CVPR 2025 Papers — Page 29
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 2871 papers
Visual Representation Learning through Causal Intervention for Controllable Image Editing
Shanshan Huang (Chongqing University), Li Liu (Chongqing University)
GenerationRepresentation LearningDiffusion modelImage
🎯 What it does: This paper proposes a causal intervention-based diffusion model (CIDiffuser) for learning interpretable causal visual representations and achieving controllable image editing.
Visual-Instructed Degradation Diffusion for All-in-One Image Restoration
Wenyang Luo (Chinese Academy of Sciences), Weiming Hu (Chinese Academy of Sciences)
RestorationDiffusion modelScore-based ModelAuto EncoderImage
🎯 What it does: This paper proposes a global image restoration framework named Defusion, which utilizes visual instructions to guide a degradation diffusion model to directly remove various image distortions in the degradation space, achieving integrated tasks such as denoising, deblurring, and defogging.
VITED: Video Temporal Evidence Distillation
Yujie Lu (University of California Santa Barbara), Tushar Nagarajan
Knowledge DistillationTransformerLarge Language ModelVision Language ModelVideoChain-of-Thought
🎯 What it does: This paper proposes a VideoQA model called VITED, which is based on video evidence chain distillation and can automatically generate and locate multi-step temporal evidence chains to answer questions in long videos.
ViUniT: Visual Unit Tests for More Robust Visual Programming
Artemis Panagopoulou (Salesforce AI Research), Juan Carlos Niebles (University of Pennsylvania)
GenerationData SynthesisOptimizationTransformerLarge Language ModelReinforcement LearningDiffusion modelImageTextMultimodality
🎯 What it does: The VuniT framework is proposed to enhance the reliability and correctness of visual programs by automatically generating visual unit tests.
VL-RewardBench: A Challenging Benchmark for Vision-Language Generative Reward Models
Lei Li (Hong Kong University), Qi Liu (Peking University)
GenerationRecommendation SystemTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmark
🎯 What it does: A benchmark for evaluating the Visual Language Generation Reward Model (VL-GenRM), called VL-RewardBench, has been constructed, which includes 1,250 manually verified multimodal preference pairs covering general multimodal requests, visual illusion detection, and complex reasoning tasks.
VL2Lite: Task-Specific Knowledge Distillation from Large Vision-Language Models to Lightweight Networks
Jinseong Jang (SK Telecom), Byeongwon Lee (SK Telecom)
ClassificationKnowledge DistillationConvolutional Neural NetworkVision Language ModelImageMultimodality
🎯 What it does: A single-stage knowledge distillation framework named VL2Lite is proposed, which directly transfers the knowledge of pre-trained vision-language models to lightweight networks to enhance classification performance.
VladVA: Discriminative Fine-tuning of LVLMs
Yassine Ouali (Samsung AI Cambridge), Georgios Tzimiropoulos (Queen Mary University of London)
RetrievalTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A Fine-tuning framework called VladVA is proposed to transform large visual-language models into discriminative models.
VLMs-Guided Representation Distillation for Efficient Vision-Based Reinforcement Learning
Haoran Xu (Shenzhen University), Yonghong Tian (Peking University)
Autonomous DrivingKnowledge DistillationTransformerReinforcement LearningPrompt EngineeringVision Language ModelImage
🎯 What it does: By transforming the common sense reasoning and referential ability of large visual language models (VLM) into supervisory signals, and combining self-supervised learning (SSL) to distill knowledge into the visual encoder of visual reinforcement learning (VRL) agents, the policy is trained using a soft actor-critic (SAC) after distillation.
VLog: Video-Language Models by Generative Retrieval of Narration Vocabulary
Kevin Qinghong Lin (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)
RetrievalTransformerLarge Language ModelVision Language ModelVideoTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes VLog, a video understanding framework based on 'narrative vocabulary' for retrieval generation;
VLOGGER: Multimodal Diffusion for Embodied Avatar Synthesis
Enric Corona (Google DeepMind), Cristian Sminchisescu (Google DeepMind)
GenerationData SynthesisPose EstimationTransformerDiffusion modelVideoMultimodalityAudio
🎯 What it does: VLOGGER achieves high-quality, variable-length videos that synthesize full-body movements, including head expressions, blinking, lip movements, as well as upper body and hand actions, from a single portrait and audio by combining audio-driven 3D pose and expression generation with a spatiotemporal diffusion model.
VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models
Byung-Kwan Lee (NVIDIA), Yueh-Hua Wu (NVIDIA)
Computational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Designed an efficient visual language model family VLsI with scales of 2B and 7B, achieving performance breakthroughs for small models through hierarchical language-based knowledge distillation.
VoCo-LLaMA: Towards Vision Compression with Large Language Models
Xubing Ye (Tsinghua University), Yansong Tang (Tsinghua University)
CompressionKnowledge DistillationTransformerLarge Language ModelImageVideo
🎯 What it does: Proposes VoCo-LLaMA, which achieves visual compression without external modules by learning to compress visual tokens within the LLM.
VODiff: Controlling Object Visibility Order in Text-to-Image Generation
Dong Liang (Tongji University), Rynson W. H. Lau (City University of Hong Kong)
GenerationData SynthesisDiffusion modelImageTextBenchmark
🎯 What it does: A training-free VODiff framework is proposed to explicitly control visibility order in text-to-image (T2I) generation, achieving precise occlusion relationships.
VolFormer: Explore More Comprehensive Cube Interaction for Hyperspectral Image Restoration and Beyond
Dabing Yu (Hohai University), Zheng Gao (Hohai University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: A Transformer structure named VolFormer is proposed for the super-resolution and denoising recovery tasks of single high-resolution hyperspectral images.
Volume Tells: Dual Cycle-Consistent Diffusion for 3D Fluorescence Microscopy De-noising and Super-Resolution
Zelin Li (City University of Hong Kong), Hong Yan (City University of Hong Kong)
RestorationSuper ResolutionDiffusion modelImage
🎯 What it does: In response to the spatial variation noise and uneven axial resolution issues in long-term live cell imaging using 3D fluorescence microscopy, this paper proposes an unsupervised learning framework that can simultaneously achieve denoising and super-resolution.
Volumetric Surfaces: Representing Fuzzy Geometries with Layered Meshes
Stefano Esposito (University of Tuebingen), Andreas Geiger (University of Tuebingen)
GenerationComputational EfficiencyMesh
🎯 What it does: A multi-layer SDF layered representation (k-SDF) is designed to achieve real-time rendering with a small number of sampling points by transforming fuzzy geometry into several translucent mesh layers.
Volumetrically Consistent 3D Gaussian Rasterization
Chinmay Talegaonkar (University of California San Diego), Nicholas Antipa (University of California San Diego)
GenerationData SynthesisComputational EfficiencyGaussian SplattingImagePoint CloudComputed Tomography
🎯 What it does: This work improves the alpha computation method of the 3D Gaussian Splatting (3DGS) efficient rasterization framework while maintaining its efficiency, by performing analytical integration of Gaussian density directly in 3D space, resulting in more physically accurate transmittance that can be directly used for color synthesis.
VoteFlow: Enforcing Local Rigidity in Self-Supervised Scene Flow
Yancong Lin (Delft University of Technology), Holger Caesar (Delft University of Technology)
Autonomous DrivingOptical FlowPoint Cloud
🎯 What it does: A self-supervised scene flow estimation method called VoteFlow is proposed, which captures local rigid motion through a voting module and embeds it into the network structure.
VoxelSplat: Dynamic Gaussian Splatting as an Effective Loss for Occupancy and Flow Prediction
Ziyue Zhu (Nankai University), Jian Yang (Nankai University)
SegmentationAutonomous DrivingGaussian SplattingOptical FlowPoint Cloud
🎯 What it does: VoxelSplat is proposed, a framework that uses 4D Gaussian Splatting as additional supervision during the training phase to simultaneously predict camera-referenced semantic occupancy maps and scene flow.
VSNet: Focusing on the Linguistic Characteristics of Sign Language
Yuhao Li (University of Electronic Science and Technology of China), Yazhou Ren (University of Electronic Science and Technology of China)
RecognitionPose EstimationGraph Neural NetworkVideo
🎯 What it does: For the word-level sign language recognition task, VSNet is proposed, which first obtains a simplified skeleton through adaptive skeleton simplification (weak joint dropping), then groups the skeleton into visual symbols (VS), and utilizes a self-attention model to capture the spatial-temporal relationships of VS, ultimately achieving sign language recognition.
