Ziyang Xu (徐子扬)

I am a Ph.D. student in Information and Communication Engineering at the School of Electronic Information and Communications (EIC), Huazhong University of Science and Technology (HUST, 华中科技大学), advised by Professors Xinggang Wang and Wenyu Liu. I began my Ph.D. studies in September 2025. Prior to this, I received my M.E. degree in Information and Communication Engineering from EIC, HUST (华中科技大学) in 2025, and my B.E. degree in Information Engineering from the School of Information Engineering (IE), Wuhan University of Technology (WHUT, 武汉理工大学) in 2022. My research spans visual generation, efficient multimodal models, visual perception and memory, and AI for science, with a current focus on generative AI.

Email  |  Google Scholar  |  GitHub  |  ORCID  |  X

Latest Update: September 1, 2026

Portrait of Ziyang Xu

Highlight

PixelHacker and Moebius, my first/co-first-authored image inpainting projects, both achieved the No. 1 daily ranking on Hugging Face.
An early quantized on-device image inpainting model that I developed during my internship at vivo AI Lab—predating PixelHacker and Moebius, and related to the later PixelHacker line of work—has been rolled out across vivo’s full smartphone lineup, serving 400 million active users worldwide. You can find it in Gallery under AI Retouching → AI Eraser.
My research on multi-frame generative low-dose DSA (Digital Subtraction Angiography) imaging spans MoSt-DSA, GaraMoSt, GenDSA, and GenDSA-V2, with results published in ECAI, AAAI, Med, and Nature Medicine. This line of work contributes to a National Key R&D Program of China and has advanced from algorithm development to clinical validation and medical-device integration. More than 100,000 patients worldwide undergo DSA-guided procedures every day. In a completed prospective randomized controlled trial involving 1,068 patients, GenDSA-V2 reduced radiation exposure for patients and healthcare professionals by approximately two-thirds. Clinical trial registration: ChiCTR2400084789.
As of September 1, 2026, my open-source research projects have received 1,150+ GitHub stars and 110+ forks in total, with a single project reaching 580+ stars.

Research

My research spans visual generation, efficient multimodal models, visual perception and memory, and AI for science, with a particular emphasis on generative AI and medical imaging. Publications are grouped by research theme; my name is bolded, * denotes equal contribution, and † denotes project leadership.

1/3 Efficient Visual Generation and Image Inpainting

I develop image generation systems that balance high-quality synthesis and structural-semantic consistency with practical model efficiency.
Input example for Moebius image inpainting Moebius image inpainting result
Representative Work

Moebius: 0.2B Lightweight Image Inpainting Framework with 10B-Level Performance

European Conference on Computer Vision (ECCV), 2026
Paper / Code / Project
Kangsheng Duan*, Ziyang Xu*†, Wenyu Liu, Xiaohu Ruan, Xiaoxin Chen, Xinggang Wang
* Equal contribution; † project leadership
Role: Co-first Author · Project Lead
This work was completed while Ziyang Xu and Kangsheng Duan were research interns at vivo AI Lab.

Moebius extends the PixelHacker line of work by asking whether high-quality image inpainting requires continued model scaling. It combines a compact architecture with knowledge distillation, using only 0.22B parameters—less than 2% of FLUX.1-Fill-Dev—while matching or surpassing it across six natural and portrait benchmarks and reducing total inference time by more than 15×.
Input example for PixelHacker image inpainting PixelHacker image inpainting result
Representative Work

PixelHacker: Image Inpainting with Structural and Semantic Consistency

arXiv preprint, 2025 Extended manuscript under review
Paper / Code / Project
Ziyang Xu†, Kangsheng Duan, Xiaolei Shen, Zhifeng Ding, Wenyu Liu, Xiaohu Ruan, Xiaoxin Chen, Xinggang Wang
† Project leadership
Role: First Author · Project Lead
This work was completed while Ziyang Xu and Kangsheng Duan were research interns at vivo AI Lab.

PixelHacker introduces Latent Categories Guidance (LCG), which separately models foreground and background semantic distributions and injects them into diffusion-based denoising. Trained with 14 million image-mask pairs, it improves structural and semantic consistency across natural and portrait inpainting benchmarks while remaining substantially smaller and faster than large industrial baselines.

