arXivSub Start free trial

CVPR 2026 Papers — Page 37

IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers

Towards Fine-Grained Attribution: Instance-Aware Preference Optimization for Aligning Diffusion Models

Jiayang Sun (University of Chinese Academy of Sciences), Ran He (University of Chinese Academy of Sciences)

GenerationOptimizationReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality

🎯 What it does: In the alignment task of text-to-image diffusion models, the Instance-Aware Preference Optimization (IAPO) method is proposed, achieving instance-level credit assignment and fine-grained preference optimization.

Towards Foundation Models for 3D Scene Understanding: Instance-Aware Self-Supervised Learning for Point Clouds

Bin Yang (Robert Bosch GmbH), Alexandru Paul Condurache (Robert Bosch GmbH)

SegmentationAutonomous DrivingKnowledge DistillationPoint Cloud

🎯 What it does: Proposes a self-supervised learning framework called PointINS that can simultaneously learn semantic consistency and instance-aware geometric reasoning for point clouds.

Towards Generalizable AI-Generated Image Detection via Image-Adaptive Prompt Learning

Yiheng Li (University of Chinese Academy of Sciences), Yang Yang (Sangfor Technologies Inc)

Anomaly DetectionTransformerPrompt EngineeringVision Language ModelImage

🎯 What it does: Proposes an AI-generated image detection framework based on Image Adaptive Prompt Learning (IAPL), which dynamically adjusts input prompts for each test image during the inference stage to enhance generalization capabilities for unknown generators.

Towards Generalized Multimodal Homography Estimation

Jinkun You (University of Macau), Yicong Zhou (University of Macau)

Data SynthesisDomain AdaptationConvolutional Neural NetworkMultimodality

🎯 What it does: Propose a training data synthesis method and a cross-scale color-invariant network to enhance the generalization of multi-modal homography estimation

Towards Generalized Representations for Low-Light Understanding: When Signal Constancy Meets Semantic Enrichment

Yifan Li (Wangxuan Institute of Computer Technology, Peking University), Jiaying Liu (Wangxuan Institute of Computer Technology, Peking University)

Image TranslationRestorationObject DetectionSegmentationTransformerVision Language ModelContrastive LearningImage

🎯 What it does: Proposes a unified prior (UniPrior) framework that simultaneously leverages illumination-invariant prior to ensure signal constancy in low-light image understanding, and enriches features with Vision Foundation Model (VFM) semantic prior without requiring any real low-light training data.

Towards GUI Agents: Vision-Language Diffusion Models for GUI Grounding

Shrinidhi Kumbhar (Arizona State University), Kunwar Yashraj Singh (AWS Agentic AI)

Object DetectionGenerationAgentic AIVision-Language-Action ModelDiffusion modelMultimodality

🎯 What it does: This paper migrates the discrete diffusion vision-language model LLaDA-V to the single-step GUI localization and action generation task.

Towards High-Quality Image Segmentation: Improving Topology Accuracy by Penalizing Neighbor Pixels

Juan Miguel Valverde (Technical University of Denmark), Anders Bjorholm Dahl (Technical University of Denmark)

SegmentationImageBiomedical Data

🎯 What it does: Propose a Same-Class Pixel Neighborhood Penalty (SCNP) method, which adjusts pixel logits during training by incorporating the worst prediction within their neighborhood, thereby enhancing topological consistency in segmentation results.

Towards High-resolution and Disentangled Reference-based Sketch Colorization

Dingkun Yan (University of Tokyo), Jiaxian Guo (University of Tokyo)

Image TranslationGenerationDiffusion modelImage

🎯 What it does: This work proposes a dual-branch framework and suppresses distribution shift between training and inference through Gram regularization loss, successfully achieving high-resolution, non-entangled reference-based sketch colorization;

Towards Highly Transferable Vision-Language Attack via Semantic-Augmented Dynamic Contrastive Interaction

Yuanbo Li (Jiangnan University), Josef Kittler (University of Surrey)

Adversarial AttackVision Language ModelContrastive LearningImageTextMultimodality

🎯 What it does: Proposes a method called SADCA for generating adversarial examples that can be highly transferable across various visual-language models and tasks.

Towards Highly-Constrained Human Motion Generation with Retrieval-Guided Diffusion Noise Optimization

Hanchao Liu (Tsinghua University), Shi-Min Hu (Tsinghua University)

GenerationDiffusion modelTime SeriesRetrieval-Augmented Generation

🎯 What it does: Propose a retrieval-guided diffusion noise optimization framework to address high-constraint human motion generation tasks.

Towards Holistic Modeling for Video Frame Interpolation with Auto-regressive Diffusion Transformers

Xinyu Peng (Shanghai Jiao Tong University), Hongkai Xiong (Fudan University)

GenerationTransformerDiffusion modelAuto EncoderVideo

🎯 What it does: Propose a video-centric autoregressive diffusion transformer framework called LDF-VFI, which can synthesize entire video segments at once while maintaining long-term temporal consistency, thus solving the issues of temporal inconsistency and motion artifacts caused by traditional frame-centric methods.

Towards Human-Imperceptible Backdoor Attacks on Text-to-Image Diffusion Models

Yiming Wu (Zhejiang University of Technology), Zhen Hong (Zhejiang University of Technology)

GenerationAdversarial AttackLarge Language ModelVision Language ModelDiffusion modelAuto EncoderImageTextMultimodality

🎯 What it does: This study proposes a label-free backdoor attack scheme targeting text-to-image diffusion models, achieving stealthy model control through bimodal subtle perturbations and semantics-preserving triggers.