VTON 360: High-Fidelity Virtual Try-On from Any Viewing Direction
Zijian He (Sun Yat-sen University), Guanbin Li (Sun Yat-sen University)
GenerationPose EstimationDiffusion modelGaussian SplattingImage
🎯 What it does: This paper proposes a 3D virtual try-on method called VTON 360 that achieves high-fidelity rendering from any viewpoint.
VTON-HandFit: Virtual Try-on for Arbitrary Hand Pose Guided by Hand Priors Embedding
Yujie Liang (Xiamen University), Rongrong Ji (Xiamen University)
Image TranslationData SynthesisPose EstimationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: To address the issue of hand occlusion in virtual try-on, we propose VTON-HandFit, which utilizes hand priors to achieve realistic clothing synthesis for arbitrary hand gestures.
Watermarking One for All: A Robust Watermarking Scheme Against Partial Image Theft
Gaozhi Liu (Fudan University), Wanli Peng (China Agricultural University)
RecognitionImage TranslationRestorationSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: A watermark embedding and extraction framework named WOFA is proposed, which can accurately extract watermarks even when only part of the original image is obtained and after geometric transformations and background blending.
Wav2Sem: Plug-and-Play Audio Semantic Decoupling for 3D Speech-Driven Facial Animation
Hao Li (Beihang University), Lei Li (University of Washington)
RecognitionGenerationTransformerSupervised Fine-TuningAudio
🎯 What it does: A plugin semantic decoupling module named Wav2Sem is proposed, which extracts semantic features from speech using global semantic information and integrates them with a self-supervised audio encoder to improve the coupling and averaging issues of near-homophones in 3D speech-driven facial animation.
WAVE: Weight Templates for Adaptive Initialization of Variable-sized Models
Fu Feng (Southeast University), Xin Geng (Southeast University)
OptimizationKnowledge DistillationTransformerImage
🎯 What it does: The WAVE method is proposed, utilizing shared weight templates and differentiable size-specific scalers to initialize visual Transformer models of different scales from a multi-task perspective.
Wavelet and Prototype Augmented Query-based Transformer for Pixel-level Surface Defect Detection
Feng Yan (Zhengzhou University), Mingliang Xu (Zhengzhou University)
SegmentationAnomaly DetectionTransformerImage
🎯 What it does: A query-based Transformer enhanced with Wavelet and Prototype is proposed for pixel-level surface defect detection, utilizing the interaction between frequency domain and spatial domain to improve defect detail segmentation accuracy.
Weakly Supervised Contrastive Adversarial Training for Learning Robust Features from Semi-supervised Data
Lilin Zhang (Sichuan University), Ning Yang (Sichuan University)
ClassificationRepresentation LearningAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper proposes Weakly Supervised Contrastive Adversarial Training (WSCAT), which generates complete perturbation adversarial samples on semi-supervised data to block the correlation between non-robust features and labels, thereby learning more robust features; it also provides theoretical analysis and experimental validation.
Weakly Supervised Semantic Segmentation via Progressive Confidence Region Expansion
Xiangfeng Xu (East China Normal University), Shaohui Lin (East China Normal University)
SegmentationTransformerImage
🎯 What it does: A weakly supervised semantic segmentation framework based on Progressive Confidence Region Expansion (PCRE) is proposed to address the issue of over-expansion in CAM generated by ViT.
Weakly Supervised Temporal Action Localization via Dual-Prior Collaborative Learning Guided by Multimodal Large Language Models
Quan Zhang (Tsinghua University), Chun Yuan (Tsinghua University)
RecognitionObject DetectionKnowledge DistillationTransformerLarge Language ModelVideoTextMultimodality
🎯 What it does: The MLLM4WTAL framework is proposed, utilizing the key semantic and complete semantic priors provided by the multimodal large language model (MLLM) to guide the weakly supervised temporal action localization (WTAL) model through two modules: Key Semantic Matching (KSM) and Complete Semantic Reconstruction (CSR).
WeakMCN: Multi-task Collaborative Network for Weakly Supervised Referring Expression Comprehension and Segmentation
Silin Cheng (Huazhong Agricultural University), Gen Luo (OpenGVLab Shanghai AI Laboratory)
RecognitionSegmentationContrastive LearningImage
🎯 What it does: Proposes the WeakMCN multi-task collaborative network, achieving joint learning of weakly supervised reference expression comprehension (WREC) and segmentation (WRES).
