🎯 What it does: Propose a test-time training method (TTT-MIM) based on masked image modeling for adapting single images under distribution shifts in image denoising.
UDA-Bench: Revisiting Common Assumptions in Unsupervised Domain Adaptation Using a Standardized Framework
Tarun Kalluri (UC San Diego), Manmohan Chandraker (UC San Diego)
CodeDomain AdaptationTransformerImageBenchmark
🎯 What it does: Proposed the UDA-Bench framework and conducted large-scale empirical studies to investigate the impact of backbone networks, the amount of unlabeled data, and pre-training data on the performance of unsupervised domain adaptation (UDA).
🎯 What it does: Proposed UDiffText, a unified framework that leverages character-level diffusion models to achieve high-quality text synthesis in arbitrary images, addressing the spelling error issues of existing text-to-image (T2I) models.
🎯 What it does: Based on the Universal Manifold Embedding (UME) framework, this paper proposes a UME-compatible feature extraction method. By combining a sampling balance module and UME contrast learning, we construct a robust point cloud registration pipeline (UMERegRobust) without relying on RANSAC. Additionally, we introduce the RotKITTI/RotnuScenes benchmark dataset tailored for large rotation scenarios.
CodeClassificationSegmentationRetrievalTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Designed and trained a unified multi-granularity CLIP model named UMG-CLIP, covering image, region, and pixel-level alignment, and built a large-scale multi-granularity dataset UMG-41M through an automated workflow to support multi-task pretraining;
Uncertainty Calibration with Energy Based Instance-wise Scaling in the Wild Dataset
Mijoo Kim (Chung-Ang University), Junseok Kwon (Chung-Ang University)
CodeClassificationDomain AdaptationImage
🎯 What it does: Propose an instance-level adaptive temperature scaling method based on an energy model for post-hoc uncertainty calibration in multi-class classification tasks, maintaining robustness under distribution shifts (including covariate shift and semantic shift).
Uncertainty-aware sign language video retrieval with probability distribution modeling
Xuan Wu (Sichuan University), Keren Fu (Peking University)
CodeRetrievalConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Proposes a sign language video retrieval method called UPRet, which represents sign language videos and text as probability distributions and uses optimal transport for matching.
CodeRepresentation LearningLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a general framework based on category theory for understanding and enhancing multi-attribute compositional learning in vision-language models (multi-compositional zero-shot learning), and implements an attention-based realization named CatCom; subsequently, the framework is used to evaluate the compositional properties of various large language models in latent spaces.
🎯 What it does: Proposes UniFS, a unified few-shot instance-aware model that integrates four major tasks—object detection, instance segmentation, pose estimation, and counting—through a dynamic point representation learning framework.
🎯 What it does: Proposed UniM AE, a multi-modal self-supervised pre-training framework that achieves image and LiDAR fusion and reconstruction through unified 3D volume space and multi-modal 3D interaction modules.
UniMD: Towards Unifying Moment Retrieval and Temporal Action Detection
Yingsen Zeng (Meituan Inc), Lin Ma (Meituan Inc)
CodeObject DetectionRetrievalConvolutional Neural NetworkVision Language ModelVideoTextMultimodality
🎯 What it does: Propose a unified framework UniMD, which uses a single model to simultaneously perform Temporal Action Detection (TAD) and Moment Retrieval (MR), achieving collaborative learning for both tasks through unified queries, shared CLIP text encoder, and query-related decoders.
CodeRestorationConvolutional Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: A unified low-level visual processing framework called UniProcessor based on text induction was constructed, capable of handling 30 types of image degradation (such as noise, blur, rainy weather, compression artifacts, etc.) within a single model, and supporting independent or step-by-step control of different degradations through natural language prompts.
CodeSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Propose a post-training defense method called UNIT, which automatically approximates upper bounds of activation distributions for each neuron and trims activations exceeding these bounds during inference, thereby eliminating the impact of backdoor triggers while maintaining model accuracy.
🎯 What it does: Proposed a unified multi-head audio-driven 3D facial animation model called UniTalker, which can learn from multiple differently annotated datasets.
🎯 What it does: Unify multi-vehicle trajectory prediction datasets, models, and evaluation methods, and study the impact of cross-domain generalization and data scale on performance within this framework.
🎯 What it does: Proposes a unified voxelization framework, UniVoxel, which explicitly learns scene geometry (SDF), material (albedo, roughness), and illumination (local lighting field) through a lightweight MLP that jointly infers them in voxel space.
