that integrates low-rank adapters (LoRA) experts into Transformer architecture search (TAS), utilizing a lightweight router to enable subnet feature learning and addressing the feature collapse problem caused by shared weights.
Task-Aware Image Signal Processor for Advanced Visual Perception
Kai Chen (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
RestorationObject DetectionSegmentationConvolutional Neural NetworkImage
🎯 What it does: Designed a lightweight task-aware RAW-to-RGB processing framework called TA-ISP, which adaptively reconstructs RAW images at global, regional, and pixel levels using a multi-scale pixel modulation module to provide optimized representations for pre-trained visual models.
Task-Driven Implicit Representations for Automated Design of LiDAR Systems
Nikhil Behari (Massachusetts Institute of Technology), Ramesh Raskar (Massachusetts Institute of Technology)
Autonomous DrivingOptimizationRepresentation LearningFlow-based ModelPoint CloudMesh
🎯 What it does: Propose a task-driven automated design framework for LiDAR systems based on a continuous six-dimensional implicit representation
Task-Oriented Data Synthesis and Control-Rectify Sampling for Remote Sensing Semantic Segmentation
Yunkai Yang (Sun Yat-Sen University), Runmin Dong (Sun Yat-Sen University)
SegmentationGenerationData SynthesisTransformerPrompt EngineeringDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Proposes a task-oriented remote sensing semantic segmentation data synthesis framework called TODSynth, which achieves text-image-mask joint attention through a multi-modal diffusion transformer (MM-DiT) and introduces a control-correction flow matching (CRFM) sampling strategy based on task feedback to generate synthetic images that better meet the needs of semantic segmentation.
TaskForce: Cooperative Multi-agent Reinforcement Learning for Multi-task Optimization
Wonhyeok Choi (DGIST), Sunghoon Im (DGIST)
OptimizationReinforcement LearningImageBiomedical Data
🎯 What it does: Optimize gradient conflicts and scale imbalance in multi-task learning through the cooperative multi-agent reinforcement learning framework TaskForce, by learning weighted strategies for each task to update the shared network.
TaskIT: Memory-Efficient Fine-Tuning of Multi-LoRA LLMs via Cross-Task Importance Transfer
Cheng Fang (City University of Hong Kong), Bin Guo (Northwestern Polytechnical University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityAudio
🎯 What it does: This paper proposes the TaskIT framework, which addresses sparse fine-tuning of multiple LoRA LLMs on memory-constrained mobile/edge devices. It leverages cross-task importance transfer to predict the significance of uninserted modules and constructs an accurate memory prediction model through block-level activation dependency analysis. This enables optimal selection of LoRA module positions, quantities, and ranks under a given memory budget, achieving efficient model fine-tuning.
Tavatar: Topology-Aware Gaussian Attribute Derivation for Animatable Human Avatars
Hailin Luo (South China University of Technology), Mingkui Tan (South China University of Technology)
GenerationGaussian SplattingMesh
🎯 What it does: This paper proposes a high-quality animatable human avatar reconstruction framework called Tavatar based on topology, which uses 3D Gaussian points bound to a variable mesh and computes Gaussian geometric attributes analytically, preserving topological consistency while ensuring mesh quality through equilateral regularization.
Taxonomy-Aware Representation Alignment for Hierarchical Visual Recognition with Large Multimodal Models
Hulingxiao He (Peking University), Yuxin Peng (Peking University)
RecognitionLarge Language ModelReinforcement LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes a framework named TARA that aligns visual and label representations within large multimodal models (LMM) by leveraging biological foundation models (BFM) to achieve hierarchical visual recognition.
TC-Pade: Trajectory-Consistent Pade Approximation for Diffusion Acceleration
Shaoxuan He (Zhejiang University), Zhou Zhao (Zhejiang University)
GenerationComputational EfficiencyDiffusion modelImageVideoTextBenchmark
🎯 What it does: Proposed a trajectory-consistent feature prediction framework, TC-Pad'e, based on Pad'e approximation to accelerate diffusion model sampling at low step sizes (20–30 steps) while maintaining high-quality generation.
