CVPR 2026 Papers — Page 41
IEEE/CVF Conference on Computer Vision and Pattern Recognition · 4071 papers
When LoRA Betrays: Backdooring Text-to-Image Models by Masquerading as Benign Adapters
Liangwei Lyu (People's Public Security University of China), Qiyao Deng (People's Public Security University of China)
GenerationAdversarial AttackTransformerDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Developed a framework named MasqLoRA for text-to-image model backdoor attacks using LoRA modules
When Numbers Speak: Aligning Textual Numerals and Visual Instances in Text-to-Video Diffusion Models
Zhengyang Sun (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
GenerationVision Language ModelDiffusion modelVideoTextMultimodality
🎯 What it does: Propose NUMINA, a training-free framework that guides video generation to precisely align with numerical instructions in text through dynamic selection, layout construction, refinement, and layout-based generation of attention heads in text-to-video diffusion models.
When Pretty Isn't Useful: Investigating Why Modern Text-to-Image Models Fail as Reliable Training Data Generators
Krzysztof Adamkiewicz (RPTU University Kaiserslautern-Landau), Andreas Dengel (RPTU University Kaiserslautern-Landau)
ClassificationData SynthesisDepth EstimationPrompt EngineeringDiffusion modelImage
🎯 What it does: Evaluated the transfer performance of synthetic data generated by 13 state-of-the-art text-to-image diffusion models when training classifiers on ImageNet subsets, and analyzed the root causes of performance degradation.
When Robots Obey the Patch: Universal Transferable Patch Attacks on Vision-Language-Action Models
Hui Lu (Nanyang Technological University), Xudong Jiang (Nanyang Technological University)
Adversarial AttackRobotic IntelligenceVision-Language-Action ModelContrastive LearningMultimodality
🎯 What it does: Proposes a generic transferable attack framework named UPA-RFAS for vision-language-action (VLA) models, and verifies its efficiency under black-box, cross-model, cross-task, and cross-domain settings.
When Robots Should Say ''I Don't Know'': Benchmarking Abstention in Embodied Question Answering
Tao Wu (Nanyang Technological University), Jianfei Yang (Nanyang Technological University)
Robotic IntelligenceLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelTextPoint CloudBenchmarkChain-of-Thought
🎯 What it does: Constructed a benchmark called AbstainEQA that enables robots to say 'I don't know' when uncertain, and systematically evaluated the performance of existing embodied QA models on this task.
When Safety Collides: Resolving Multi-Category Harmful Conflicts in Text-to-Image Diffusion via Adaptive Safety Guidance
Yongli Xiang (University of Sydney), Tongliang Liu (University of Sydney)
GenerationSafty and PrivacyDiffusion modelImageTextBenchmark
🎯 What it does: Proposed a training-agnostic conflict-aware adaptive safety guidance framework, CASG, to address multi-class harmful conflicts in text-to-image diffusion models.
When to Think and When to Look: Uncertainty-Guided Lookback
Jing Bi (University of Rochester), Chenliang Xu (University of Rochester)
Computational EfficiencyVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper systematically evaluates the 'thinking' (chain-of-thought) patterns of large-scale vision-language models during inference and proposes an uncertainty-guided 'lookback' decoding strategy that improves visual reasoning accuracy while maintaining or reducing generation length.
When Token Pruning is Worse than Random: Understanding Visual Token Information in VLLMs
Yahong Wang (Tongji University), Yuyin Zhou (University Of California Santa Cruz)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper introduces the concept of 'information horizon' to analyze and quantify information decay in visual large language models (VLLM), revealing that information in deeper visual tokens tends to become uniform and vanish, leading to existing training-agnostic pruning methods performing similarly to random pruning in deeper layers.
When Transformers Meet Mamba: A Hybrid Transformer-Mamba Network for Video Object Detection
Qiang Qi (Qingdao University of Science and Technology), Yu Zhang (Qingdao University of Science and Technology)
Object DetectionTransformerVideo
🎯 What it does: Proposed TMambaDet, a hybrid network integrating Transformer and Mamba for video object detection.
