🎯 What it does: Propose a two-stage segmentation framework ViT-P: first generate masks using a class-agnostic mask generator, then use a Vision Transformer-based point classifier to precisely classify the masks, thereby significantly improving segmentation accuracy.
🎯 What it does: Propose Contrastive Fusion (ConFu), a contrastive learning framework that jointly aligns single-modal representations with their fused representations.
🎯 What it does: This paper improves upon adversarial attack algorithms based on sign gradients by proposing the MDCS (Monotonically Decreasing Coordinate Step Size) strategy, which is embedded into classical methods such as I-FGSM and MI-FGSM, significantly enhancing the stability and transferability of attacks.
The Power of Prior: Training-Free Open-Vocabulary Semantic Segmentation with LLaVA
Bingfeng Zhang (China University Of Petroleum East China), Jimin Xiao (Xjtlu)
CodeSegmentationLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageMultimodalityBenchmark
🎯 What it does: This paper proposes a fully training-free open-source lexical semantic segmentation method called FSeg-LLaVA, which leverages the internal prior knowledge of the multimodal large language model LLaVA. It identifies foreground classes through a question-answering pipeline, extracts visual-textual responses, refines them through prototyping, and generates point/box prompts for SAM to produce the final mask.
TherA: Thermal-Aware Visual-Language Prompting for Controllable RGB-to-Thermal Infrared Translation
Dong-Guw Lee (Seoul National University), Ayoung Kim (Seoul National University)
CodeImage TranslationTransformerPrompt EngineeringVision Language ModelDiffusion modelMultimodality
🎯 What it does: Proposed a controllable RGB-to-thermal infrared (TIR) image translation framework called TherA, capable of generating TIR images that comply with thermal physics laws, and supporting control via text and reference images
CodeGenerationTransformerReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes ThinkGen, a visual generation framework that guides diffusion models to generate high-quality images through chain-of-thought (CoT) mechanisms of multimodal large language models.
Thinking with Programming Vision: Towards a Unified View for Thinking with Images
Zirun Guo (Zhejiang University), Tao Jin (Zhejiang University)
CodeAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposed the CodeVision framework, enabling multimodal large language models to call arbitrary image tools through code generation, achieving more flexible and scalable image thinking capabilities.
CodeGenerationPrompt EngineeringVision Language ModelDiffusion modelVideoTextMultimodalityBenchmark
🎯 What it does: Propose the 'Thinking with Video' paradigm, using video generation models for multimodal reasoning and constructing the VideoThinkBench benchmark.
TIM: Temporal Decoupling with Iterative Mutual-Refinement Model for Longitudinal Radiology Report Generation
Yiheng Dong (Huazhong University of Science and Technology), Xin Yang (Huazhong University of Science and Technology)
CodeGenerationTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityElectronic Health Records
🎯 What it does: Proposed the TIM model, utilizing a two-stage framework with temporal decoupling and mutual iterative refinement to generate longitudinal X-ray imaging reports.
🎯 What it does: Propose a text-agnostic reverse attack method called TINA, which utilizes DDIM inversion and self-consistent optimization to recover erased concepts in text-to-image diffusion models where text has been erased;
Too Vivid to Be Real? Benchmarking and Calibrating Generative Color Fidelity
Zhengyao Fang (Harbin Institute of Technology), Wenjie Pei (Harbin Institute of Technology)
CodeGenerationVision Language ModelDiffusion modelImageMultimodalityBenchmark
🎯 What it does: This paper addresses the issue of color authenticity in text-to-image generation by proposing a large-scale color authenticity dataset (CFD), a multi-modal color authenticity metric (CFM), and an untrained dynamic regulator (CFR) to enhance the color authenticity of generated images.
CodeRetrievalRepresentation LearningTransformerContrastive LearningBiomedical Data
🎯 What it does: Propose TopoSlide, a self-supervised framework for whole-slide image representation learning that leverages persistent homology (topological data analysis) to simultaneously capture local pathological features and global spatial organization;
🎯 What it does: Proposed a low-cost and efficient time learning method (LETL), integrated into the single-stream visual tracker LETrack based on DeiT-Tiny, to achieve real-time robustness for UAV visual tracking.
Toward Real-world Infrared Image Super-Resolution: A Unified Autoregressive Framework and Benchmark Dataset
Yang Zou (Northwestern Polytechnical University), Jinyuan Liu (Dalian University of Technology)
CodeSuper ResolutionAuto EncoderImageBenchmark
🎯 What it does: This paper proposes a unified autoregressive framework, Real-IISR, for real-world infrared image super-resolution, and constructs the FLIR-IISR dataset.