WeatherGen: A Unified Diverse Weather Generator for LiDAR Point Clouds via Spider Mamba Diffusion
Yang Wu (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Object DetectionGenerationData SynthesisAutonomous DrivingDiffusion modelContrastive LearningPoint Cloud
🎯 What it does: This paper presents WeatherGen, a unified multi-weather LiDAR data generation framework that utilizes a learning-based simulator to generate high-quality multi-weather point clouds. It introduces the Spider Mamba generator, aligner, and contrastive learning controller in a diffusion network to achieve high-fidelity and controllable LiDAR generation.
WeGen: A Unified Model for Interactive Multimodal Generation as We Chat
Zhipeng Huang (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
GenerationTransformerLarge Language ModelPrompt EngineeringDiffusion modelVideoMultimodality
🎯 What it does: We propose WeGen, a unified multimodal generation model that supports various visual generation tasks through natural dialogue.
WF-VAE: Enhancing Video VAE by Wavelet-Driven Energy Flow for Latent Video Diffusion Model
Zongjian Li (Peking University), Li Yuan (Peng Cheng Laboratory)
GenerationCompressionComputational EfficiencyDiffusion modelAuto EncoderVideo
🎯 What it does: A WF-VAE based on multi-layer wavelet transform is designed, utilizing low-frequency energy flow to compress video information into latent space, and a Causal Cache is proposed to achieve lossless block-level inference.
What Makes a Good Dataset for Knowledge Distillation?
Logan Frank (Ohio State University), Jim Davis (Ohio State University)
Knowledge DistillationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: The study investigates which alternative data can effectively transfer teacher knowledge during knowledge distillation when the original training data is unavailable, and proposes criteria for measuring the quality of distillation data.
What's in the Image? A Deep-Dive into the Vision of Vision Language Models
Omri Kaduri (Weizmann Institute of Science), Tali Dekel (Weizmann Institute of Science)
RetrievalCompressionTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: An empirical analysis of the internal visual information processing mechanisms of large visual language models reveals that query words compress global visual information, while intermediate layers are crucial for local retrieval of key and fine-grained information.
When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach
Vaibhav Rathore (Indian Institute of Technology Bombay), Biplab Banerjee (Indian Institute of Technology Bombay)
Domain AdaptationTransformerContrastive LearningImage
🎯 What it does: A DG-GCD setting is proposed that can achieve cross-domain general category discovery using only source domain data, and DG-2 CD-Net is designed to accomplish this task.
When the Future Becomes the Past: Taming Temporal Correspondence for Self-supervised Video Representation Learning
Yang Liu (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
SegmentationKnowledge DistillationRepresentation LearningTransformerContrastive LearningVideo
🎯 What it does: A self-supervised video representation learning framework T-CoRe is proposed, which guides the recovery of masked video models through temporal correspondence, significantly improving the performance of video downstream tasks.
Where the Devil Hides: Deepfake Detectors Can No Longer Be Trusted
Shuaiwei Yuan (Ocean University of China), Yuezun Li (Ocean University of China)
RecognitionAdversarial AttackGenerative Adversarial NetworkImageVideo
🎯 What it does: This paper implements a backdoor attack on deepfake detectors by embedding triggers in the training data;
Where's the Liability in the Generative Era? Recovery-based Black-Box Detection of AI-Generated Content
Haoyue Bai (University of Wisconsin-Madison), Haifeng Chen (NEC Laboratories America)
GenerationAnomaly DetectionDiffusion modelImage
🎯 What it does: A black-box AI-generated image detection framework that can be implemented solely through API access is proposed, which determines whether an image is generated by the target model based on the differences after masking and recovery.
Which Viewpoint Shows it Best? Language for Weakly Supervising View Selection in Multi-view Instructional Videos
Sagnik Majumder (University of Texas at Austin), Kristen Grauman (University of Texas at Austin)
GenerationPose EstimationTransformerVision Language ModelVideo
🎯 What it does: In multi-view teaching videos, a method for weakly supervised optimal viewpoint selection using language descriptions is proposed.