When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm
Ye Leng (CISPA Helmholtz Center for Information Security), Yang Zhang (Flexera)
GenerationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: The system compared the performance of diffusion models and multimodal large language models (MLLMs) on two major security risks: unsafe content generation and forged image detection. By generating 82,880 images from 1,184 unsafe prompts and calculating unsafe scores using the OpenAI Moderation API; then generating 14,000 images from 2,000 safe prompts (from MSCOCO and Flickr30k) to evaluate the accuracy of four public and commercial detectors.
When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought
Yiyang Zhou (ByteDance Seed), Qinghao Ye (ByteDance Seed)
TransformerLarge Language ModelVision Language ModelImageMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Proposes the MIRA benchmark to evaluate the ability of multimodal large language models to generate and utilize intermediate visual images for reasoning.
Where Culture Fades: Revealing the Cultural Gap in Text-to-Image Generation
Chuancheng Shi (University of Sydney), Tat-Seng Chua (National University of Singapore)
GenerationExplainability and InterpretabilityRepresentation LearningTransformerVision Language ModelAuto EncoderImageTextMultimodalityBenchmark
🎯 What it does: Analyze the phenomenon of multilingual text-to-image models generating culturally neutral or English-biased outputs, proposing to locate culture-related neurons via attention and sparse autoencoders (SAE), and designing two methods for culture activation with zero training or lightweight hierarchical fine-tuning.
Where Does Vision Meet Language? Understanding and Refining Visual Fusion in MLLMs via Contrastive Attention
Shezheng Song, Jie Yu (Hunan University)
Computational EfficiencyTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Systematically study the hierarchical process of visual and textual information fusion in multimodal large language models (MLLMs), and propose a training-free contrastive attention mechanism to enhance the quality of visual attention.
Where MLLMs Attend and What They Rely On: Explaining Autoregressive Token Generation
Ruoyu Chen (Chinese Academy of Sciences), Xiaochun Cao (Sun Yat-sen University)
Explainability and InterpretabilityComputational EfficiencyLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Propose a black-box attribution framework named EAGLE to explain the autoregressive token generation process in multimodal large language models (MLLMs), capable of locating visual regions that generated tokens depend on, quantifying the relative importance of linguistic priors and visual evidence, and diagnosing and reducing hallucinations.
Where, What, Why: Toward Explainable 3D-GS Watermarking
Mingshu Cai (Waseda University), Yixuan Li (Nanyang Technological University)
Safty and PrivacyExplainability and InterpretabilityGaussian SplattingPoint Cloud
🎯 What it does: Propose an interpretable 3D Gaussian Splatting (3D-GS) watermarking framework, which selects secure and reliable Gaussian primitives through Trio-Experts, ensures coexistence of visual quality and watermark robustness via SBAG gate allocation, Channel-wise Group Mask, and Decoupled Finetuning.
Which Concepts to Forget and How to Refuse? Decomposing Concepts for Continual Unlearning in Large Vision-Language Models
Hyundong Jin (Chung-Ang University), Eunwoo Kim (Chung-Ang University)
Safty and PrivacyLarge Language ModelMixture of ExpertsVision Language ModelContrastive LearningMultimodality
🎯 What it does: Propose a continuous forgetting framework CORE based on visual-language concept decomposition, which can precisely generate context-related refusal responses to user recursive deletion requests while maintaining the generality of large-scale visual-language models.
WhisperNet: A Scalable Solution for Bandwidth-Efficient Collaboration
Gong Chen (Tianjin University), Xinyan Zhao (Tianjin University)
Autonomous DrivingComputational EfficiencyMixture of ExpertsPoint Cloud
🎯 What it does: Proposes the WhisperNet framework, achieving a bandwidth-aware collaborative perception system based on receiver-side global coordination.