Unleashing the Power of Prompt-driven Nucleus Instance Segmentation
Zhongyi Shui (Zhejiang University), Lin Yang (Westlake University)
CodeSegmentationTransformerSupervised Fine-TuningPrompt EngineeringImageBiomedical Data
🎯 What it does: Designed the PromptNucSeg framework, achieving nuclear instance segmentation without post-processing by utilizing an automated point prompter and fine-tuned SAM.
Unlocking Attributes' Contribution to Successful Camouflage: A Combined Textual and Visual Analysis Strategy
Hong Zhang (Beihang University), Yifan Yang (Beihang University)
CodeSegmentationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Investigated the impact of camouflage attributes on segmentation performance in concealed object segmentation (COS), constructed the first COD-TAX dataset containing image descriptions and attribute contributions, and proposed the ACUMEN framework that integrates textual and visual information.
🎯 What it does: Propose a server-side data distillation framework FedDGM, allowing clients to train only small proxy models, while the server utilizes a pre-trained deep generative model (StyleGAN-XL) and matching training trajectory techniques to generate synthetic data and train a larger global model on this data.
Hu Yu (University of Science and Technology of China), Feng Zhao (Alibaba Group)
CodeGenerationDiffusion modelImage
🎯 What it does: This paper systematically analyzes the bias caused by the commonly used constant weight loss in diffusion model training, and proposes a weight strategy based on the inverse square root of the signal-to-noise ratio, significantly improving sample quality and training/sampling efficiency.
🎯 What it does: Proposed an unsupervised learning-based low-light video enhancement framework called UDU-Net, which expands the maximum a posteriori (MAP) problem into a trainable deep network, optimizing spatial (Intra) and temporal (Inter) visual priors in separate stages;
🎯 What it does: Propose the UDI (Unsqueeze [CLS] Bottleneck) method, leveraging self-distillation and multi-scale target distributions in Vision Transformers to enhance the diversity and richness of visual representations.
Ruodai Cui (Qualcomm Technologies, Inc.), Guosheng Hu (University of Bristol)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: Proposes a fully unsupervised exposure correction framework that utilizes images from multi-exposure sequences to mutually serve as training targets, achieving pixel-level exposure adjustment.
🎯 What it does: Propose an unsupervised multimodal medical image registration method called INNReg, which achieves image modality conversion through invertible neural networks and performs registration using the converted monomodal images.
Unveiling Advanced Frequency Disentanglement Paradigm for Low-Light Image Enhancement
Kun Zhou (CUHK-Shenzhen), Jiangbo Lu (SmartMore Corporation)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: Proposed a general frequency decoupling optimization paradigm for low-light image enhancement, which can improve performance by adding only a small number of parameters to existing models.
CodeRestorationFederated LearningSafty and PrivacyImage
🎯 What it does: This paper studies gradient inversion attacks on spiking neural networks (SNN) in federated learning and proposes the ISG attack method that can reconstruct training data from gradients.
Unveiling Typographic Deceptions: Insights of the Typographic Vulnerability in Large Vision-Language Models
Hao Cheng (Hong Kong University of Science and Technology), Renjing Xu (Hong Kong University of Science and Technology)
CodeExplainability and InterpretabilityAdversarial AttackLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Systematically evaluate and mitigate the vulnerability of large vision-language models (LVLM) to typographic attacks; propose the largest-scale typographic dataset TypoD, and experimentally verify the mechanism by which typographic attacks affect LVLM; further propose methods to reduce the attack's impact through enriched text prompts and cross-modal attention matching.
🎯 What it does: This paper addresses the challenge of urban flood detection by constructing the UW-Bench dataset and proposing the Large-Small Model Co-Adapter framework.
Using My Artistic Style? You Must Obtain My Authorization
Xiuli Bi (Chongqing University of Posts and Telecommunications), Bin Xiao (Chongqing University of Posts and Telecommunications)
CodeImage TranslationSafty and PrivacyAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: Propose an adversarial perturbation-based artistic style protection scheme (ASPS) that reduces the style transfer effectiveness of unauthorized models without compromising the output quality of authorized models.
Vary: Scaling up the Vision Vocabulary for Large Vision-Language Models
Haoran Wei (MEGVII Technology), Xiangyu Zhang (Huazhong University of Science and Technology)
CodeTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposed a method called Vary to expand the visual vocabulary of large vision-language models (LVLM), aiming to improve performance on specific tasks, especially those requiring fine-grained perception.