🎯 What it does: Propose an incremental unified multimodal anomaly detection framework named IB-IUMAD, addressing catastrophic forgetting caused by pseudo feature interference and redundant information in cross-modal features.
Towards Balanced Multi-Modal Learning in 3D Human Pose Estimation
Mengshi Qi (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
CodePose EstimationTransformerMultimodality
🎯 What it does: Proposed a balanced multi-modal learning framework that evaluates the contribution of each modality in 3D pose estimation using Shapley values, and balances the learning rates of different modalities early in training through adaptive weight constraints (AWC) based on the Fisher information matrix, ultimately achieving collaborative learning across four modalities: RGB, LiDAR, mmWave, and WiFi.
Towards Cross-Modal Preservation, Consistency and Alignment for Privacy-Preserving Visible-Infrared Person Re-Identification
Yudi Xie (Wuhan University), Mang Ye (Wuhan University)
CodeRecognitionPose EstimationRetrievalSafty and PrivacyKnowledge DistillationConvolutional Neural NetworkGenerative Adversarial NetworkImageMultimodality
🎯 What it does: Propose a cross-modal anonymizable person re-identification framework named PPA, achieving re-identification of visible-infrared images while protecting privacy.
Towards Efficient Medical Reasoning with Minimal Fine-Tuning Data
Xinlin Zhuang (East China Normal University), Imran Razzak (MBZUAI)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningBiomedical Data
🎯 What it does: For medical reasoning tasks, we propose a data selection framework called DIQ (Difficulty-Influence Quadrant), which selects a subset that significantly improves model performance under small sample sizes by jointly measuring the reasoning difficulty of samples and gradient influence, thus achieving results comparable to full training with only a small amount of fine-tuning.
🎯 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)
CodeAdversarial 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 Intrinsic-Aware Monocular 3D Object Detection
Zhihao Zhang (Michigan State University), Xiaoming Liu (Michigan State University)
CodeObject 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.
CodeReinforcement 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)
CodeRecognitionData 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.
🎯 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 Multimodal Large Language Models Against Jailbreak Attacks
Ziyi Yin (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
CodeSafty 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 Stable Federated Continual Test-Time Adaptation in Wild World
Liwen Wang (Anhui University), Zhe Jin (University of Nottingham)
CodeClassificationSegmentationDomain 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 Streaming Referring Video Segmentation via Large Language Model
Wenkang Zhang (Imperial College London), Jiankang Deng (Alibaba Group)
CodeSegmentationLarge 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.
🎯 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.
🎯 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.
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)
CodeDepth 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.
🎯 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)
CodeAutonomous 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.
Training High-Level Schedulers with Execution-Feedback Reinforcement Learning for Long-Horizon GUI Automation
Zehao Deng (Soochow University), Gongshen Liu (Shanghai Jiao Tong University)
CodeTransformerLarge 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-free Motion Factorization for Compositional Video Generation
Zixuan Wang (Sichuan University), Yinjie Lei (Sichuan University)
CodeGenerationLarge 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-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)
CodeClassificationKnowledge 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.
🎯 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.
TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model
Ao Li (Shandong University), Hu Wang (Mohamed bin Zayed University of Artificial Intelligence)
CodeComputational 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
TRivia: Self-supervised Fine-tuning of Vision-Language Models for Table Recognition
Junyuan Zhang (University of Hong Kong), Conghui He (Shanghai AI Laboratory)
CodeRecognitionReinforcement 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.
🎯 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.
🎯 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)
CodeDomain 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.
🎯 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.
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)
CodeClassificationTransformerSupervised 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.
🎯 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.
U-Mind: A Unified Framework for Real-Time Multimodal Interaction with Audiovisual Generation
Xiang Deng (Tsinghua University), Yebin Liu (Tsinghua University)
CodeGenerationData 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;
Xunpei Sun (Sun Yat-sen University), Gang Chen (Sun Yat-sen University)
CodeRecurrent 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.
🎯 What it does: Design and implement a lightweight unified convolutional attention network (UCAN), which expands the receptive field while maintaining efficiency by fusing convolution and attention mechanisms for single-image super-resolution.