Towards Human-Like Robot Handwriting via Contour-Aware Generation

Yutao Qin (South China University of Technology), Shuangping Huang (South China University of Technology)

GenerationRobotic IntelligenceConvolutional Neural NetworkGraph Neural NetworkImageGraph

🎯 What it does: Proposed the Contour-aware Handwriting Trajectory Reconstruction (CHTR) task and constructed the first large-scale handwriting trajectory dataset CHTR-110K containing contour information; simultaneously designed the G-HTR method based on multi-scale graph neural networks, achieving complete trajectory generation from character images to bandwidth, stroke order, and stroke details;

Towards Intrinsic-Aware Monocular 3D Object Detection

Zhihao Zhang (Michigan State University), Xiaoming Liu (Michigan State University)

Object DetectionAutonomous DrivingTransformerLarge Language ModelVision Language ModelImageText

🎯 What it does: Propose a unified MonoIA framework that converts camera intrinsic parameters into language-oriented semantic embeddings and layer-by-layer fuses them in the detection network, achieving adaptability and robustness of monocular 3D object detection to camera intrinsic parameters.

Towards Knowledge-augmented Bayesian Deep Learning For Computer Vision

Wang Ma (Rensselaer Polytechnic Institute), Qiang Ji (Rensselaer Polytechnic Institute)

Pose EstimationImage

🎯 What it does: This paper proposes a two-stage knowledge-enhanced Bayesian deep learning framework, where domain knowledge is used in the prior learning stage to generate information-rich priors, and in the posterior inference stage, the model is continuously constrained through adaptive knowledge likelihood.

Towards Motion Turing Test: Evaluating Human-Likeness in Humanoid Robots

Mingzhe Li (Xiamen University), Cheng Wang (Xiamen University)

Robotic IntelligenceRecurrent Neural NetworkGraph Neural NetworkLarge Language ModelVideoTextBenchmark

🎯 What it does: Propose the Motion Turing Test framework, construct the HHMotion dataset, annotate the 'human likeness' of human and humanoid robot motions, and introduce PTR-Net as a baseline model for automatically evaluating motion human likeness.

Towards Multimodal Domain Generalization with Few Labels

Hongzhao Li (Zhengzhou University), Muhammad Haris Khan (MBZUAI)

Domain AdaptationMultimodalityBenchmark

🎯 What it does: This paper proposes the semi-supervised multi-modal domain generalization (SSMDG) problem, constructing a unified framework that jointly utilizes limited labeled and abundant unlabeled multi-modal data from multiple source domains to achieve generalization to unseen domains;

Towards Open Environments and Instructions: General Vision-Language Navigation via Fast-Slow Interactive Reasoning

Yang Li (Tianjin University), Yahong Han (Tianjin University)

Autonomous DrivingTransformerLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelImageTextMultimodalityChain-of-Thought

🎯 What it does: Propose a dual-speed interactive reasoning framework (slow4fast-VLN), which generates actions in real-time through a fast reasoning module while recording memories, and then enhances rapid decision-making dynamically by reflecting and extracting experiences from a slow reasoning module, addressing adaptability issues in unknown scenarios and diverse instructions within the GSA-VLN task.

Towards Open-Vocabulary Industrial Defect Understanding with a Large-Scale Multimodal Dataset

Tsai-Ching Ni (National Yang Ming Chiao Tung University), Yuan-Fu Yang (National Yang Ming Chiao Tung University)

Object DetectionSegmentationGenerationAnomaly DetectionConvolutional Neural NetworkTransformerLarge Language ModelDiffusion modelContrastive LearningImageMultimodality

🎯 What it does: Proposed the IMDD-1M 1 million aligned industrial defect image-text dataset, and trained a diffusion-based vision-language foundation model on it, achieving unified inference for multiple tasks such as defect generation, segmentation, and detection.

Towards Persistence: Learning Topological Constraints for Event-based Small Object Detection

Shiman He (National University of Defense Technology), Miao Li (National University of Defense Technology)

Object DetectionConvolutional Neural NetworkGraph Neural NetworkPoint CloudBenchmark

🎯 What it does: This paper proposes a topology-constrained sparse convolutional network called SpTopoNet for detecting extremely small targets under event cameras.

Towards Photorealistic and Efficient Bokeh Rendering via Diffusion Framework

Linxiao Shi (Shenzhen Institutes of Advanced Technology), Peng-Tao Jiang (vivo Mobile Communication Co., Ltd.)

RestorationGenerationDepth EstimationSuper ResolutionTransformerSupervised Fine-TuningDiffusion modelImage

🎯 What it does: Integrates super-resolution and bokeh rendering in a single step, providing a unified diffusion framework called MagicBokeh for high-zoom mobile image processing.

Towards Policy-Adaptive Image Guardrail: Benchmark and Method

Caiyong Piao (Fudan University), Shuigeng Zhou (Fudan University)

Reinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark

🎯 What it does: Proposes a cross-strategy visual safety gatekeeping framework, SafeGuard-VL, and constructs a benchmark of safe/unsafe image pairs based on image editing, SafeEditBench, to evaluate the adaptability of Vision-Language Models (VLMs) under different safety strategies.

Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training

Gengluo Li (Institute of Information Engineering, Chinese Academy of Sciences), Yu Zhou (Nankai University)

RecognitionData SynthesisTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark

🎯 What it does: This paper proposes a data-training collaborative framework that generates large-scale full-page end-to-end document parsing data using realistic scene synthesis, and improves the model's parsing quality and robustness through structure-aware training.

Towards Realistic and Consistent Orbital Video Generation via 3D Foundation Priors

Rong Wang (Australian National University), Hongdong Li (Australian National University)

GenerationDiffusion modelVideoMesh

🎯 What it does: Propose a single-image orbit video generation method based on the prior of a 3D foundation model, capable of generating geometrically consistent and shape-realistic orbit videos.

Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models

Hongji Li (MBZUAI), Lijie Hu (MBZUAI)

Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought

🎯 What it does: Proposes a training-agnostic, inference-time intervention framework called R-MUSE, which enables machine forgetting at the reasoning level in multi-modal large language models while preserving the model's overall reasoning capabilities.