WildAvatar: Learning In-the-wild 3D Avatars from the Web
Zihao Huang (Huazhong University of Science and Technology), Ziwei Liu (Nanyang Technological University)
Object DetectionSegmentationGenerationPose EstimationVideo
🎯 What it does: Proposes an automated labeling and screening process without human intervention, and based on this, collects a dataset of over 10k human subjects' wild videos from YouTube for 3D avatar generation;
WildGS-SLAM: Monocular Gaussian Splatting SLAM in Dynamic Environments
Jianhao Zheng (Stanford University), Iro Armeni (Stanford University)
Object DetectionPose EstimationDepth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: A monocular RGB SLAM system called WildGS-SLAM is proposed, which can automatically eliminate dynamic objects through uncertainty prediction and construct static scene maps in dynamic environments based on 3D Gaussian splatting.
WiLoR: End-to-end 3D Hand Localization and Reconstruction in-the-wild
Rolandos Alexandros Potamias (Imperial College London), Stefanos Zafeiriou (Imperial College London)
Object DetectionPose EstimationTransformerImage
🎯 What it does: An end-to-end hand detection and 3D reconstruction framework called WiLoR is proposed, achieving real-time multi-hand localization and high-precision 3D hand model reconstruction in complex environments.
WISE: A Framework for Gigapixel Whole-Slide-Image Lossless Compression
Yu Mao (Mohamed bin Zayed University of Artificial Intelligence), Chun Jason Xue (Mohamed bin Zayed University of Artificial Intelligence)
CompressionImage
🎯 What it does: This paper proposes a lossless compression framework WISE for whole slide images (WSI), aiming to reduce the storage and transmission costs of extremely large images.
WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels
Hyeokjun Kweon (Chung-Ang University), Kuk-Jin Yoon (KAIST)
Object DetectionSegmentationPrompt EngineeringImage
🎯 What it does: A framework for instance segmentation called WISH is proposed, which can simultaneously utilize multiple types of weak labels and supports heterogeneous weak annotations.
WISNet: Pseudo Label Generation on Unbalanced and Patch Annotated Waste Images
Shifan Zhang (Shanghai Jiao Tong University), Shan Chang (Donghua University)
SegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A dataset called ShanghaiWaste was constructed, containing 12,208 waste images and 40,392 patch-level annotations, and a weakly supervised pseudo-label generation framework WISNet was proposed for achieving pixel-level semantic segmentation of waste images.
Wonderland: Navigating 3D Scenes from a Single Image
Hanwen Liang (University of Toronto), Jian Ren (Snap Inc.)
GenerationData SynthesisTransformerDiffusion modelGaussian SplattingImageVideo
🎯 What it does: This paper proposes a framework named WonderLan, which can quickly generate high-quality, wide-angle 3D scenes from a single image.
WonderWorld: Interactive 3D Scene Generation from a Single Image
Hong-Xing Yu (Stanford University), Jiajun Wu (MIT)
GenerationData SynthesisDiffusion modelGaussian SplattingImage
🎯 What it does: An interactive 3D scene generation framework named WonderWorld is proposed, which allows users to generate connected and diverse 3D scenes in real-time and with low latency based on a single image.
Words or Vision: Do Vision-Language Models Have Blind Faith in Text?
Ailin Deng (National University of Singapore), Bryan Hooi (National University of Singapore)
TransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This study investigates the modal preference of visual-language models (VLM) when there is inconsistency between visual and textual information. It constructs a benchmark with four types of tasks, including matching, corruption, and irrelevant text, to evaluate 10 VLMs. The findings reveal a phenomenon of 'blind trust in text' and explore the impacts of factors such as prompts, model size, text relevance, token order, and unimodal confidence. Subsequently, it proposes to mitigate this bias through supervised fine-tuning combined with text augmentation.
World-consistent Video Diffusion with Explicit 3D Modeling
Qihang Zhang (Apple), Jiatao Gu (The Chinese University of Hong Kong)
GenerationData SynthesisDepth EstimationTransformerDiffusion modelImageVideoPoint Cloud
🎯 What it does: Proposes the WVD (World-consistent Video Diffusion) model, which learns both RGB and XYZ images within a diffusion framework, achieving 3D consistent video and image generation, supporting multi-tasks such as single image to 3D, depth estimation, and camera controllable video.