TDATR: Improving End-to-End Table Recognition via Table Detail-Aware Learning and Cell-Level Visual Alignment
Chunxia Qin (University of Science and Technology of China), Cong Liu (iFLYTEK Research)
RecognitionTransformerImageTextMultimodality
🎯 What it does: Propose the TDATR end-to-end table recognition framework, which first captures structure and content through table detail perception learning, and then achieves precise parsing via structure-guided cell localization.
Tea-Adapter: Teacher Adapter for Efficient Conditional Generation
Yinhan Zhang (Hong Kong University of Science and Technology), Zeyu Wang (Hong Kong University of Science and Technology)
GenerationKnowledge DistillationTransformerMixture of ExpertsDiffusion modelVideo
🎯 What it does: Proposes Tea-Adapter, a pluggable adapter that utilizes knowledge from a smaller teacher model to perform reverse distillation on large video diffusion models, achieving high-fidelity video generation under multi-conditions (single or combined conditions) with extremely low GPU memory usage.
Teacher-Guided Routing for Sparse Vision Mixture-of-Experts
Masahiro Kada (Institute of Science Tokyo), Ikuro Sato (Institute of Science Tokyo)
ClassificationKnowledge DistillationTransformerMixture of ExpertsImage
🎯 What it does: This paper proposes a Teacher-Guided Sparse Visual Expert Mixture-of-Experts Network (TGR-MoE), which trains a lightweight router on a pre-trained dense teacher model to provide pseudo-supervision for sparse MoE student routers, thereby stabilizing routing learning and reducing routing fluctuations during training.
Teaching DINOv3 About Partial 3D Geometry: A Self-Supervised Geometry-Aware Approach
Viktoria Ehm (Technical University of Munich), Daniel Cremers (Technical University of Munich)
Data SynthesisRepresentation LearningTransformerContrastive LearningPoint CloudMesh
🎯 What it does: This work injects 3D geometry awareness into the DINOv3 visual foundation model through a self-supervised LoRA method, achieving more robust feature extraction for partial 3D shapes.
TeamHOI: Learning a Unified Policy for Cooperative Human-Object Interactions with Any Team Size
Stefan Lionar, Gim Hee Lee
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: Propose TeamHOI, which utilizes a single decentralized strategy to achieve collaborative control for human-object interaction (HOI) of any scale.
TEAR: Temporal-aware Automated Red-teaming for Text-to-Video Models
Jiaming He (University of Electronic Science and Technology of China), Tianwei Zhang (Nanyang Technological University)
Safty and PrivacyAdversarial AttackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: Developed TEAR, an automated time-aware red teaming framework for text-to-video (T2V) models, designed to identify security vulnerabilities in time-series generation of models;
TeFlow: Enabling Multi-frame Supervision for Self-Supervised Feed-forward Scene Flow Estimation
Qingwen Zhang (Kth Royal Institute Of Technology), Patric Jensfelt (Kth Royal Institute Of Technology)
Autonomous DrivingOptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkOptical FlowImageVideo
🎯 What it does: Proposes TeFlow, a real-time scene flow estimation framework based on multi-frame self-supervised learning.
TeHOR: Text-Guided 3D Human and Object Reconstruction with Textures
Hyeongjin Nam, Kyoung Mu Lee (Seoul National University)
GenerationData SynthesisLarge Language ModelVision Language ModelDiffusion modelScore-based ModelGaussian SplattingImageTextPoint Cloud
🎯 What it does: This paper proposes TeHOR, a text-guided single-image 3D human and object simultaneous reconstruction framework.
Tell Model Where to Look: Mitigating Hallucinations in MLLMs by Vision-Guided Attention
Jianfei Zhao (Beijing Institute of Technology), Zhixing Tan (Zhongguancun Laboratory)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelMultimodality
🎯 What it does: Proposed a visual-guided attention (VGA) mechanism based on visual semantic confidence (VSC) to reduce hallucination generation in multimodal large language models.