Towards Reliable Evaluation of Adversarial Robustness for Spiking Neural Networks

Jihang Wang (Chinese Academy of Sciences), Yi Zeng (Chinese Academy of Sciences)

ClassificationAdversarial AttackSpiking Neural NetworkImage

🎯 What it does: This paper proposes a reliable framework for evaluating the adversarial robustness of Spiking Neural Networks (SNNs), which includes Adaptive Sharpness Substitute Gradient (ASSG) and Stable Adaptive PGD (SA-PGD) attacks.

Towards Robust Multi-Modal Semantic Segmentation with Teacher-Student Framework and Hybrid Prototype Distillation

Jiaqi Tan (Beijing University of Posts and Telecommunications), Yang Liu (Beijing University of Posts and Telecommunications)

SegmentationKnowledge DistillationTransformerMultimodality

🎯 What it does: Proposes RobustSeg, a teacher-student self-distillation framework, leveraging cross-modal prototype distillation (Hybrid Prototype Distillation, HPD) and feedback mechanisms to enhance robustness in multi-modal semantic segmentation under missing modality scenarios, while maintaining high accuracy when all modalities are present.

Towards Robust Multimodal Large Language Models Against Jailbreak Attacks

Ziyi Yin (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)

Safty and PrivacyAdversarial AttackLarge Language ModelVision Language ModelContrastive LearningMultimodality

🎯 What it does: Propose SAFEMLLM, an adversarial training framework for multi-modal large language models (MLLMs), which alternately iterates adversarial attacks and defenses by injecting adversarial noise into the embedding layer and combining it with contrastive loss;

Towards Robust Sequential Decomposition for Complex Image Editing

Zilai Zeng (Brown University), Peng Wang (University Of Tokyo)

GenerationData SynthesisVision Language ModelFlow-based ModelRectified FlowImageTextMultimodalitySequential

🎯 What it does: Studies how to achieve complex image editing through sequential decomposition (step-by-step editing), proposes a unified in-context editing framework, and builds a synthetic data pipeline on Blender to generate multi-step editing tasks.

Towards Robust Vision Transformers: Path Dependency Analysis and a Simple Two-Stage Adversarial Training

Seongmin Kim (Inha University), Byung Cheol Song (Inha University)

ClassificationKnowledge DistillationAdversarial AttackTransformerImage

🎯 What it does: This paper reveals the path dependency, semantic priors, and global-local relationships of ViT during adversarial training through Gradient Path Masking (GPM) analysis and multi-dimensional observations such as attention maps and patch relationships. Based on these findings, a two-phase adversarial training framework is proposed, significantly enhancing the robustness and generalization ability of ViT.

Towards Sparse Video Understanding and Reasoning

Chenwei Xu (Northwestern University), Han Liu (Northwestern University)

Computational EfficiencyLarge Language ModelReinforcement LearningVision Language ModelVideo

🎯 What it does: Propose a multi-round, frame-sparse question-answering framework called REVISE, which can incrementally select key frames, maintain summary states, and implement early stopping;

Towards Stable Federated Continual Test-Time Adaptation in Wild World

Liwen Wang (Anhui University), Zhe Jin (University of Nottingham)

ClassificationSegmentationDomain AdaptationFederated LearningImageBiomedical Data

🎯 What it does: Under the federated learning framework, for scenarios with continuous, unlabeled client arrivals, BPFedCTTA is proposed, which employs Bayesian Prior-guided Adaptation (BPA) for locally stable test-time adaptation and Uncertainty-Gated Single-client Aggregation (UGSA) to achieve secure global model updates.

Towards Stable Self-Supervised Object Representations in Unconstrained Egocentric Video

Yuting Tan (Communication University of China), Jingjing Zhang (Chinese University of Hong Kong)

RecognitionObject DetectionSegmentationKnowledge DistillationRepresentation LearningTransformerContrastive LearningVideo

🎯 What it does: This paper proposes EgoViT, which learns stable self-supervised object representations using unlabeled first-person videos, enabling continuous object recognition and segmentation under extreme conditions such as occlusion, motion, and cluttered backgrounds.

Towards Stealthy and Effective Backdoor Attacks on Lane Detection: A Naturalistic Data Poisoning Approach

Yifan Liao (Changan Automobile), Jin Song Dong (National University of Singapore)

Autonomous DrivingAdversarial AttackDiffusion modelImage

🎯 What it does: Propose a naturalized data poisoning method DBALD based on diffusion models, designed to implant stealthy and efficient backdoor triggers into lane detection models.

Towards Storytelling Animations: Joint Synthesis of Human and Camera Motions

Boyuan Cheng (Bournemouth University), Xiaosong Yang (Bournemouth University)

GenerationData SynthesisTransformerDiffusion modelSequential

🎯 What it does: This paper proposes a joint diffusion-based generative model that can simultaneously generate 3D motion trajectories of two characters and the camera's motion trajectory, enabling story-driven animation creation.

Towards Streaming Referring Video Segmentation via Large Language Model

Wenkang Zhang (Imperial College London), Jiankang Deng (Alibaba Group)

SegmentationLarge Language ModelVision Language ModelVideoMultimodality

🎯 What it does: Propose a StreamingRVOS framework that utilizes large language models to achieve online reference video segmentation, avoiding the traditional multi-step process of offline sampling, image segmentation, and subsequent mask propagation.

Towards Training-free Scene Text Editing

Yubo Li (Chinese Academy of Sciences), Kexin Zhang (Nanjing University of Science and Technology)

Image TranslationImage HarmonizationGenerationTransformerDiffusion modelFlow-based ModelImageText

🎯 What it does: Proposed a training-free scene text editing framework called TextFlow, which can achieve high-quality modification of text content while preserving the original image structure and font style.