White-Balance First, Adjust Later: Cross-Camera Color Constancy via Vision-Language Evaluation
Shuwei Li (National University of Singapore), Robby T. Tan (National University of Singapore)
RestorationTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: By employing a vision-language model (VLM) for iterative white balance and illumination estimation, the color constancy problem is reformulated as an interactive perceptual feedback process;
WHU-MARS: A Multispectral Aerial-Ground Benchmark Towards Any-Scenario Person Re-Identification
Yuxuan Zhao (Wuhan University), Mang Ye (Wuhan University)
RetrievalTransformerContrastive LearningImageMultimodalityBenchmark
🎯 What it does: This paper proposes a novel Any-Scenario ReID (AS-ReID) task, constructs a large-scale multispectral aerial-ground human image dataset WHU-MARS, and designs a Unified Alignment and Discrimination (UAD) framework based on a single network to achieve unified retrieval across modalities and viewpoints.
Why Does RL Generalize Better Than SFT? A Data-Centric Perspective on VLM Post-Training
Aojun Lu (Sichuan University), Yanan Sun (Sichuan University)
Computational EfficiencyData-Centric LearningSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Investigated the fundamental reasons for the generalization gap between reinforcement learning (RL) and supervised fine-tuning (SFT) in vision-language models (VLMs), and proposed a Difficulty-Curated SFT (DC-SFT) method based on difficulty screening;
Why Not Hyperparameter-Friendly Optimisation? A Monotonic Adaptive Norm Rescaling Approach For Long-Tailed Recognition
Shuo Zhang (University of Oxford), Tingting Zhu (University of Oxford)
ClassificationRecognitionImage
🎯 What it does: This paper proposes an adaptive monotonic normalization (SAMN) method, which addresses the weight scale imbalance problem in long-tailed recognition by directly applying monotonic constraints on the weight norms of each class during two-stage retraining.
Widget2Code: From Visual Widgets to UI Code via Multimodal LLMs
Houston H. Zhang (McMaster University), Zhixiang Chi (University of Toronto)
AI Code AssistantTransformerLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes Widget2Code, which automatically generates visual-to-code for small, context-free application widgets, constructing the first image-only widget dataset and providing fine-grained evaluation metrics.
WikiCLIP: An Efficient Contrastive Baseline for Open-domain Visual Entity Recognition
Shan Ning (ShanghaiTech University), Xuming He (ShanghaiTech University)
RecognitionTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes WikiCLIP, an efficient open-domain visual entity recognition framework based on contrastive learning, aiming to address the problem of entity retrieval in large-scale knowledge bases.
WildCap: Facial Albedo Capture in the Wild via Hybrid Inverse Rendering
Yuxuan Han (Tsinghua University), Feng Xu (Tsinghua University)
RestorationConvolutional Neural NetworkDiffusion modelVideo
🎯 What it does: High-quality human face diffuse maps were reconstructed in wild environments by first preprocessing facial images from mobile phone videos using data-driven SwitchLight, and then combining model-driven inverse rendering (employing a texture grid lighting model and diffusion prior).
WildPose: A Unified Framework for Robust Pose Estimation in the Wild
Jianhao Zheng (Stanford University), Iro Armeni (Stanford University)
Pose EstimationConvolutional Neural NetworkRecurrent Neural NetworkSimultaneous Localization and MappingImageVideo
🎯 What it does: Proposed WildPose, a unified monocular pose estimation framework that maintains robustness in dynamic environments while achieving state-of-the-art performance in static environments.
WildRayZer: Self-supervised Large View Synthesis in Dynamic Environments
Xuweiyi Chen (University of Virginia), Zezhou Cheng (University of Virginia)
GenerationData SynthesisTransformerNeural Radiance FieldContrastive LearningGaussian SplattingImageVideo
🎯 What it does: Proposes a fully self-supervised WildRayZer framework that generates static new views of non-transient (dynamic) objects from sparse uncalibrated perspectives in dynamic environments where both the camera and objects are moving.
Will Multimodal Models Be Dazzled by Multi-Image Visual Puzzles?