Tell2Adapt: A Unified Framework for Source Free Unsupervised Domain Adaptation via Vision Foundation Model
Yulong Shi (Northeastern University), Lin Qi (Northeastern University)
SegmentationDomain AdaptationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringBiomedical DataMagnetic Resonance ImagingUltrasound
🎯 What it does: Proposes a unified source-free unsupervised domain adaptation framework called Tell2Adapt, which leverages visual foundation models (VFM) to generate high-quality pseudo labels and achieves cross-modal, multi-target medical image segmentation domain adaptation through knowledge distillation and visual plausibility refinement.
TempoControl: Temporal Attention Guidance for Text-to-Video Models
Shira Schiber (Bar-Ilan University), Idan Schwartz (Bar-Ilan University)
GenerationDiffusion modelVideoText
🎯 What it does: Propose a training-free, inference-time temporal control method called TEMPOCONTROL, enabling text-video diffusion models to make specific visual concepts appear or disappear at user-specified time points.
TempoMaster: Efficient Long Video Generation via Next-Frame-Rate Prediction
Yukuo Ma (Fudan University), Xuelong Li (Institute of Artificial Intelligence (TeleAI), China Telecom)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Propose the TempoMaster framework, reformulating long video generation as multi-rate frame prediction, first generating low-frame-rate global structures and then progressively refining to high-frame-rate details.
Temporal Equilibrium MeanFlow: Bridging the Scale Gap for One-Step Generation
Yuanpeng Tu (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
GenerationTransformerDiffusion modelFlow-based ModelImage
🎯 What it does: This paper proposes the Temporal Equilibrium MeanFlow (TEMF) framework, which addresses the time-scale imbalance issue in MeanFlow during first-order generation, achieving high-quality image generation.
Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning
Jinge Ma (Purdue University), Fengqing Zhu (Purdue University)
ClassificationImage
🎯 What it does: Propose Temporal-Adjusted Loss (TAL) to address prediction bias caused by temporal imbalance in class-incremental learning.
Temporal Interaction in Spiking Transformers with Multi-Delay Mixer
Kexin Shi (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
ClassificationRecognitionSpiking Neural NetworkTransformerImageVideoSequential
🎯 What it does: Introduce a Multi-Delay Mixer (MD-Mixer) in Spiking Transformer to explicitly model temporal information, and propose Temporal Interaction Coefficient (TIC) to evaluate the temporal interaction capability of attention mechanisms.
Temporal Inversion for Learning Interval Change in Chest X-Rays
Hanbin Ko (Seoul National University), Chang Min Park (Seoul National University)
ClassificationRetrievalSupervised Fine-TuningVision Language ModelContrastive LearningImageBiomedical Data
🎯 What it does: Proposed and implemented a complete framework that uniformly leverages time-reversed supervision for identifying and retrieving interval changes in chest X-ray images.
Temporal Representation Enhancement (TRE): Learning to Forget Dominant Patterns for Enhanced Temporal Spiking Features
Wei Liu (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
ClassificationRepresentation LearningConvolutional Neural NetworkSpiking Neural NetworkImageTime Series
🎯 What it does: Propose Temporal Representation Enhancement (TRE), which dynamically suppresses redundant dominant features in Spiking Neural Networks (SNNs) through a learning-based forgetting mechanism, enhancing feature diversity and discriminability across multiple time steps.
TempR1: Improving Temporal Understanding of MLLMs via Temporal-Aware Multi-Task Reinforcement Learning
Tao Wu (Nanjing University), Limin Wang (Nanjing University)
TransformerReinforcement LearningVision Language ModelVideoTextMultimodality
🎯 What it does: Propose TempR1, a multi-task reinforcement learning framework based on GRPO, to enhance the reasoning and localization capabilities of multimodal large models in video temporal understanding tasks.
TerraScope: Pixel-Grounded Visual Reasoning for Earth Observation
Yan Shu (University of Trento), Paolo Rota (University of Trento)
SegmentationTransformerVision Language ModelImageMultimodalityTime SeriesBenchmarkChain-of-Thought
🎯 What it does: Introduce a unified visual language model called TerraScope capable of performing multimodal and multitemporal reasoning at the pixel level for Earth observation.
TerraSeg: Self-Supervised Ground Segmentation for Any LiDAR
Ted Lentsch (Delft University of Technology), Dariu M. Gavrila (Delft University of Technology)
SegmentationAutonomous DrivingTransformerPoint Cloud
🎯 What it does: Propose a self-supervised, domain-agnostic LiDAR ground segmentation model called TerraSeg, and construct the OmniLiDAR unified dataset and PseudoLabeler pseudo-label generator.