Towards Uncertainty-aware Unsupervised Domain Adaptation for Videos and Time-Series with Causal Optimal Transport

Khushboo Mishra (Indian Institute of Technology Varanasi), Tanima Dutta (Indian Institute of Technology Varanasi)

Domain AdaptationConvolutional Neural NetworkRecurrent Neural NetworkVideoTime Series

🎯 What it does: Propose a unified unsupervised domain adaptation framework called Causal-OT, combining Granger causal graphs with optimal transport, addressing temporal feature alignment, causal structure preservation, and prediction uncertainty handling.

Towards Unified Human Perception and Machine Understanding: Token Flow Guided Compression Framework

Li Xu (Xidian University), Yunsong Li (Xidian University)

CompressionTransformerVision Language ModelFlow-based ModelMultimodality

🎯 What it does: Propose a variable bitrate compression framework called TFGC based on 1D tokens, which supports the unified utilization of human perception and machine reasoning, and the compressed tokens can be directly used by large vision-language models.

Towards Universal Computational Aberration Correction in Photographic Cameras: A Comprehensive Benchmark Analysis

Xiaolong Qian (Zhejiang University), Kaiwei Wang (Zhejiang University)

RestorationConvolutional Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkImageBenchmarkPhysics RelatedOrdinary Differential Equation

🎯 What it does: Investigated the benchmark for general computational aberration correction (CAC), constructed a large-scale lens library and unified evaluation metrics, and evaluated 24 image restoration and CAC algorithms.

Towards Visual Query Localization in the 3D World

Liang Peng, Libo Zhang (Wuhan University)

Object TrackingPose EstimationTransformerImageMultimodalityPoint CloudBenchmark

🎯 What it does: Proposed the 3D Visual Query Localization (3DVQL) task and benchmark dataset

TR2M: Transferring Monocular Relative Depth to Metric Depth with Language Descriptions and Dual-Level Scale-Oriented Contrast

Beilei Cui (Chinese University of Hong Kong), Hongliang Ren (Chinese University of Hong Kong)

Depth EstimationTransformerVision Language ModelContrastive LearningMultimodality

🎯 What it does: This paper proposes a unified framework TR2M, which predicts pixel-level scaling (scale) and offset maps using images and corresponding textual descriptions, subsequently converting pre-trained relative depth images into metric depth via pixel-level linear transformation.

TraceGen: World Modeling in 3D Trace Space Enables Learning from Cross-Embodiment Videos

Seungjae Lee (University of Maryland College Park), Furong Huang (University of Maryland College Park)

Data SynthesisRobotic IntelligenceTransformerVision Language ModelVision-Language-Action ModelDiffusion modelWorld ModelOptical FlowVideoMultimodality

🎯 What it does: Propose TraceGen, a framework that learns world models in 3D trajectory space, along with the TraceForge data generation pipeline, supporting cross-body and cross-environment video learning.

Tracking by Predicting 3-D Gaussians Over Time

Tanish Baranwal (University of California Berkeley), Jitendra Malik (University of California Berkeley)

Object TrackingTransformerAuto EncoderGaussian SplattingVideo

🎯 What it does: Propose a self-supervised video pre-training framework called Video-GMAE, which models videos by predicting the evolution of 3D Gaussian primitives over time;

Tracking through Severe Occlusion via Event-Derived Transient Cues

Hao Dong (Huazhong University of Science and Technology), Luxin Yan (Huazhong University of Science and Technology)

Object TrackingTransformerImageVideo

🎯 What it does: Propose the EvoTrack framework, combining microsecond-level temporal information from event cameras to achieve event-driven motion autoregression and target-aware appearance matching, specifically designed for nonlinear high-speed object tracking under severe occlusion.

Tracking-Guided 4D Generation: Foundation-Tracker Motion Priors for 3D Model Animation

Su Sun (Purdue University), Mei Chen (Purdue University)

GenerationData SynthesisDiffusion modelScore-based ModelGaussian SplattingVideoPoint Cloud

🎯 What it does: Propose Track4DGen, a two-stage framework that animates arbitrary static 3D models by leveraging multi-view video diffusion models, a base point tracker, and 4D Gaussian Splatting (4D-GS).

TrackMAE: Video Representation Learning via Track Mask and Predict

Renaud Vandeghen (University of Li' ege), Bernard Ghanem (KAUST)

Object TrackingRepresentation LearningTransformerAuto EncoderVideo

🎯 What it does: This paper proposes TrackMAE, a self-supervised video pre-training framework based on trajectory masking and prediction, which uses motion trajectories obtained from sparse point tracking as additional supervisory signals and further enhances temporal dynamic modeling through motion-aware masking;

TrafficAlign: Aligning Large Language Models for Traffic Scenario Generation

Zhi Tu (Purdue University), Tianyi Zhang (Purdue University)

Autonomous DrivingLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodality

🎯 What it does: Proposed and implemented the TRAFFICALIGN framework, which automatically generates, validates, and aligns LLMs using real driving videos to create safe testing scenarios that conform to actual traffic distributions.

Trainable Log-linear Sparse Attention for Efficient Diffusion Transformers

Yifan Zhou, Xingang Pan

GenerationComputational EfficiencyTransformerDiffusion modelImage

🎯 What it does: Propose Log-linear Sparse Attention (LLSA), a sparse attention mechanism for diffusion Transformers, which achieves O(N log N) computational complexity on long sequences.

Training High-Level Schedulers with Execution-Feedback Reinforcement Learning for Long-Horizon GUI Automation

Zehao Deng (Soochow University), Gongshen Liu (Shanghai Jiao Tong University)

TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningBenchmark

🎯 What it does: Propose a staged execution feedback reinforcement learning framework (CES), which improves long-sequence GUI automation performance by decomposing task scheduling and state tracking into two high-level modules (Coordinator and State Tracker) and decoupling them from the low-level Executor.