Zhi Zhu (Nanjing University), Tong Lu (Nanjing University)
Large Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper proposes the MIRACLE benchmark to evaluate the capabilities of multimodal large language models in multi-graph visual reasoning tasks and systematically evaluates existing models.
WiseEdit: Benchmarking Cognition- and Creativity-Informed Image Editing
Kaihang Pan (Zhejiang University), Siliang Tang (Zhejiang University)
GenerationTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Constructed the WiseEdit benchmark to evaluate cognition- and creativity-driven image editing.
WISER: Wider Search, Deeper Thinking, and Adaptive Fusion for Training-Free Zero-Shot Composed Image Retrieval
Tianyue Wang (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)
RetrievalLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a training-agnostic zero-shot compositional image retrieval framework called WISER, which enhances retrieval quality by fusing T2I and I2I retrieval paths through a retrieval-validate-improve loop.
WiTTA-Bench: Benchmarking Test-Time Adaptation for WiFi Sensing
Bing Li (UESTC), Wei Cui (A*STAR)
Domain AdaptationConvolutional Neural NetworkBenchmark
🎯 What it does: Constructed WiTTA-Bench, a systematic benchmark for evaluating test-time adaptation (TTA) methods in WiFi human activity recognition (HAR), providing two adaptation protocols (OTTA and TTDA), 20 representative methods, three categories of physics-induced domain shifts (cross-environment, cross-subject, cross-device), and releasing a new dataset, WiHAR-Dual, to cover cross-device scenarios.
WOD-E2E: Waymo Open Dataset for End-to-End Driving in Challenging Long-tail Scenarios
Runsheng Xu (Waymo LLC), Dragomir Anguelov (Waymo LLC)
Autonomous DrivingRecurrent Neural NetworkTransformerLarge Language ModelReinforcement LearningDiffusion modelVideoTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the Waymo Open Dataset for End-to-End Driving (WOD-E2E) dataset, containing 4,021 segments of approximately 12 hours of rare long-tail driving scenarios, accompanied by human expert rating labels and a novel Rater Feedback Score (RFS) evaluation metric, aiming to assess the performance of visual end-to-end driving models in safety-critical scenarios;
WonderZoom: Multi-Scale 3D World Generation
Jin Cao (Stanford University), Jiajun Wu (Stanford University)
GenerationDiffusion modelGaussian SplattingImage
🎯 What it does: Generate a multi-scale 3D world from a single image, supporting user interactive zooming and generating new details at any scale level.
World in a Frame: Understanding Culture Mixing as a New Challenge for Vision-Language Models
Eunsu Kim (KAIST), Alice Oh (KAIST)
Large Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the CultureMix benchmark to evaluate the understanding capabilities of large vision-language models in culturally mixed scenarios;
WorldGen: From Text to Traversable and Interactive 3D Worlds
Dilin Wang (Reality Labs, Meta), Andrea Vedaldi (Reality Labs, Meta)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityMesh
🎯 What it does: Based on text prompts, an LLM-driven procedural generator creates navigation meshes and rough layouts, followed by a depth-conditioned image generator to build scene themes. Subsequently, a 3D gamma diffusion model conditioned on navigation meshes performs global 3D reconstruction, which is then decomposed into individual objects. Details are enhanced through high-resolution images, mesh refinement, and texture generation, ultimately producing a complete, navigable, and interactive 3D world directly usable in game engines.
WorldLens: Full-Spectrum Evaluations of Driving World Models in Real World
Ao Liang (National University Of Singapore), Ziwei Liu (Nanyang Technological University)
Autonomous DrivingExplainability and InterpretabilityTransformerSupervised Fine-TuningVision Language ModelWorld ModelVideoMultimodalityBenchmark
🎯 What it does: This work constructs the WorldLens benchmark to uniformly evaluate generation, reconstruction, behavior following, downstream tasks, and human preferences of driving world models, providing corresponding quantitative metrics and human-annotated data;
WorldMM: Dynamic Multimodal Memory Agent for Long Video Reasoning
Woongyeong Yeo (Korea Advanced Institute Of Science And Technology), Sung Ju Hwang (Korea Advanced Institute Of Science And Technology)
RetrievalComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelAgentic AIVision Language ModelVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the WorldMM framework, leveraging multi-modal multi-scale memory (episodic, semantic, visual) and adaptive retrieval agents to achieve efficient reasoning and question answering for long videos spanning hours to weeks.