TESO: Online Tracking of Essential Matrix by Stochastic Optimization
Jaroslav Moravec (Czech Technical University in Prague), Akihiro Sugimoto (National Institute of Informatics)
Pose EstimationDepth EstimationAutonomous DrivingOptimizationImageVideo
🎯 What it does: This paper proposes an algorithm called TESO for online tracking of the extrinsic and intrinsic parameters of stereo cameras, which uses random optimization on the essential matrix manifold and a robust essential matrix error based on kernel correlation to real-time correct camera rotation drift.
TESSERA: Temporal Embeddings of Surface Spectra for Earth Representation and Analysis
Zhengpeng Feng (University of Cambridge), Srinivasan Keshav (University of Cambridge)
ClassificationSegmentationRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningMultimodalityTime SeriesAgriculture Related
🎯 What it does: Proposes TESSERA—a pixel-level multimodal (Sentinel-1/2) time series foundation model that generates efficient embeddings for sparse satellite observations using self-supervised learning;
Test-Time 3D Occupancy Prediction
Fengyi Zhang (University of Queensland), Yadan Luo (University of Queensland)
Autonomous DrivingGaussian SplattingSimultaneous Localization and MappingOptical FlowImagePoint Cloud
🎯 What it does: Proposes a 3D occupancy prediction framework TT-Occ that generates time-aware 3D Gaussians using visual foundation models during testing and supports voxelization at arbitrary resolutions.
Test-Time Alignment of Text-to-Image Diffusion Models via Null-Text Embedding Optimisation
Taehoon Kim (University of Edinburgh), Timothy Hospedales (University of Edinburgh)
GenerationDiffusion modelImageText
🎯 What it does: Propose a framework called Null-TTA that achieves alignment in text-to-image diffusion models during inference by optimizing unconditional text embeddings (null-text).
Test-Time Attention Purification for Backdoored Large Vision Language Models
Zhifang Zhang (University of Queensland), Miao Xu (University of Queensland)
Adversarial AttackTransformerVision Language ModelMultimodality
🎯 What it does: Proposes a training-free test-time attention purification method called CleanSight to defend against backdoor attacks on vision-language models.
Test-time Ego-Exo-centric Adaptation for Action Anticipation via Multi-Label Prototype Growing and Dual-Clue Consistency
Zhaofeng Shi (University of Electronic Science and Technology of China), Hongliang Li (University of Electronic Science and Technology of China)
ClassificationDomain AdaptationRecurrent Neural NetworkContrastive LearningVideoTextMultimodality
🎯 What it does: Proposes the task TE2A, which performs instant adaptation for action prediction between egocentric and exocentric views during testing, and introduces the Dual-Clue enhanced Prototype Growing Network (DCPGN) scheme
Test-Time Instance-Specific Parameter Composition: A New Paradigm for Adaptive Generative Modeling
Minh-Tuan Tran (Monash University), Trung Le (Monash University)
GenerationTransformerImage
🎯 What it does: Dynamically generate low-rank parameter updates for pre-trained generative models during the inference phase, and combine them with the original weights to achieve input-specific adaptive parameter combinations.
Test-Time Multi-Prompt Adaptation for Open-Vocabulary Remote Sensing Image Segmentation
Ting Yang (Tianjin University), Qinghua Hu (Tianjin University)
SegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
🎯 What it does: Propose a test-time multi-prompt adaptation method (TMPA), which generates diverse context-aware text descriptions and adaptively calibrates prompt embeddings during inference to alleviate text ambiguity in open-vocabulary remote sensing image semantic segmentation, thereby improving segmentation accuracy.
Test-Time Perturbation Tuning with Delayed Feedback for Vision-Language-Action Models
Zehua Zang (Institute of Software, Chinese Academy of Sciences), Jiangmeng Li (Institute of Software, Chinese Academy of Sciences)
Computational EfficiencyRobotic IntelligenceReinforcement LearningVision-Language-Action ModelVideoMultimodality
🎯 What it does: Propose a validator-free test-time adaptation framework PDF, which leverages uncertainty-driven data augmentation voting and delay feedback-guided lightweight perturbation learning to enhance the robustness of vision-language-action models under minor environmental changes.