VCD-Texture: Variance Alignment based 3D-2D Co-Denoising for Text-Guided Texturing
Shang Liu (DAMO Academy Alibaba Group), Fan Wang (DAMO Academy Alibaba Group)
CodeRestorationGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageTextMesh
🎯 What it does: Developed a 3D-2D collaborative denoising framework called VCD-Texture based on variance alignment, which utilizes a pre-trained diffusion model to achieve high-fidelity, view-consistent 3D texture synthesis, and further designs mutation alignment and pixel-level repair strategies;
VeCLIP: Improving CLIP Training via Visual-enriched Captions
Zhengfeng Lai (University of California, Davis), Meng Cao (Apple AI/ML)
CodeClassificationRetrievalTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Built an expandable visual rich re-annotation (VeCap) pipeline, which extracts visual concepts from AltText crawled from the web using a multi-modal LLM, then fuses the original AltText with these visual concepts through an open-source LLM to generate new 'visual rich captions' (VeCap), followed by a hybrid training scheme VeCLIP pre-trained on CLIP.
🎯 What it does: Propose a more general incremental learning scenario VIL (which can be class-incremental, domain-incremental, or both simultaneously), and design the ICON framework (comprising CAST regularization and IC dynamic classifier expansion) to address internal class-domain confusion and cross-domain class confusion in class-domain hybrid incremental learning.
🎯 What it does: Propose ViC-MAE, a self-supervised video and image representation learning method that combines Masked Autoencoder with contrastive learning.
🎯 What it does: This paper proposes VideoMamba, an efficient video recognition framework based on the Mamba pure state space model, which processes spatiotemporal information in videos by implementing forward and backward scanning along the temporal dimension and combining spatial-temporal bidirectional SSM.
VideoMamba: State Space Model for Efficient Video Understanding
Kunchang Li (Shenzhen Institute of Advanced Technology Chinese Academy of Sciences), Yu Qiao (OpenGVLab Shanghai AI Laboratory State Key Laboratory for Novel Software Technology Nanjing University)
🎯 What it does: Proposed a fully state-space model-based video understanding framework called VideoMamba, which simultaneously handles short-term action recognition and long-term video reasoning.
VideoStudio: Generating Consistent-Content and Multi-Scene Videos
Fuchen Long (HiDream.ai Inc), Tao Mei (HiDream.ai Inc)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Designed and implemented the VideoStudio framework, which generates multi-scenario scripts using a large language model, generates entity reference images, and produces visually consistent multi-scenario videos through an improved diffusion model.
Viewpoint textual inversion: discovering scene representations and 3D view control in 2D diffusion models
James Burgess (Stanford University), Serena Yeung-Levy (Stanford University)
CodeGenerationRepresentation LearningPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelImageText
🎯 What it does: By learning a '3D perspective token' in the embedding space of Stable Diffusion, the camera perspective of generated images can be controlled without modifying the model weights, and this method is used to examine whether the model internally contains an implicit 3D scene representation.
ViGoR: Improving Visual Grounding of Large Vision Language Models with Fine-Grained Reward Modeling
Siming Yan (University of Texas at Austin), Li Erran Li (AWS AI)
CodeObject DetectionReinforcement Learning from Human FeedbackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the ViGoR framework, which efficiently fine-tunes large vision-language models (e.g., LLaVA) for visual grounding through fine-grained reward modeling and automated evaluation methods, significantly reducing hallucinations and information omissions in text.
ViLA: Efficient Video-Language Alignment for Video Question Answering
Xijun Wang (University of Maryland College Park), Shan Yang (Amazon)
CodeKnowledge DistillationTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Designed and implemented the ViLA network, which includes a text-guided Frame-Prompter and a cross-modal QFormer-Distiller, to efficiently sample key frames from videos and align them with pre-trained LLMs, thereby completing video question answering tasks.
VISA: Reasoning Video Object Segmentation via Large Language Model
Cilin Yan (Beihang University University Of Amsterdam), Efstratios Gavves (Shanghai Jiao Tong University)
CodeSegmentationTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodality
🎯 What it does: The study proposes the ReasonVOS task and designs the VISA model, achieving video object segmentation and reasoning based on large language models.
🎯 What it does: Propose the VISAGE method, which generates appearance queries through mask pooling in video instance segmentation, improves query matching via contrastive learning, reduces dependence on positional information, and implements a simplified tracker and memory pool to enhance instance tracking accuracy.
🎯 What it does: Propose a self-reconstruction based tiny object detection framework SR-TOD, which enhances the visibility of tiny objects by adding a reconstruction head in the detector's neck layer to generate difference maps, thereby guiding feature enhancement.