Training One Model to Master Cross-Level Agentic Actions via Reinforcement Learning

Kaichen He (Peking University), Yitao Liang (Peking University)

TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision-Language-Action ModelTextMultimodality

🎯 What it does: This paper proposes CrossHA, a native agent model capable of dynamically switching among multiple heterogeneous action spaces (such as high-level APIs, low-level atomic commands, GUI events, etc.) within the same policy.

Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods

Omer Ben Hayun (Technion Israel Institute of Technology), Guy Gilboa (Technion Israel Institute of Technology)

Anomaly DetectionTransformerVision Language ModelContrastive LearningVideo

🎯 What it does: Propose a zero-shot, no-training generative video detection framework named STALL, which distinguishes real and synthetic videos based on Gaussian likelihood estimation in spatial and temporal dimensions.

Training-free Mixed-Resolution Latent Upsampling for Spatially Accelerated Diffusion Transformers

Wongi Jeong (Seoul National University), Se Young Chun (Seoul National University)

GenerationComputational EfficiencyTransformerDiffusion modelAuto EncoderImage

🎯 What it does: Proposes a training-free spatial acceleration framework RALU, achieving efficient inference for Diffusion Transformers via hybrid-resolution latent upscaling.

Training-free Motion Factorization for Compositional Video Generation

Zixuan Wang (Sichuan University), Yinjie Lei (Sichuan University)

GenerationLarge Language ModelVision Language ModelDiffusion modelOptical FlowVideoBenchmark

🎯 What it does: Proposes a training-agnostic motion factorization framework that decomposes motion into three categories—static, rigid, and non-rigid—in multi-instance video generation, utilizing structured motion graphs for planning and disentangled motion guidance.

Training-Free Open-Vocabulary Camouflaged Object Segmentation via Fine-Grained Object Binding and Adaptive Hybrid Prompt

Peng Ren (Jilin University), Tian Bai (Jilin University)

SegmentationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark

🎯 What it does: Propose a training-free open-source vocabulary hidden object segmentation framework, utilizing fine-grained object binding and adaptive hybrid prompting to achieve precise text-visual binding and segmentation;

Training-free, Perceptually Consistent Low-Resolution Previews with High-Resolution Image for Efficient Workflows of Diffusion Models

Wongi Jeong (Seoul National University), Se Young Chun (Seoul National University)

GenerationComputational EfficiencyDiffusion modelFlow-based ModelRectified FlowImage

🎯 What it does: Proposed a training-agnostic low-resolution preview generation method that can quickly screen candidate images while maintaining perceptual consistency of high-resolution images.

Training-Only Heterogeneous Image-Patch-Text Graph Supervision for Advancing Few-Shot Learning Adapters

Mohammed Rahman Sherif Khan Mohammad (Edge Hill University), Amr Ahmed (Edge Hill University)

ClassificationKnowledge DistillationMeta LearningGraph Neural NetworkTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark

🎯 What it does: Propose a heterogeneous graph teacher model used during training, which constructs a graph by leveraging multi-scale image patches and text labels, and enhances the representational power of Tip-Adapter's cache keys through deep cross-modal reasoning with Modality-aware Graph Transformer, without increasing inference overhead.

TrajRAG: Retrieving Geometric-Semantic Experience for Zero-Shot Object Navigation

Yiyao Wang (Institute of Computing Technology, Chinese Academy of Sciences), Shuqiang Jiang (Institute of Computing Technology, Chinese Academy of Sciences)

RetrievalAutonomous DrivingLarge Language ModelContrastive LearningMultimodalityRetrieval-Augmented Generation

🎯 What it does: This paper proposes a framework called TrajRAG based on retrieval-augmented generation, which retrieves geometric semantic experiences to assist large models in decision-making for zero-shot object navigation.

TrajTok: Learning Trajectory Tokens Enhances Video Understanding

Chenhao Zheng (University of Washington), Ranjay Krishna (University of Washington)

RetrievalRepresentation LearningConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningVideo

🎯 What it does: Propose an end-to-end differentiable trajectory tokenizer, TrajTok, to convert videos into trajectory-level tokens and apply it to pretraining video encoders, pretraining feature adaptation, and vision-language models.

Transform to Transfer: Boosting Adversarial Attack Transferability on Vision-Language Pre-training Models

Yang Li (Fujian Province Key Laboratory of Information Security and Network Systems), Wei Lin (Fujian University of Technology)

Adversarial AttackTransformerVision Language ModelMultimodality

🎯 What it does: Propose and implement a new black-box attack method called Transform to Transfer Attack (TTA), which enhances the transferability of vision-language pretraining models (VLP) across different downstream tasks through a learnable block-level input transformation and boosted integral gradients (boosted IG).

Transition Matching Distillation for Fast Video Generation

Weili Nie (Nvidia), Arash Vahdat (Nvidia)

GenerationKnowledge DistillationDiffusion modelGenerative Adversarial NetworkVideoText

🎯 What it does: Propose a framework called Transition Matching Distillation (TMD), which distills large video diffusion models into lightweight generators capable of producing high-quality videos in just a few steps.

Transition Models: Rethinking the Generative Learning Objective

Zidong Wang (Chinese University of Hong Kong), Lei Bai (Shanghai AI Lab)

GenerationTransformerDiffusion modelScore-based ModelFlow-based ModelImageTextMultimodalityOrdinary Differential Equation

🎯 What it does: Designed a Transition Models (TiM) that learns state transitions over arbitrary time intervals, replacing the fixed-interval training of traditional PF-ODE or consistency models, enabling a single model to generate images under any sampling step count from single-step to multi-step.

Translating Signals to Languages for sEMG-Based Activity Recognition

Ming Wang (Lancaster University), Jun Liu (Lancaster University)

RecognitionTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderBiomedical Data

🎯 What it does: Convert continuous surface electromyography (sEMG) signals into discrete sequences similar to language, and use large language models (LLMs) to recognize motor activities.