WorldReel: 4D Video Generation with Consistent Geometry and Motion Modeling
Shaoheng Fang (University of Texas at Austin), Qixing Huang (University of Texas at Austin)
GenerationDepth EstimationTransformerDiffusion modelAuto EncoderOptical FlowVideoPoint Cloud
🎯 What it does: Trained a 4D video generation model called WorldReel, capable of simultaneously generating RGB videos, point clouds per frame, camera trajectories, optical flow, and scene flow while ensuring spatiotemporal consistency;
WorldStereo: Bridging Camera-Guided Video Generation and Scene Reconstruction via 3D Geometric Memories
Yisu Zhang (Zhejiang University), Chunchao Guo (Tencent Hunyuan)
GenerationDepth EstimationDiffusion modelOptical FlowVideoPoint Cloud
🎯 What it does: Propose the WorldStereo framework, leveraging two geometric memories (Global Geometric Memory GGM and Spatial Stereo Memory SSM) to achieve high-quality video generation from single views and consistent 3D reconstruction.
WPT: World-to-Policy Transfer via Online World Model Distillation
Guangfeng Jiang (University of Science and Technology of China), Xu Yan (Huawei Foundation Model Department)
Autonomous DrivingKnowledge DistillationTransformerWorld ModelMultimodality
🎯 What it does: Propose a training framework named WPT, which utilizes a pre-trained world model to perform online distillation of strategies during training, ultimately achieving a lightweight driving policy capable of real-time inference without relying on the world model.
Write Where It Matters: Policy-Guided Watermarks for 3D Gaussian Splatting
Nan Li (Tianjin University), Liang Wan (Tianjin University)
Safty and PrivacyReinforcement LearningAuto EncoderGaussian SplattingImage
🎯 What it does: Introduces Write Where It Matters (W2M), a reinforcement learning-based 3D Gaussian Splatting watermarking framework that adaptively selects writing positions and intensity within the scene to embed copyright information into 3D Gaussian representations without compromising rendering quality.
WRIVINDER: Towards Spatial Intelligence for Geo-locating Ground Images onto Satellite Imagery
Chandrakanth Gudavalli (Mayachitra Inc), B.S. Manjunath (Mayachitra Inc)
Pose EstimationDepth EstimationConvolutional Neural NetworkGaussian SplattingSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: A zero-shot training framework for aligning ground images to satellite images called Wrivinder was constructed. It utilizes multi-view ground photos to reconstruct sparse geometry via SfM, then densifies it using 3D Gaussian Splatting, and generates a consistent zenith view through semantic ground planes and monocular depth scale information. Finally, it achieves geometrically accurate GPS positioning by performing self-supervised depth template matching with satellite images.
X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection
Youngseo Kim (KAIST), Junyong Noh (KAIST)
ClassificationAnomaly DetectionRepresentation LearningConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelContrastive LearningVideoMultimodalityAudio
🎯 What it does: To address deepfake videos, this paper proposes X-AVDT, which extracts video composites and audio-visual cross-attention features by performing DDIM inversion on a pre-trained diffusion model, and then classifies the fused features.
X-band Radar Non-Line-of-Sight Imaging
Dongyu Du (Princeton University), Felix Heide (Princeton University)
TransformerImagePoint CloudPhysics Related
🎯 What it does: Propose a non-line-of-sight (NLOS) imaging method based on 10 GHz X-band radar, utilizing long wavelengths to convert diffuse reflection into specular reflection for large-scale hidden scene perception.