2/3 Generative Low-Dose DSA Imaging

This research line progresses from direct multi-frame interpolation and efficient generation to large-scale pretraining, multi-center validation, and prospective randomized clinical evaluation.
Standard DSA image used for comparison with GenDSA-V2 Low-dose DSA image generated by GenDSA-V2
Representative Work

Generative AI-based low-dose digital subtraction angiography for intra-operative radiation dose reduction: a randomized controlled trial

Nature Medicine, 2026
Paper / Code (GenDSA) / Data / Trial Registration
Huangxuan Zhao*, Yaowei Bai*, Lei Chen*, Jinqiang Ma*, Yu Lei, Tao Sun, Linxia Wu, Ruiheng Zhang, Ziyang Xu, Xiaoyun Liang, Yi Li, Yan Huang, Yun Feng, Cheng Hong, Zhongrong Miao, Lin Long, Haidong Zhu, Jiahe Zheng, Lin Fan, Zhuting Fang, Peng Dong, Lefei Zhang, Xiaoyu Han, Bin Wang, Bin Liang, Xiangwen Xia, Xuefeng Kan, Chengcheng Zhu, Bo Du, Xinggang Wang, Chuansheng Zheng
* Equal contribution
Role: Algorithm Lead

GenDSA-V2 iterates our previously developed GenDSA system using data from 46,829 patients, more than 5 million DSA images, and 70 centers. In a completed prospective randomized controlled trial of 1,068 patients, it reduced radiation exposure for patients and healthcare professionals by approximately two-thirds while maintaining noninferior operation time and similar complication rates.
Standard DSA image used for comparison with GenDSA Low-dose DSA sequence generated by GenDSA
Representative Work

Large-scale pretrained frame generative model enables real-time low-dose DSA imaging: An AI system development and multi-center validation study

Med (Cell Press), 2025
Paper / Code
Huangxuan Zhao*, Ziyang Xu*, Lei Chen*, Linxia Wu*, Ziwei Cui, Jinqiang Ma, Tao Sun, Yu Lei, Nan Wang, Hongyao Hu, Yiqing Tan, Wei Lu, Wenzhong Yang, Kaibing Liao, Gaojun Teng, Xiaoyun Liang, Yi Li, Congcong Feng, Tong Nie, Xiaoyu Han, Dongqiao Xiang, Charles B.L.M. Majoie, Wim H. van Zwam, Aad van der Lugt, P. Matthijs van der Sluijs, Theo van Walsum, Yun Feng, Guoli Liu, Yan Huang, Wenyu Liu, Xuefeng Kan, Ruisheng Su, Weihua Zhang, Xinggang Wang, Chuansheng Zheng
* Equal contribution
Role: Co-first Author · Algorithm Lead

GenDSA is a large-scale pretrained multi-frame generative system for real-time low-dose DSA imaging. It was developed on approximately 3 million DSA images from 27,117 patients at 10 hospitals and evaluated on two additional datasets from 25 hospitals; using one-third of the clinical radiation dose, it generated frames in 0.07 seconds and achieved image quality comparable to full-sampled videos.
Input DSA frames for GaraMoSt interpolation DSA frames interpolated by GaraMoSt

GaraMoSt: Parallel Multi-Granularity Motion and Structural Modeling for Efficient Multi-Frame Interpolation in DSA Images

AAAI Conference on Artificial Intelligence (AAAI), 2025
Paper / Code
Ziyang Xu, Huangxuan Zhao, Wenyu Liu, Xinggang Wang
Role: First Author

GaraMoSt extends MoSt-DSA with parallel multi-granularity motion and structural feature extraction. The design improves noise suppression, accuracy, and visual quality within the same computational time scale, outperforming MoSt-DSA and natural-scene video frame interpolation baselines on DSA interpolation.
Input DSA frames for MoSt-DSA interpolation DSA frames interpolated by MoSt-DSA

MoSt-DSA: Modeling Motion and Structural Interactions for Direct Multi-Frame Interpolation in DSA Images

European Conference on Artificial Intelligence (ECAI), 2024
Paper / Code
Ziyang Xu, Huangxuan Zhao, Ziwei Cui, Wenyu Liu, Chuansheng Zheng, Xinggang Wang
Role: First Author

MoSt-DSA is the first deep learning method designed specifically for DSA frame interpolation. It models motion-structure interactions and directly generates an arbitrary number of intermediate frames at arbitrary time steps in a single forward pass, establishing the algorithmic foundation for the subsequent GaraMoSt and GenDSA systems.