TRANSPORTER: Transferring Visual Semantics from VLM Manifolds

Alexandros Stergiou (University of Twente)

GenerationExplainability and InterpretabilityRepresentation LearningTransformerVision Language ModelFlow-based ModelVideoText

🎯 What it does: This paper proposes the "Logits‑to‑Video (L2V)" task and the TRANSPORTER model, which visually interprets Vision‑Language Models (VLM) by generating videos that intuitively demonstrate changes in token logits.

TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model

Ao Li (Shandong University), Hu Wang (Mohamed bin Zayed University of Artificial Intelligence)

Computational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality

🎯 What it does: Propose a training-free prefix pruning method called TransPrune, based on dynamic transfer of visual tokens, for efficiently accelerating inference in large vision-language models

TRCoRSurg: Temporal-Relational Co-Reasoning for Surgical Video Triplet Recognition

Fang Li (Beihang University), Aimin Hao (Beihang University)

RecognitionGraph Neural NetworkVideo

🎯 What it does: This paper proposes a unified spatiotemporal relational reasoning framework for precise identification of triplets (instrument-action-target) in surgical videos.

TreeTeaming: Autonomous Red-Teaming of Vision-Language Models via Hierarchical Strategy Exploration

Chunxiao Li (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)

Adversarial AttackLarge Language ModelVision Language ModelMultimodality

🎯 What it does: This paper proposes TreeTeaming, an automatic red teaming framework based on policy trees, for discovering and evolving attack paths in vision-language models;

Tri-Modal Fusion Transformers for UAV-based Object Detection

Craig Iaboni (New Jersey Institute of Technology), Pramod Abichandani (New Jersey Institute of Technology)

Object DetectionTransformerMultimodality

🎯 What it does: Propose a three-modal (RGB, LWIR thermal, and event camera) joint detection framework that utilizes a dual-stream hierarchical vision Transformer to achieve cross-modal interaction and fusion.

Tri-Subspaces Disentanglement for Multimodal Sentiment Analysis

Chunlei Meng (Fudan University), Chun Ouyang (Fudan University)

ClassificationRecognitionTransformerMultimodality

🎯 What it does: Propose a three-subspace decoupling (Common, Submodally-Shared, Private) framework to model global consistency, partial shared interactions, and single-modal features in multimodal sentiment analysis;

TRIDENT: A Trimodal Cascade Generative Framework for Drug and RNA-Conditioned Cellular Morphology Synthesis

Rui Peng (Peking University), Jinzhuo Wang (Peking University)

GenerationData SynthesisDrug DiscoveryTransformerDiffusion modelAuto EncoderImageMultimodalityBiomedical Data

🎯 What it does: Propose a cascading generation framework TRIDENT that generates cell morphology images using drug and RNA information;

TriDF: Evaluating Perception, Detection, and Hallucination for Interpretable DeepFake Detection

Jian-Yu Jiang-Lin (National Taiwan University), Wen-Huang Cheng (National Taiwan University)

Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelImageVideoMultimodalityBenchmarkAudio

🎯 What it does: Proposed the TriDF benchmark to systematically evaluate the three-dimensional performance of DeepFake detection models in perception, detection, and interpretability.

TriLite: Efficient Weakly Supervised Object Localization with Universal Visual Features and Tri-Region Disentanglement

Arian Sabaghi (University of Antwerp), Jose Oramas (University of Antwerp)

Object DetectionComputational EfficiencyRepresentation LearningTransformerGenerative Adversarial NetworkContrastive LearningImage

🎯 What it does: Proposes TriLite, a single-stage weakly supervised object localization framework that trains only a small number of parameters.

TriSim: Tri-Dimensional Similarity Modeling with Extreme Value Theory for False-Negative Mitigation in Remote Sensing Image-Text Retrieval

Chengyu Zheng (University of British Columbia Okanagan), Shan Du (University of British Columbia Okanagan)

RetrievalTransformerVision Language ModelContrastive LearningMultimodalityBenchmark

🎯 What it does: Propose the TriSim framework, which constructs a three-dimensional similarity space (image-image, image-text, text-text), combines extreme value theory and Mahalanobis distance dual thresholds to identify and eliminate false negative samples in remote sensing image-text retrieval, and utilizes inner-model saliency differences to generate a mask for learning a gain matrix to strengthen discriminative regions.

TRivia: Self-supervised Fine-tuning of Vision-Language Models for Table Recognition

Junyuan Zhang (University of Hong Kong), Conghui He (Shanghai AI Laboratory)

RecognitionReinforcement LearningVision Language ModelImageTextMultimodality

🎯 What it does: This paper proposes TRivia, which performs self-supervised fine-tuning on pre-trained vision-language models (VLMs) using unlabeled table images to achieve high-quality table recognition.

TRM-VLA: Temporal-Aware Chain-of-Thought Reasoning and Memorization for Vision-Language-Action Models

Xiang Li (Tsinghua University), Shengjin Wang (Tsinghua University)

Robotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelMultimodalityTime SeriesChain-of-Thought

🎯 What it does: Propose the TRM-VLA framework, achieving temporal awareness and memory-based reasoning in vision-language-action models for robot manipulation through keyframe-triggered hierarchical chain-of-thought reasoning (KTR) and adjustable-grained context memory (GCM).

TROPHIES: Temporal Reconstruction of Places, Humans, and Cameras from Multi-view Videos

Jinpeng Liu (National University of Singapore), Xingyu Liu (National University of Singapore)

Pose EstimationOptimizationTransformerVideo

🎯 What it does: This paper proposes a unified framework named TROPHIES, which can simultaneously reconstruct dynamic humans, static scene geometry, and camera trajectories in multi-view videos, achieving physical consistency among them in the same global 4D space.