X-Part: High Fidelity And Structure Coherent Shape Decomposition And Completion
Xinhao Yan (Shanghaitech), Chunchao Guo (Tencent Hunyuan)
SegmentationGenerationTransformerDiffusion modelFlow-based ModelAuto EncoderPoint CloudMesh
🎯 What it does: Propose X-Part, a diffusion-based framework that decomposes complete 3D shapes into high-fidelity, structurally consistent editable parts using bounding boxes and point-level semantic features, and provides an interactive editing pipeline.
X-PCR: A Benchmark for Cross-modality Progressive Clinical Reasoning in Ophthalmic Diagnosis
Gui Wang (Shenzhen University), Linlin Shen (Shenzhen University)
TransformerLarge Language ModelVision Language ModelImageMultimodalityBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Proposed the X-PCR benchmark, using a six-stage progressive reasoning chain and cross-modal reasoning tasks to evaluate the capabilities of multimodal large language models in ophthalmic diagnosis.
X-WIN: Building Chest Radiograph World Model via Predictive Sensing
Zefan Yang (Rensselaer Polytechnic Institute), Pingkun Yan (Massachusetts General Hospital)
Image TranslationKnowledge DistillationAuto EncoderContrastive LearningWorld ModelBiomedical DataComputed Tomography
🎯 What it does: Propose X-WIN, a chest X-ray world model that achieves 3D CT knowledge distillation into 2D X-ray representations by predicting CT projections in the latent space.
x^2-Fusion: Cross-Modality and Cross-Dimension Flow Estimation in Event Edge Space
Ruishan Guo (Tsinghua University), Xinlei Chen (Harbin Institute of Technology)
Autonomous DrivingConvolutional Neural NetworkTransformerContrastive LearningOptical FlowImageMultimodalityPoint Cloud
🎯 What it does: Proposes a framework that unifies image, LiDAR, and event camera tri-modal data by mapping them into a shared event edge space, where adaptive fusion and joint estimation of optical flow/scene flow are achieved.
XPaintNet: An eXtreme Lightweight Framework for Stereoscopic Conversion without Inpainting Network
Kihwan Yoon (BLUEDOT), Minyong Jeon (BLUEDOT)
Image TranslationGenerationConvolutional Neural NetworkOptical FlowImageVideo
🎯 What it does: Propose an interpolation-free bidirectional warp and fusion framework named XPaintNet, achieving the generation of stereo right views from single-view images.
XSeg: A Large-scale X-ray Contraband Segmentation Benchmark For Real-World Security Screening
Hongxia Gao (Xi'an Jiaotong University), Kaijie Zhang (South China University Of Technology)
SegmentationTransformerSupervised Fine-TuningImageBiomedical DataBenchmark
🎯 What it does: This paper proposes the largest X-ray prohibited item segmentation dataset, XSeg, and constructs an APSAM model based on SAM with an energy-aware encoder and an adaptive point generator for high-precision occlusion and cross-domain segmentation.
YieldSAT: A Multimodal Benchmark Dataset for High-Resolution Crop Yield Prediction
Miro Miranda (RPTU Kaiserslautern-Landau), Andreas Dengel (University of Groningen)
Domain AdaptationRecurrent Neural NetworkImageMultimodalityTabularTime SeriesBenchmarkAgriculture Related
🎯 What it does: The study proposes and releases the large-scale, multi-modal, high-resolution YieldSAT dataset, conducts benchmark experiments with various deep learning models on this dataset, and explores the impact of distribution shift on model performance.
Yo'City: Personalized and Boundless 3D Realistic City Scene Generation via Self-Critic Expansion
Keyang Lu (Peking University), Ming Li (Guangdong Laboratory of Artificial Intelligence and Digital Economy)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextRetrieval-Augmented Generation
🎯 What it does: Propose the Yo'City multi-agent framework to achieve user-defined, infinitely scalable 3D city generation.
YOLO-Master: MOE-Accelerated with Specialized Transformers for Enhanced Real-time Detection
Xu Lin (Tencent Youtu Lab), Jun Liu (Tencent Youtu Lab)
Object DetectionConvolutional Neural NetworkTransformerMixture of ExpertsImage
🎯 What it does: Propose YOLO-Master, a real-time object detection model integrating Efficient Sparse Mixture-of-Experts (ES-MoE) into the YOLO framework, which dynamically allocates computational resources based on image complexity.