UETrack: A Unified and Efficient Framework for Single Object Tracking
Ben Kang (Dalian University of Technology), Huchuan Lu (City University of Hong Kong)
CodeObject TrackingKnowledge DistillationTransformerMixture of ExpertsVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose a unified and efficient single-object tracking framework called UETrack, which supports multi-modal inputs such as RGB, Depth, Thermal, Event, and Language;
🎯 What it does: Propose a block-based video coding framework DCVC-UF, which utilizes a cross-frame interaction module to compress multiple frames into a single latent representation, and achieves parallel decoding through frame-specific decoders, significantly improving encoding and decoding throughput.
🎯 What it does: Propose the AEIC framework, utilizing a shallow encoder and first-order diffusion decoder to achieve ultra-low bitrate (<0.05 bpp) image compression, balancing encoding efficiency and perceptual quality;
Ultrasound-CLIP: Semantic-Aware Contrastive Pre-training for Ultrasound Image-Text Understanding
Jiayun Jin (Hangzhou City University), Binbin Zhou (Hangzhou City University)
CodeClassificationRetrievalTransformerVision Language ModelContrastive LearningMultimodalityBiomedical DataUltrasound
🎯 What it does: Constructed the first large-scale ultrasound image-text dataset US-365K, and proposed the Ultrasound-CLIP semantic-aware contrastive learning framework, which uses the UDT semantic ontology to guide image-text alignment.
Qi Zhang (University of Amsterdam), Martin R. Oswald (University of Amsterdam)
CodeConvolutional Neural NetworkGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: Propose an online SLAM system, Unblur-SLAM, which simultaneously handles motion blur and defocus blur, and achieves clear 3D reconstruction from blurred inputs through a single-frame deblurring network, 3D Gaussian splatting, and adaptive optimization.
Uncertainty-Aware Exploratory Direct Preference Optimization for Multimodal Large Language Models
Huatian Zhang (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
CodeOptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: Improve direct preference optimization (DPO) by introducing an exploration mechanism based on knowledge uncertainty to reduce visual hallucinations in multimodal large language models (MLLMs).
CodeKnowledge DistillationTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose a uncertainty-aware knowledge distillation framework named Beta-KD, which distills large multimodal large language models (e.g., MobileVLM, Qwen-VL) into smaller student models.
🎯 What it does: Investigate the theoretical foundations of task arithmetic, propose Task-Feature Specialization (TFS), and demonstrate its relationship with weight decoupling and weight vector orthogonality, followed by designing an orthogonal regularization method called OrthoReg to improve task arithmetic performance.
🎯 What it does: This paper systematically analyzes the interaction between instantaneous velocity (v) and average velocity (u) in MeanFlow training, proposing to improve one-shot generation performance by first accelerating the learning of v and gradually weighting the time intervals of u.
🎯 What it does: This paper proposes a single-stage unified distillation and adaptation framework, Uni-DAD, for rapidly generating high-quality images in few-step and few-sample scenarios.
🎯 What it does: This paper proposes a two-stage heterogeneous method called UniME for precise segmentation of brain tumors under missing modality conditions.
UniChange: Unifying Change Detection with Multimodal Large Language Model
Xu Zhang (Nankai University), Xiang Li (Nankai University)
CodeSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Designed a unified change detection framework called UniChange based on a multimodal large language model, which can simultaneously perform binary change detection and semantic change detection, and supports joint training with multi-source data.
UniComp: Rethinking Video Compression Through Informational Uniqueness
Chao Yuan (Meituan Inc.), Lin Ma (Meituan Inc.)
CodeCompressionTransformerVideo
🎯 What it does: Proposes a video compression framework called UniComp based on information uniqueness, aiming to maximize the fidelity of video representations under limited computational budgets.
Unified Camera Positional Encoding for Controlled Video Generation
Cheng Zhang (Monash University), Jianfei Cai (Monash University)
CodeGenerationTransformerDiffusion modelVideoText
🎯 What it does: This paper proposes a Unified Camera Position Encoding (UCPE), which enables video diffusion Transformers to precisely control camera parameters such as pose, focal length, and distortion by combining relative ray encoding and absolute orientation encoding, thereby achieving high-quality, controllable text-to-video generation.
🎯 What it does: Proposed the USO unified model, achieving unified generation for subject customization, style customization, and joint customization of both, while enhancing generation quality through cross-task co-decomposition and auxiliary style rewards.