TruckDrive: Long-Range Autonomous Highway Driving Dataset

Filippo Ghilotti (Torc Robotics), Felix Heide (Torc Robotics)

Object DetectionObject TrackingDepth EstimationAutonomous DrivingImageVideoMultimodalityPoint Cloud

🎯 What it does: Built and released TruckDrive — a long-range multimodal dataset covering highway scenarios, containing over 475k synchronized LiDAR, radar, and high-resolution camera frames, along with 1km 2D annotations and 400m 3D annotations.

Trust-calibrated Collaborative Learning for Long-Tailed Visual Recognition

Hao Zhou (Naval University of Engineering), Tingjin Luo (National University of Defense Technology)

ClassificationRecognitionKnowledge DistillationMixture of ExpertsImage

🎯 What it does: Propose a multi-expert learning framework named Trust-Calibrated Collaborative Learning (TCL), specifically designed to address knowledge distillation errors and bias propagation in long-tail visual recognition.

TSTM: Temporal Segmentation for Task-relevant Mask in Visual Reinforcement Learning Generalization

Weicheng Du (Shandong University), Xiankai Lu (Shandong University)

SegmentationKnowledge DistillationRepresentation LearningRecurrent Neural NetworkReinforcement LearningVideo

🎯 What it does: Construct a segmentation network (TSTM) that utilizes temporal observations to extract task-related regions, and combine invariant representation learning with lightweight teacher-student distillation to enhance the generalization ability of visual reinforcement learning.

TTAPFormer: Robust Arbitrary Point Tracking via Transient Asynchronous Fusion of Frames and Events

Jiaxiong Liu (National University of Defense Technology), Dewen Hu (National University of Defense Technology)

Object TrackingTransformerMultimodality

🎯 What it does: This paper proposes a TAPFormer framework based on Transformer, achieving asynchronous temporal consistent fusion between frames and events for robust arbitrary point tracking.

TTL: Test-time Textual Learning for OOD Detection with Pretrained Vision-Language Models

Jinlun Ye (Sun Yat-sen University), Ruixuan Wang (Hong Kong University of Science and Technology)

Anomaly DetectionRepresentation LearningTransformerPrompt EngineeringContrastive LearningImageTextRetrieval-Augmented Generation

🎯 What it does: Propose the TTL framework, which dynamically captures OOD semantics during testing by learning adjustable text prompts without requiring pre-defined OOD labels, applicable to unlabelled test streams;

TTP: Test-Time Padding for Adversarial Detection and Robust Adaptation on Vision-Language Models

Zhiwei Li (NLPR & MAIS Institute of Automation Chinese Academy of Sciences), Qi Li (NLPR & MAIS Institute of Automation Chinese Academy of Sciences)

Domain AdaptationAdversarial AttackVision Language ModelContrastive LearningMultimodality

🎯 What it does: Propose a lightweight test-time padding (Test-Time Padding, TTP) defense framework to enhance the adversarial robustness of vision-language models such as CLIP against adversarial attacks without retraining.

TTRV: Test-Time Reinforcement Learning for Vision Language Models

Akshit Singh (Independent Researcher), M. Jehanzeb Mirza (MIT CSAIL)

Domain AdaptationLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality

🎯 What it does: Proposes TTRV — a framework that adaptively fine-tunes vision-language models using reinforcement learning during testing, completely independent of labeled data.

tttLRM: Test-Time Training for Long Context and Autoregressive 3D Reconstruction

Chen Wang (University of Pennsylvania), Yiwei Hu (Adobe Research)

GenerationData SynthesisNeural Radiance FieldGaussian SplattingImage

🎯 What it does: This paper proposes a large-scale reconstruction model called tttLRM based on Test-Time Training, which efficiently and scalably generates explicit 3D representations from multi-view images.

TUDSR: Twice Upsampling-Diffusion for Higher Super-Resolution

Zhiqiang Wu (East China Normal University), Xian Wei (East China Normal University)

Super ResolutionTransformerDiffusion modelGenerative Adversarial NetworkImage

🎯 What it does: Proposed the TUDSR (Twice Upsampling-Diffusion) framework, utilizing two-stage LoRA adaptation and cyclic block training to achieve super-resolution from low-resolution to high-resolution (e.g., 256×256→2048×2048) based on Stable Diffusion 2.1-base.

TUNA: Taming Unified Visual Representations for Native Unified Multimodal Models

Zhiheng Liu (Meta BizAI), Yuren Cong (Meta BizAI)

GenerationTransformerLarge Language ModelVision Language ModelFlow-based ModelAuto EncoderImageVideoTextMultimodality

🎯 What it does: Proposed a local unified multimodal model called TUNA, which constructs a continuous unified visual representation by combining a VAE encoder with a representation encoder, and achieves image/video understanding, generation, and editing through an LLM decoder and a flow-matching head.

Tunable Soft Equivariance with Guarantees

Md Ashiqur Rahman (Purdue University), Raymond A. Yeh (Purdue University)

ClassificationSegmentationConvolutional Neural NetworkTransformerImageSequential

🎯 What it does: Propose a tunable soft equivariant layer, which controls equivariance by projecting network weights into a specific subspace, and provides an error upper bound. Subsequently, this layer is applied to pre-trained models such as ViT, ResNet, and SegFormer to achieve tasks including classification, segmentation, and trajectory prediction.

Turbo-GS: Accelerating 3D Gaussian Fitting for High-Resolution Radiance Fields

Ankit Dhiman (Indian Institute of Science), Srinath Sridhar (Brown University)

Computational EfficiencyGaussian SplattingImage

🎯 What it does: This paper proposes Turbo-GS, a training acceleration framework for 3D Gaussian Splatting in high-resolution (4K) scenarios. It significantly reduces computational costs and gradient sparsity per iteration through dilated rendering, convergence-aware Gaussian budget control, and adaptive densification strategies based on position information and color gradients, thereby improving training speed and image quality.