Vision-Language Dual-Pattern Matching for Out-of-Distribution Detection
Zihan Zhang (Huazhong University of Science and Technology), Xiang Xiang (Huazhong University of Science and Technology)
CodeAnomaly DetectionTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a dual-modal matching framework based on CLIP (Dual-Pattern Matching, DPM), which constructs visual patterns and text patterns by simultaneously utilizing visual features and text features of ID (in-distribution) samples to enhance the accuracy of OOD (out-of-distribution) detection. Two implementations are provided: no-training (DPM-F) and lightweight training (DPM-T).
CodeClassificationSegmentationGenerationTransformerLarge Language ModelDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: Propose VisionLLaMA—a unified vision transformer architecture based on LLaMA, applicable to multiple visual tasks such as image generation, classification, segmentation, and detection.
🎯 What it does: Propose the VISTA3D framework, which employs a two-stage coarse-to-fine grid generation approach (first rapidly constructing a rough geometry using Gaussian Splatting, then converting it into an SDF and refining it via FlexiCubes), while achieving texture decoupling and angular composition from multi-source diffusion models, generating diverse and consistent high-quality 3D objects from a single image within 5–20 minutes.
Visual Grounding for Object-Level Generalization in Reinforcement Learning
Haobin Jiang (Peking University), Zongqing Lu (Peking University)
CodeReinforcement LearningVision Language ModelVideo
🎯 What it does: This study proposes the COPL method, which leverages the vision-language knowledge of MineCLIP for visual localization and reward design, thereby achieving zero-shot generalization for unseen targets in Minecraft tasks.
VividDreamer: Invariant Score Distillation for Hyper-Realistic Text-to-3D Generation
Wenjie Zhuo (State Key Laboratory of Brain-machine Intelligence, Zhejiang University), Yi Yang (State Key Laboratory of Brain-machine Intelligence, Zhejiang University)
🎯 What it does: Proposes the Invariant Score Distillation (ISD) method to address the issues of over-saturation and over-smoothing in Score Distillation Sampling (SDS) for text-to-3D generation.
🎯 What it does: Propose VLAD-BuFF, improving traditional VLAD aggregation by suppressing feature burstiness through self-similarity soft counting weights, and accelerating aggregation using PCA-initialized low-dimensional projection.
🎯 What it does: Proposes a self-supervised multi-object tracking model called Walker, which can perform detection and tracking using only sparse bounding box annotations without requiring instance ID labels.
🎯 What it does: This paper proposes a target-aware representation learning framework for low-light environments, leveraging bidirectional domain alignment and target saliency mechanisms to enhance the performance of high-level visual tasks.
🎯 What it does: Proposed a wavelet transform-based convolutional layer (WTConv) that achieves an extremely large receptive field without increasing the number of parameters, and can replace traditional depth convolutions in convolutional neural networks.
🎯 What it does: Propose a task-agnostic backdoor attack during training, named WBP, which induces binary bit flips in pre-trained model weights using hardware row-hammer, enabling backdoor implantation during fine-tuning.
🎯 What it does: This paper proposes the WebRPG task, which involves automatically generating rendering parameters for each element based on HTML code to achieve web page visualization without CSS code.
🎯 What it does: Proposes the WeConvene framework, integrating discrete wavelet transform into convolutional layers and entropy models to achieve more efficient learned image compression.
🎯 What it does: This paper proposes a continuous learning method that utilizes model averaging of pre-trained weights to balance the adaptability of new tasks and the memorization of old tasks, introducing CoMA and its improved version CoFiMA.
🎯 What it does: This paper proposes a method for weighting pseudo-labels in semi-supervised semantic segmentation. By jointly training a semantic segmentation model and an object detection model, reliable pseudo-label pixels are identified. Learning weights are assigned to each pseudo-label pixel using the ranking similarity of high-activation feature dimensions, thereby mitigating the negative impact of pseudo-label noise on training.
🎯 What it does: Propose and verify the 'Scaling on Scales' (S²) method, which enhances visual model performance by extracting features from multi-scale images while keeping the model size unchanged.
🎯 What it does: Proposes an efficient image restoration framework named SFHformer, which integrates Fast Fourier Transform (FFT) with Transformer, achieving the fusion of local and global features in a dual perception structure across spatial and frequency domains, capable of uniformly handling various image degradation problems.
🎯 What it does: Developed a general multi-modal pedestrian detection model called MMPedestron, and constructed a large-scale benchmark MMPD containing multiple modalities such as RGB, IR, Depth, LiDAR, and Event.
Which Model Generated This Image? A Model-Agnostic Approach for Origin Attribution
Fengyuan Liu (University of Oxford), Jindong Gu (University of Oxford)
CodeClassificationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: Propose a few-shot single-class classification framework called OCC-CLIP, which determines whether an image originates from the same generative model, thereby achieving source attribution of generated images.