3/3 Multimodal Generation and Visual Perception

I also study multimodal world generation and robust visual perception, including joint video-LiDAR generation and extremely small object detection in videos.
Driving scene inputs for Genesis Multimodal driving scene generated by Genesis

Genesis: Multimodal Driving Scene Generation with Spatio-Temporal and Cross-Modal Consistency

Conference on Neural Information Processing Systems (NeurIPS), 2025
Paper / Code / Project
Xiangyu Guo*, Zhanqian Wu*, Kaixin Xiong*, Ziyang Xu, Lijun Zhou, Gangwei Xu, Shaoqing Xu, Haiyang Sun, Bing Wang, Guang Chen, Hangjun Ye, Wenyu Liu, Xinggang Wang
* Equal contribution

Genesis jointly generates multi-view driving videos and LiDAR sequences with spatio-temporal and cross-modal consistency. It couples a DiT-based video diffusion model and a BEV-aware LiDAR generator through a shared latent space, achieving strong generation results on nuScenes and improving downstream segmentation and 3D detection.
Aerial video frame from the XS-VID dataset Extremely small objects detected in an XS-VID video frame

XS-VID: An Extremely Small Video Object Detection Dataset

arXiv preprint, 2024 Extended manuscript under review
Paper / Code / Dataset
Jiahao Guo, Ziyang Xu, Lianjun Wu, Fei Gao, Wenyu Liu, Xinggang Wang

XS-VID introduces a diverse aerial video benchmark for extremely small object detection, covering eight categories and objects below 32² pixels together with track and motion annotations. The accompanying YOLOFT baseline strengthens local feature association and temporal motion modeling, improving the accuracy and stability of small video object detection.

Internship

Research Intern, vivo AI Lab
October 2024 – August 2025 · Hangzhou, China
Worked on efficient on-device image inpainting and multimodal content generation, with a focus on model distillation and deployment-oriented optimization. This work led to PixelHacker and Moebius.

Selected Honors & Awards

Outstanding Master’s Graduate, Huazhong University of Science and Technology (HUST) (优秀硕士毕业生), 2025.
National Scholarship for Graduate Students (Master’s Level) (研究生国家奖学金，硕士研究生), 2024.
Awarded to only 0.2% of candidates nationwide.
Outstanding Graduate Student, HUST (三好硕士生), 2024.
Outstanding Graduate, Wuhan University of Technology (WHUT) (优秀本科毕业生), 2022.
Provincial Undergraduate Innovation and Entrepreneurship Training Program — Excellent Completion (优秀结题), 2021.
Team Lead “Research and Implementation of a Video-Based Vehicle-Assisted Driving Device.”
Outstanding Undergraduate Student, WHUT (三好本科生), 2021.

Selected Competitions

Team Lead Huawei Cup 19th China Graduate Mathematical Modeling Competition (华为杯第十九届中国研究生数学建模竞赛), National Second Prize (国家级二等奖), 2022.
Top 13% (前13%) of all teams.
Team Lead Artificial Intelligence Track of the 14th China College Student Computer Design Competition (中国大学生计算机设计大赛人工智能赛道), National Third Prize (国家级三等奖), 2021.
Among the 4,886 entries that advanced from the provincial round to the national competition, 8.5% received this award.
National College Student Integrated Circuit Innovation and Entrepreneurship Competition (全国大学生集成电路创新创业大赛), Third Prize in Hubei Province (湖北省三等奖), 2021.
Team Lead The 20th Innovation Cup College Students Extracurricular Academic Science and Technology Works Competition (创新杯大学生课外学术科技作品竞赛), Special Prize (特等奖) in the Mathematical Information Group (数理信息组), 2020.
I led the team to win the sole Special Prize (唯一特等奖) in the Mathematical Information category.
Team Lead National College Student Mathematical Modeling Competition (全国大学生数学建模竞赛), Second Prize in Hubei Province (湖北省二等奖), 2020.

Academic Service

Conference Reviewer: CVPR 2026; ECCV 2024 and 2026.
Journal Reviewer: IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) and Image and Vision Computing, 2026; International Journal of Computer Vision (IJCV), 2024.