Turning Pre-Trained Vision Transformers into End-to-End Histopathology Whole Slide Image Models for Survival Prediction

Jiawen Li (Tsinghua Shenzhen International Graduate School (SIGS), Tsinghua University), Yonghong He (Tsinghua Shenzhen International Graduate School (SIGS), Tsinghua University)

ClassificationTransformerSupervised Fine-TuningImageBiomedical Data

🎯 What it does: Proposed an E2E-ViT conversion strategy that directly transforms pre-trained low-resolution ViT models into end-to-end models capable of processing full-resolution pathological whole slide images (WSI), achieving survival prediction through fine-tuning.

Tutor-Student Reinforcement Learning: A Dynamic Curriculum for Robust Deepfake Detection

Zhanhe Lei (School Of Computer Science Wuhan University), Dengpan Ye (School Of Computer Science Wuhan University)

Anomaly DetectionReinforcement LearningImageVideo

🎯 What it does: Proposed and implemented the Tutor-Student Reinforcement Learning (TSRL) framework, which dynamically assigns continuous weights to each training sample, forming an adaptive curriculum in the training of deepfake detection models.

TV2TV: A Unified Framework for Interleaved Language and Video Generation

Xiaochuang Han (Meta FAIR), Emily Dinan (Meta FAIR)

GenerationTransformerLarge Language ModelMixture of ExpertsFlow-based ModelAuto EncoderVideoTextMultimodalityOrdinary Differential Equation

🎯 What it does: Propose TV2TV, a unified generative framework that decomposes video generation into an alternating process of text generation and video frame generation.

TVHighlights: LLM-Guided Human-Free Collaborative Training for Video Highlight Detection in Movies and TV Dramas

Qi Qiu (HUJING Digital Media & Entertainment Group), Yanlong Du (HUJING Digital Media & Entertainment Group)

Object DetectionTransformerLarge Language ModelVideoTextMultimodalityBenchmark

🎯 What it does: This paper constructs a new movie/TV show video highlight detection dataset called TVHighlights and proposes the LTV-HD framework, which achieves fully unsupervised highlight detection by leveraging LLMs to generate pseudo-labels, noise-robust training, and highlight pattern induction.

TWEO: Transformers Without Extreme Outliers Enables FP8 Training And Quantization For Dummies

Guang Liang (Nanjing University), Jianxin Wu (Nanjing University)

Computational EfficiencyRepresentation LearningTransformerImageText

🎯 What it does: Propose the TWEO loss function to eliminate extreme activation outliers in Transformers, making FP8 training and quantization feasible.

Twin-T & TwintVQA: A Reliable Structure-Detail Separating VLM and a Comprehensive Benchmark for Chart and Table Tasks

Jiahua Bao (Harbin Institute of Technology), Jie Liu (Harbin Institute of Technology)

Vision Language ModelContrastive LearningImageTextBenchmark

🎯 What it does: Propose Twin-T, a two-stage expert visual language model designed for chart and table tasks;

TWINGS: Thin Plate Splines Warp-aligned Initialization for Sparse-View Gaussian Splatting

Hyeseong Kim (Yonsei University), Dosik Hwang (Yonsei University)

GenerationOptimizationGaussian SplattingImageBenchmark

🎯 What it does: TWINGS aligns monocular depth back-projected points with control points derived from multi-view triangulation using Thin Plate Splines (TPS), providing dense and geometrically consistent initialization for 3D Gaussian Splatting under sparse views;

U-Mind: A Unified Framework for Real-Time Multimodal Interaction with Audiovisual Generation

Xiang Deng (Tsinghua University), Yebin Liu (Tsinghua University)

GenerationData SynthesisPose EstimationComputational EfficiencyTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelAuto EncoderGaussian SplattingVideoTextMultimodalityChain-of-ThoughtAudio

🎯 What it does: Propose U-Mind, a real-time full-stack multimodal interaction framework capable of generating text, speech, pose, and synthetic video within a single loop;

U^2Flow: Uncertainty-Aware Unsupervised Optical Flow Estimation

Xunpei Sun (Sun Yat-sen University), Gang Chen (Sun Yat-sen University)

Recurrent Neural NetworkOptical FlowVideo

🎯 What it does: Propose a self-supervised recursive optical flow network called U Flow, which can simultaneously predict pixel-level optical flow and corresponding uncertainty, and adaptively refine the optical flow using the predicted uncertainty.

U4D: Uncertainty-Aware 4D World Modeling from LiDAR Sequences

Xiang Xu (Nanjing University of Aeronautics and Astronautics), Qingshan Liu (Nanjing University of Posts and Telecommunications)

GenerationData SynthesisDiffusion modelWorld ModelPoint Cloud

🎯 What it does: Propose the U4D framework, which utilizes a pre-trained segmentation model to estimate uncertainty and generates a four-dimensional LiDAR world through a two-stage diffusion process (first reconstructing high-entropy regions, then completing the overall scene), achieving geometrically detailed and temporally consistent scene synthesis.

UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement

Weiqi Li (Peking University), Shijie Zhao (ByteDance Inc)

RestorationSuper ResolutionTransformerVision Language ModelRectified FlowAuto EncoderImageTextMultimodality

🎯 What it does: Designed and trained a unified audio-visual language model UARE, capable of simultaneously performing image quality assessment, image restoration, and enhancement, and achieving guidance of the restoration process through quality assessment text via two-stage training.

UAST: Unified Active Search and Tracking for Arbitrary Targets with UAVs

Liang Qin (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)

Object TrackingRobotic IntelligenceConvolutional Neural NetworkImageMultimodality

🎯 What it does: Proposed the UAST framework, achieving active search and continuous tracking of any target by drones using RGB-D sensors without mapping or multi-stage planning;