YOLO-ULM: Ultra-Lightweight Models for Real-Time Object Detection
Shasha Han (Ocean University of China), Xuebo Li (Ocean University of China)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: Proposed the YOLO-ULM ultra-lightweight real-time object detection model, which significantly reduces model parameters and computational cost while maintaining high accuracy.
YOSE: You Only Select Essential Tokens for Efficient DiT-based Video Object Removal
Chenyang Wu (Nankai University), Chongyi Li (Nankai University)
RestorationComputational EfficiencyTransformerDiffusion modelFlow-based ModelVideo
🎯 What it does: Proposes the YOSE framework, achieving efficient video object removal by processing only masked regions based on the Diffusion Transformer;
You Only Erase Once: Erasing Anything without Bringing Unexpected Content
Yixing Zhu (Sun Yat-sen University), Wei-Shi Zheng (Sun Yat-sen University)
Image HarmonizationRestorationDiffusion modelGenerative Adversarial NetworkImage
🎯 What it does: Propose a few-step diffusion model named YOEO for single-erase target objects while maintaining context consistency.
Your Classifier Can Do More: Towards Balancing the Gaps in Classification, Robustness, and Generation
Kaichao Jiang (Hefei University of Technology), Richang Hong (Hefei University of Technology)
ClassificationGenerationAdversarial AttackImage
🎯 What it does: Proposed the Energy-based Joint Distribution Adversarial Training (EB-JDAT) method, which unifies the tasks of classification, robustness, and generation.
Your Dissimilarities Define You: Complementary Learning Exploiting Class Diversities
Dimitrios Katsikas (Aristotle University of Thessaloniki), Anastasios Tefas (Aristotle University of Thessaloniki)
ClassificationKnowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Propose a new supervised loss called Complementary Dissimilarity Loss (CDL), which explicitly models the dissimilarity of non-target classes, providing complementary learning signals to traditional cross-entropy.
Your Latent Mask is Wrong: Pixel-Equivalent Latent Compositing for Diffusion Models
Rowan Bradbury (Bradbury Group), Dazhi Zhong (Bradbury Group)
Image HarmonizationRestorationGenerationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Propose the Pixel-Equivalent Latent Compositing (PELC) principle and implement a lightweight Transformer-based DecFormer to perform pixel-equivalent mixing in the latent space within Diffusion models, enhancing the boundary quality of masked editing and inpainting.
Your One-Stop Solution for AI-Generated Video Detection
Long Ma (University of Science and Technology of China), Zhen Bi (Huzhou University)
Data SynthesisAnomaly DetectionVideoTextMultimodalityBenchmark
🎯 What it does: Proposes AIGVDBench, an AI-generated video detection benchmark containing 31 generative models and over 440,000 videos, covering multiple tasks such as text-video, image-video, and video-video, and evaluates 33 existing detectors 1,500 times.
Yume1.5: A Text-Controlled Interactive World Generation Model
Xiaofeng Mao (Shanghai AI Laboratory), Kaipeng Zhang (Shanghai AI Laboratory)
GenerationData SynthesisTransformerDiffusion modelVideoTextMultimodality
🎯 What it does: This paper proposes Yume1.5, a long-sequence interactive virtual world video model that can be controlled via keyboard, generate text instructions, and perform real-time inference.
Z-Order Transformer for Feed-Forward Gaussian Splatting
Can Wang (University of Hong Kong), Dong Xu (University of Hong Kong)
GenerationDepth EstimationTransformerGaussian SplattingImage
🎯 What it does: Propose a feed-forward 3D Gaussian Splatting framework based on Z-order Transformer, achieving high-quality novel view rendering with a single forward pass.
Zero-Shot Depth Completion with Vision-Language Model
Zhiqiang Yan (National University of Singapore), Gim Hee Lee (National University of Singapore)
Depth EstimationLarge Language ModelSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: This paper proposes a zero-shot depth completion framework that leverages a vision-language model to convert sparse depth images into full depth.