X-Dyna: Expressive Dynamic Human Image Animation
Di Chang (University of Southern California), Mohammad Soleymani
GenerationData SynthesisPose EstimationDiffusion modelVideo
🎯 What it does: X-Dyna is a zero-shot diffusion model pipeline that can generate realistic, dynamic animations for a single portrait in real-time by utilizing the poses and expressions driven from videos.
XLRS-Bench: Could Your Multimodal LLMs Understand Extremely Large Ultra-High-Resolution Remote Sensing Imagery?
Fengxiang Wang (National University of Defense Technology), Maosong Sun (Tsinghua University)
RecognitionObject DetectionSegmentationGenerationTransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes XLRS-Bench, a benchmark dataset specifically designed to evaluate the perception and reasoning capabilities of multimodal large language models in ultra-high-resolution remote sensing images.
Yo'Chameleon: Personalized Vision and Language Generation
Thao Nguyen (University of Wisconsin Madison), Yuheng Li (Adobe Research)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Personalizing large multimodal models using 3-5 images and soft prompts, enabling them to possess exclusive capabilities for specific concepts in understanding and generation.
You See it, You Got it: Learning 3D Creation on Pose-Free Videos at Scale
Baorui Ma (Beijing Academy of Artificial Intelligence), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationDepth EstimationDiffusion modelGaussian SplattingOptical FlowVideo
🎯 What it does: A visual conditional multi-view diffusion model See3D has been constructed, which can be trained on network videos without pose annotations, and achieves high-fidelity reconstruction from single or sparse views to 3D scenes using a distortion-based generative framework.
Your Large Vision-Language Model Only Needs A Few Attention Heads For Visual Grounding
Seil Kang (Yonsei University), Seong Jae Hwang (Yonsei University)
Object DetectionSegmentationTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper analyzes the attention mechanism of large visual language models (LVLM) and finds that only a few attention heads are needed to achieve zero-shot visual grounding (localization and segmentation).
Your Scale Factors are My Weapon: Targeted Bit-Flip Attacks on Vision Transformers via Scale Factor Manipulation
Jialai Wang (National University of Singapore), Zhenkai Liang (Tsinghua University)
Adversarial AttackTransformerImage
🎯 What it does: This paper proposes a target bit-flip attack method Flip-S aimed at quantized Transformers (ViT), which utilizes the flippable bits of the scale factor in quantized models to implant backdoors and induce the model to output a specified category for target inputs.
Your ViT is Secretly an Image Segmentation Model
Tommie Kerssies (Eindhoven University of Technology), Daan de Geus (RWTH Aachen University)
SegmentationTransformerLarge Language ModelImage
🎯 What it does: Proposes the Encoder-only Mask Transformer (EoMT), which uses only ViT for image segmentation, removing task-specific modules such as ViT-Adapter, pixel decoder, multi-scale, and Transformer decoder.
Z-Magic: Zero-shot Multiple Attributes Guided Image Creator
Yingying Deng (Chinese Academy of Sciences), Weiming Dong (Chinese Academy of Sciences)
GenerationData SynthesisDiffusion modelScore-based ModelImage
🎯 What it does: This paper studies multi-attribute zero-shot image generation and proposes the Z-Magic method, which explicitly models the conditional dependencies between attributes to make the generated images more consistent.
Zero-1-to-A: Zero-Shot One Image to Animatable Head Avatars Using Video Diffusion
Zhenglin Zhou (Zhejiang University), Tat-Seng Chua (National University of Singapore)
GenerationData SynthesisDiffusion modelScore-based ModelGaussian SplattingImageVideo
🎯 What it does: This paper proposes a method for generating animatable 4D avatars from a single portrait image in a zero-shot manner.
Zero-shot 3D Question Answering via Voxel-based Dynamic Token Compression
Hsiang-Wei Huang (University of Washington), Cheng-Hao Kuo (Amazon)
Object DetectionCompressionTransformerVision Language ModelImagePoint Cloud
🎯 What it does: For multi-view 3D scenes, a voxel-based Dynamic Token Compression (DTC) method is proposed to significantly reduce the number of visual tokens while maintaining 3D question-answering performance.
Zero-Shot 4D Lidar Panoptic Segmentation
Yushan Zhang (NVIDIA), Tim Meinhardt (NVIDIA)
Object DetectionObject TrackingSegmentationAutonomous DrivingTransformerVision Language ModelVideoPoint Cloud
🎯 What it does: The SAL-4D method is proposed, achieving zero-shot segmentation, tracking, and recognition in Lidar 4D sequences.
Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling
Yuhui Quan (South China University of Technology), Hui Ji (National University of Singapore)
RestorationGenerative Adversarial NetworkImage
🎯 What it does: A zero-shot blind spot denoising method addressing the noise correlation of adjacent pixels in the real world, utilizing implicit neural representation to resample visible pixels, thereby reducing noise correlation while preserving pixel correlation, and jointly training a denoising network for denoising.
Zero-Shot Head Swapping in Real-World Scenarios
Taewoong Kang (KAIST), Jaegul Choo (KAIST)
Image HarmonizationGenerationDiffusion modelImage
🎯 What it does: A zero-shot head swapping method called HID is proposed, which utilizes IOMask to automatically generate masks and combines a hair injection module to achieve seamless integration of the head and body.
Zero-Shot Image Restoration Using Few-Step Guidance of Consistency Models (and Beyond)
Tomer Garber (Bar Ilan University), Tom Tirer (Bar Ilan University)
RestorationSuper ResolutionDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a zero-shot image restoration framework based on a Consistency Model (CM4IR), capable of completing super-resolution, deblurring, and inpainting tasks with only 4 neural function evaluations (NFE).
Zero-Shot Monocular Scene Flow Estimation in the Wild
Yiqing Liang (NVIDIA), Orazio Gallo (NVIDIA)
Depth EstimationAutonomous DrivingTransformerOptical FlowVideo
🎯 What it does: A zero-shot monocular scene flow estimation framework is proposed, capable of directly predicting 3D point displacements and motion offsets in unseen outdoor videos.
Zero-Shot Novel View and Depth Synthesis with Multi-View Geometric Diffusion
Vitor Guizilini (Toyota Research Institute), Rares Ambrus (Toyota Research Institute)
GenerationData SynthesisDepth EstimationAutonomous DrivingTransformerDiffusion modelImage
🎯 What it does: A diffusion model architecture named MVGD is proposed, which can directly generate multi-view consistent and scale-distinguishable RGB images and depth maps from any number of pose-known images, without the need for intermediate 3D representations.
Zero-shot RGB-D Point Cloud Registration with Pre-trained Large Vision Model
Haobo Jiang (Nanyang Technological University), Jianmin Zheng (Nanjing University)
RecognitionPose EstimationDiffusion modelPoint Cloud
🎯 What it does: A zero-shot RGB-D point cloud registration framework called ZeroMatch is proposed, which can perform point cloud registration on unseen data without task-specific training.
Zero-Shot Styled Text Image Generation, but Make It Autoregressive
Vittorio Pippi (University of Modena and Reggio Emilia), Rita Cucchiara (University of Modena and Reggio Emilia)
GenerationData SynthesisTransformerAuto EncoderImageText
🎯 What it does: This paper proposes an offline, autoregressive text image generation model named Emuru, which can synthesize background-free, style-faithful handwritten or typed text images given any length of text and reference style samples.
ZeroGrasp: Zero-Shot Shape Reconstruction Enabled Robotic Grasping
Shun Iwase (Carnegie Mellon University), Sergey Zakharov (Toyota Research Institute)
Pose EstimationRobotic IntelligenceConvolutional Neural NetworkTransformerAuto EncoderImage
🎯 What it does: We propose ZeroGrasp, an end-to-end framework that can perform real-time 3D reconstruction and 6D grasp pose prediction from a single-view RGB-D image, and we release a large-scale synthetic dataset, ZeroGrasp-11B.
ZeroVO: Visual Odometry with Minimal Assumptions
Lei Lai (Boston University), Eshed Ohn-Bar (Boston University)
Depth EstimationAutonomous DrivingTransformerSimultaneous Localization and MappingOptical FlowVideoMultimodality
🎯 What it does: This paper proposes ZeroVO, an algorithm for visual odometry that does not require camera calibration and can achieve zero-shot transfer in various environments.
ZoomLDM: Latent Diffusion Model for Multi-scale Image Generation
Srikar Yellapragada (Stony Brook University), Dimitris Samaras (Stony Brook University)
GenerationData SynthesisSuper ResolutionTransformerDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: We propose ZoomLDM, a conditional latent diffusion model capable of generating high-quality medical images at multiple scales.