UniFusion: A Unified Image Fusion Framework with Robust Representation and Source-Aware Preservation
Xingyuan Li (Zhejiang University), Jinyuan Liu (Dalian Maritime University)
CodeTransformerAuto EncoderContrastive LearningImageMultimodalityBiomedical Data
🎯 What it does: Propose a unified image fusion framework named UniFusion, leveraging DINOv3 as a general-purpose semantic feature extractor, combined with adaptive Adapter, reconstruction alignment mechanism, and dual-layer optimization to achieve high-quality cross-task fusion.
Unifying Language-Action Understanding and Generation for Autonomous Driving
Xinyang Wang (Zhejiang University), Wei Chen (Li Auto)
CodeAutonomous DrivingTransformerLarge Language ModelVision-Language-Action ModelImageVideoTextMultimodality
🎯 What it does: Designed and implemented the LinkVLA model, unifying language instructions and trajectory actions into a single discrete codebook, and achieving efficient, aligned closed-loop driving through bidirectional learning and coarse-to-fine generation.
UniGenDet: A Unified Generative-Discriminative Framework for Co-Evolutionary Image Generation and Generated Image Detection
Yanran Zhang (Tsinghua University), Jie Zhou (Tsinghua University)
CodeGenerationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelFlow-based ModelImageTextMultimodality
🎯 What it does: Proposed a unified generative-discriminative framework called UniGenDet, achieving synergistic evolution between image generation and generated image detection.
UniGeoSeg: Towards Unified Open-World Segmentation for Geospatial Scenes
Shuo Ni (Beijing Institute of Technology), Jing Zhang (Wuhan University)
CodeSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Proposed the GeoSeg-1M million-scale remote sensing instruction-driven segmentation dataset and the GeoSeg-Bench benchmark, along with the unified framework UniGeoSeg;
UniMMAD: Unified Multi-Modal and Multi-Class Anomaly Detection via MoE-Driven Feature Decompression
Yuan Zhao (Dalian University of Technology), Xiaoqi Zhao (Nanyang Technological University)
CodeAnomaly DetectionMixture of ExpertsMultimodalityBenchmark
🎯 What it does: Propose UniMMAD, a unified multi-modal multi-class unsupervised anomaly detection framework that can handle different sensor combinations and classes within a single model, supporting continual learning.
UniRain: Unified Image Deraining with RAG-based Dataset Distillation and Multi-objective Reweighted Optimization
Qianfeng Yang (Dalian Polytechnic University), Jiangxin Dong (Nanjing University of Science and Technology)
CodeRestorationMixture of ExpertsVision Language ModelImageRetrieval-Augmented Generation
🎯 What it does: The paper proposes UniRain, a unified image de-raining framework capable of handling both raindrops and rain streaks during the day and night.
🎯 What it does: Propose a two-stage 3D texture generation framework called UniTEX, which first generates multi-view images using a large-scale Diffusion Transformer, and then directly regresses complete textures in 3D space through a Large Texturing Model.
Junjian Li (Central South University), Jianxin Wang (Central South University)
CodeClassificationTransformerLarge Language ModelVision Language ModelBiomedical Data
🎯 What it does: This paper proposes a dynamic knowledge-guided multi-instance learning framework called DyKo for whole slide image (WSI) classification under few-shot scenarios.
UniVerse: A Unified Modulation Framework for Segmentation-Free, Disentangled Multi-Concept Personalization
Quynh Phung (University of Maryland, College Park), Jia-Bin Huang (University of Maryland, College Park)
CodeGenerationTransformerVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Proposes UniVerse, a unified modulation framework that can extract visual and textual conditions from reference images and text without requiring segmentation masks, enabling the decomposition and combination of multiple concepts;
🎯 What it does: Implant a hidden backdoor into a single diffusion model, enabling the synthetic data it generates to automatically propagate the backdoor during subsequent classifier training;
🎯 What it does: Propose MonoCoP, a monocular 3D object detection framework that adaptively leverages geometric correlations among 3D attributes (size, orientation, depth).
🎯 What it does: Propose Parameter Inheritance HyperNetwork (PITH), which uses a learnable projection matrix to directly map weights from publicly pre-trained models to the target model, achieving zero-shot initialization;
🎯 What it does: Proposed an efficient pixel-level feature upsampling method called UPLiFT, which maps low-resolution visual backbones to high-resolution features using iterative convolutions and a local attender.
UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking
Hao Wu (Eastern Institute of Technology), Xiaoyu Shen (Eastern Institute of Technology)
CodeObject TrackingComputational EfficiencyTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: Proposes UTPTrack, a unified token pruning framework that simultaneously compresses the search region, dynamic template, and static template in state-of-the-art Transformer visual tracking, achieving efficient real-time tracking.
V-Attack: Targeting Disentangled Value Features for Controllable Adversarial Attacks on LVLMs
Sen Nie (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences), Xilin Chen (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences)
CodeAdversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a controllable adversarial attack method called V-Attack based on the internal value features of Transformer.
🎯 What it does: Proposes the V2U4Real dataset and benchmark, which records real-world large-scale multimodal data from vehicle and drone collaborative perception, supporting single-machine and collaborative 3D detection and tracking tasks.
🎯 What it does: Lightweight post-training of existing autoregressive image generation models, aligning the generator directly in pixel space using variational inference and reinforcement learning to enhance image quality and diversity.
🎯 What it does: Diagnose the asynchronous policy conflict problem in visual autoregressive (VAR) models within reinforcement learning and propose an improved RL framework, significantly enhancing the quality of text rendering and human preference scores.
CodeCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoBenchmark
🎯 What it does: Proposes a dynamic pruning method (Variation-aware Vision Token Dropping, V Drop 2) based on the variation degree of visual tokens within LLMs, discarding layer-by-layer the tokens with the least variation during inference to significantly reduce the number of visual tokens and improve inference speed.
🎯 What it does: This paper proposes a training-agnostic acceleration method called VDE (Velocity Decomposition and Estimation) to speed up the sampling process of Rectified Flow models;
Venus: Benchmarking and Empowering Multimodal Large Language Models for Aesthetic Guidance and Cropping
Tianxiang Du (Peking University), Yuxin Peng (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a two-phase framework Venus, which first trains an MLLM to achieve aesthetic guidance using the AesGuide dataset, then activates its cropping capability through Chain-of-Thought aesthetic reasoning, enabling interpretable interactive aesthetic cropping.
🎯 What it does: This paper proposes a 3D lung vessel segmentation model named VesMamba, which can automatically extract vascular structures from CT images;
CodeGenerationData SynthesisRobotic IntelligenceTransformerLarge Language ModelMixture of ExpertsVision-Language-Action ModelVideoTextMeshAudio
🎯 What it does: Built ViBES, a 3D conversational agent capable of simultaneously understanding speech and text while generating corresponding facial expressions and full-body movements.
🎯 What it does: Propose a 1D Transformer image tokenizer VibeToken and its corresponding autoregressive generative model VibeToken-Gen, which can adapt to arbitrary resolutions and aspect ratios, achieving efficient multi-resolution image generation.
VideoCoF: Unified Video Editing with Temporal Reasoner
Xiangpeng Yang (University of Technology Sydney), Qiang Wu (University of Technology Sydney)
CodeGenerationTransformerLarge Language ModelDiffusion modelAuto EncoderVideoMultimodalityBenchmarkChain-of-ThoughtOrdinary Differential Equation
🎯 What it does: Propose the VideoCoF framework, leveraging Chain-of-Frames (insight → reasoning → editing) to achieve unified and precise video editing;
VideoRealBench: A Chain-of-Thought Realism Evaluation Benchmark for Generated Human-Centric Videos
Min Yang (Nanjing University), Limin Wang (Nanjing University)
CodeExplainability and InterpretabilitySupervised Fine-TuningVision Language ModelVideoTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a high-quality human-centered video authenticity evaluation benchmark, VideoRealBench, which includes the VideoRealDataset dataset and an automatic evaluator, VideoRealEval.
VideoSSR: Video Self-Supervised Reinforcement Learning
Zefeng He (Shanghai Artificial Intelligence Laboratory), Yu Cheng (Chinese University of Hong Kong)
CodeTransformerLarge Language ModelReinforcement LearningVideoMultimodalityBenchmark
🎯 What it does: Proposed a self-supervised reinforcement learning framework called VideoSSR that leverages video's own information, addressing the data scarcity issue in training large multimodal language models for video understanding.
VidPrism: Heterogeneous Mixture of Experts for Image-to-Video Transfer
Rui Lin (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
CodeRecognitionTransformerMixture of ExpertsVision Language ModelVideo
🎯 What it does: Propose VidPrism, a heterogeneous Mixture-of-Experts (MoE) framework, to migrate pre-trained vision-language models to video understanding tasks, addressing the homogenization problem of traditional MoE experts.