Zero-shot Detection of AI-Generated Image via RAW-RGB Alignment
Haiwei Wu (University of Electronic Science and Technology of China), Jiantao Zhou (University of Macau)
ClassificationAnomaly DetectionGraph Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Achieve zero-shot detection of AI-generated images without prior information by constructing RAW-RGB alignment traces, while maintaining recognition capability for physical remapping (print+scan, screen+shot).
Zero-Shot Image Denoising via Hybrid Prior-Guided Pseudo Sample Generation
Xiaole Zhao (Southwest Jiaotong University), Tianrui Li (Southwest Jiaotong University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a zero-shot image denoising framework that generates pseudo-samples using local gradient aggregation downsampling and global Gaussian-constrained random sampling, and trains the model with a spectral-weighted loss.
Zero-Shot Reconstruction of Animatable 3D Avatars with Cloth Dynamics from a Single Image
Joohyun Kwon, Gyeongsik Moon
Image TranslationGenerationTransformerSupervised Fine-TuningGaussian SplattingOptical FlowImageVideo
🎯 What it does: Proposed DynaAvatar, which can generate animatable 3D human avatars under zero-shot conditions using only a single image, and simulate non-rigid dynamics of clothing with motion.
ZeroIDIR: Zero-Reference Illumination Degradation Image Restoration with Perturbed Consistency Diffusion Models
Hai Jiang (Sichuan University), Shuaicheng Liu (University of Electronic Science and Technology of China)
RestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Proposes ZeroIDIR, a zero-reference (without requiring paired high-quality images) diffusion model framework, which first performs adaptive gamma correction (AGCM) on illumination-degraded images and then restores details and suppresses noise through a perturbation-consistency diffusion model (PCDM).
ZINA: Multimodal Fine-grained Hallucination Detection and Editing
Yuiga Wada (Keio AI Research Center), Graham Neubig (Carnegie Mellon University)
Anomaly DetectionGraph Neural NetworkLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This paper proposes the multi-modal fine-grained hallucination detection and editing task, and realizes detection, classification, and editing through the ZINA system;
ZipMap: Linear-Time Stateful 3D Reconstruction via Test-Time Training
Haian Jin (Google DeepMind), Aleksander Hołyński (Google DeepMind)
Pose EstimationDepth EstimationTransformerImagePoint Cloud
🎯 What it does: Propose ZipMap, a linear-time, stateful Transformer model that generates camera poses, depth maps, and point clouds from large-scale image collections in a single pass, and can perform real-time queries on implicit scene representations.
ZOO-Prune: Training-Free Token Pruning via Zeroth-Order Gradient Estimation in Vision-Language Models
Youngeun Kim (Amazon), Sungeun Hong (Sungkyunkwan University)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Proposes a training-agnostic, attention-agnostic visual-language model (token pruning) framework called ZOO-Prune, which measures the sensitivity of visual tokens using zeroth-order gradient estimation in the projection layer and combines it with diversity selection to obtain an efficient subset of tokens.
Zoo3D: Zero-Shot 3D Object Detection at Scene Level
Andrey Lemeshko (Higher School Of Economics), Maksim Kolodiazhnyi (Lomonosov Moscow State University)
Object DetectionComputational EfficiencyTransformerVision Language ModelImagePoint Cloud
🎯 What it does: Proposed two zero-training/self-supervised open-vocabulary 3D object detection frameworks, Zoo3D0 and Zoo3D1, capable of directly performing object detection on point clouds, images with known or unknown camera poses.
ZoomEarth: Active Perception for Ultra-High-Resolution Geospatial Vision-Language Tasks
Ruixun Liu (Xi'an Jiaotong University), Xiangyong Cao (Xi'an Jiaotong University)
Large Language ModelReinforcement LearningVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the ZoomEarth framework, achieving active perception (active cropping-zooming) to handle visual-language tasks on ultra-high-resolution remote sensing images;