WildRefer: 3D Object Localization in Large-scale Dynamic Scenes with Multi-modal Visual Data and Natural Language
Zhenxiang Lin (ShanghaiTech University), Yuexin Ma (Shanghai AI Laboratory)
CodeObject DetectionConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodalityPoint Cloud
🎯 What it does: This paper proposes the 3D Visual Grounding in the Wild (3DVGW) task, which utilizes synchronized LiDAR point clouds and industrial camera images along with natural language descriptions to locate target objects in large-scale dynamic scenes, and constructs two large-scale datasets, STRefer and LifeRefer, based on this task;
🎯 What it does: This paper proposes WiMANS, a novel dual-band WiFi CSI and synchronized video combined multi-user activity sensing benchmark dataset, and conducts benchmark experiments on multi-user identification, localization, and activity recognition tasks using this dataset.
WoVoGen: World Volume-aware Diffusion for Controllable Multi-camera Driving Scene Generation
Jiachen Lu (Fudan University), Li Zhang (Fudan University)
CodeAutonomous DrivingTransformerVision Language ModelDiffusion modelAuto EncoderVideoTextMultimodality
🎯 What it does: Propose the WoVoGen framework, which first predicts future world states using a 4D world volume (time + space + height), and then generates multi-camera street-level videos from this volume, supporting controllable edits for weather, lighting, urban style, etc.
WPS-SAM: Towards Weakly-Supervised Part Segmentation with Foundation Models
Xin-Jian Wu (State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences), Cheng-Lin Liu (State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences)
🎯 What it does: Propose a weakly supervised part segmentation framework called WPS-SAM, which leverages the pre-trained Segment Anything Model (SAM) to automatically generate prompts and complete pixel-level part segmentation during training using only bounding box or point-level annotations.
WSI-VQA: Interpreting Whole Slide Images by Generative Visual Question Answering
Pingyi Chen (Zhejiang University), Lin Yang (Westlake University)
CodeData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelMultimodalityBiomedical Data
🎯 What it does: Proposed the WSI-VQA framework, transforming various whole slide image (WSI)-level tasks into generative visual question answering (VQA) tasks, and constructed the Wsi2Text Transformer (W2T) model; simultaneously automatically generated and made publicly available the WSI-VQA dataset containing 977 WSI and 8,672 question-answer pairs.
XPSR: Cross-modal Priors for Diffusion-based Image Super-Resolution
Qu Yunpeng, Chao Zhou (Kuaishou Technology)
CodeSuper ResolutionVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: This paper proposes the XPSR framework, which utilizes cross-modal priors (high-level semantics and low-level degradation information) to drive diffusion models for image super-resolution, generating high-fidelity and realistic high-resolution images.
🎯 What it does: This paper proposes Programmable Gradient Information (PGI) and Generalized Efficient Layer Aggregation Network (GELAN), integrating them into YOLOv9, aiming to significantly enhance the training effectiveness and inference performance of lightweight and large-scale object detection models by reprogramming gradient transmission through reversible branches and multi-level auxiliary information;
🎯 What it does: This paper proposes the HQNet framework, which achieves single-stage multi-task human-centric perception (detection, segmentation, pose, attribute recognition, and 3D mesh restoration) by learning a unified human query (Human Query) to share cross-task features;
Zero-Shot Adaptation for Approximate Posterior Sampling of Diffusion Models in Inverse Problems
Yasar U Alcalar, Mehmet Akcakaya
CodeDiffusion modelScore-based ModelImagePhysics Related
🎯 What it does: Proposes the Zero-shot Approximate Posterior Sampling (ZAPS) method for efficient posterior sampling in inverse problems using diffusion models.
Huilin Zhu (Wuhan University of Technology), Shengfeng He (Singapore Management University)
CodeObject DetectionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Proposed the VA-Count framework, combining an example enhancement module (EEM) and a noise suppression module (NSM) to achieve zero-shot object counting.
ZeroI2V: Zero-Cost Adaptation of Pre-Trained Transformers from Image to Video
Xinhao Li (State Key Laboratory for Novel Software Technology, Nanjing University), Limin Wang (State Key Laboratory for Novel Software Technology, Nanjing University)
CodeRecognitionDomain AdaptationTransformerVideo
🎯 What it does: Propose a zero-cost method for migrating a pre-trained image Transformer to video tasks, named ZeroI2V, which utilizes spatiotemporal dual-head attention (STDHA) to achieve spatiotemporal modeling, employs a linear adapter for parameter-efficient transfer, and integrates it into the original model during inference through structural reparameterization, achieving no additional parameters or computations during inference.