Test-time Sparsity for Extreme Fast Action Diffusion
Kangye Ji (Tsinghua University), Zhi Wang (Tsinghua University)
Computational EfficiencyRobotic IntelligenceTransformerDiffusion modelSequentialBenchmark
🎯 What it does: Proposes a method for accelerating action diffusion models during testing by predicting which residual calculations can be skipped at each step and reusing historical features.
Test-Time Training for LiDAR Semantic Segmentation under Corruption via Geometric Inlier Discrimination
Hyeonseong Kim (KAIST), Kuk-Jin Yoon (KAIST)
SegmentationDomain AdaptationConvolutional Neural NetworkPoint Cloud
🎯 What it does: Propose a self-supervised test-time training framework based on geometry-inlier discrimination (GeoID) to adapt to distribution drift caused by sensor and environmental damage in LiDAR semantic segmentation tasks.
Text-Driven 3D Hand Motion Generation from Sign Language Data
Léore Bensabath (École des Ponts ParisTech), Gül Varol (École des Ponts ParisTech)
GenerationData SynthesisPose EstimationRetrievalLarge Language ModelVision-Language-Action ModelDiffusion modelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Constructed a dataset named BOBSL3DT, which aligns 3D hand actions with text descriptions based on British Sign Language (BSL) videos, and trained a text-driven hand action diffusion model called HandMDM to generate 3D hand actions that conform to natural language descriptions.
Text-guided Feature Disentanglement for Cross-modal Gait Recognition
Zhiyang Lu (Xiamen University), Ming Cheng (Xiamen University)
RecognitionConvolutional Neural NetworkLarge Language ModelSupervised Fine-TuningVision Language ModelImagePoint CloudRetrieval-Augmented Generation
🎯 What it does: Designed a text-guided cross-modal gait recognition framework called TCFDNet, which can decouple modal features from camera videos and LiDAR point clouds, and learn modality-shared and modality-specific representations.
Text-Image Conditioned 3D Generation
Jiazhong Cen (Shanghai Jiao Tong University), Qi Tian (Huawei Inc)
GenerationTransformerVision Language ModelDiffusion modelRectified FlowImageTextMultimodality
🎯 What it does: Propose a text-image joint conditional 3D generation task and provide a diagnostic study;
Text-Phase Synergy Network with Dual Priors for Unsupervised Cross-Domain Image Retrieval
Jing Yang (Southeast University), Pengfei Fang (Southeast University)
RetrievalDomain AdaptationVision Language ModelContrastive LearningImageText
🎯 What it does: Proposed TPSNet, which leverages text and phase dual priors to achieve unsupervised cross-domain image retrieval
Text-Printed Image: Bridging the Image-Text Modality Gap for Text-centric Training of Large Vision-Language Models
Shojiro Yamabe, Tsubasa Takahashi
Data SynthesisLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes using Text Printed Images (TPI) to generate visual inputs in text-centric training to address the image-text modality gap;
TextFM: Robust Semi-dense Feature Matching with Language Guidance
Zhihao Zheng (Lehigh University), Mooi Choo Chuah (Lehigh University)
Pose EstimationRetrievalTransformerPrompt EngineeringVision Language ModelImagePoint CloudBenchmark
🎯 What it does: Propose TextFM, a semi-dense feature matching framework that leverages vision-language model (VLM) text embeddings to guide coarse matching, fine-tunes visual foundation models with LoRA, and incorporates illumination-invariant physical priors;
TextOVSR: Text-Guided Real-World Opera Video Super-Resolution
Hua Chang (Wuhan University of Science and Technology), Qi Tian (Huawei Technologies Ltd)
Super ResolutionConvolutional Neural NetworkLarge Language ModelVision Language ModelGenerative Adversarial NetworkVideoTextMultimodality
🎯 What it does: Proposed a text-guided dual-branch network called TextOVSR for super-resolution reconstruction of real-world Peking opera videos.