Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, Hongsheng Li
Thirteenth International Conference on Learning Representations. ICLR 2025.
We conducted an in-depth and meticulous theoretical analysis and empirical validation of flow matching, rectified flow, and the rectification operation. We demonstrated that the rectification operation is also applicable to general diffusion models, and that flow matching is fundamentally no different from the traditional noise addition methods in DDPM. Our related blog post on ZHIHU garnered over 10k views and approximately 400 likes.
arXiv  •   GitHub  •   Poster

Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, Xiaogang Wang, Hongsheng Li
Conference on Neural Information Processing Systems. NeurIPS 2024.
We validated and enhanced the effectiveness of consistency models for text-to-image and text-to-video generation. Our method has been adopted by the FastVideo project, successfully accelerating SoTA video diffusion models including HunyuanVideo and WAN.
Project Page  •   Github  •   Paper   •   Poster

Fu-Yun Wang, Yunhao Shui, Jingtan Piao, Keqiang Sun, Hongsheng Li
Thirteenth International Conference on Learning Representations. ICLR 2025.
We proposed a general, simple yet effective method for strengthened diffusion preference optimization, improving the alignment of generated outputs with user preferences.
Paper  •   Poster   •   Github

Xiaoyu Shi*, Zhaoyang Huang*, Fu-Yun Wang*, Weikang Bian*, Dasong Li, Yi Zhang, Manyuan Zhang, Kachun Cheung, Simon See, Hongwei Qin, Jifeng Dai, Hongsheng Li
Special Interest Group on GRAPHics and Interactive Techniques.SIGGRAPH 2024.
SIGGRAPH 2024 Technical Papers Trailer
Project Page  •   GitHub  •   arXiv

Fu-Yun Wang, Zhaoyang Huang, Qiang Ma, Xudong Lu, Weikang Bian, Yijin Li, Yu Liu, Hongsheng Li
European Conference on Computer Vision. ECCV 2024.
ECCV 2024 Oral Presentation
Project Page  •   Paper

Da-Wei Zhou*, Fu-Yun Wang*, Han-Jia Ye, De-Chuan Zhan
SCIENCE CHINA Information Sciences. SCIS.
PyCIL stands out as a comprehensive and user-friendly Python toolbox for Class-Incremental Learning. Boasting nearly 1000 stars on GitHub, it is currently the most widely collected CIL toolkit, adopted by researchers worldwide. It provides a standardized framework for implementing and evaluating various CIL algorithms, fostering reproducible research and accelerating advancements in the field.
Github  •   arXiv  •   Media   •  

Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan
European Conference on Computer Vision. ECCV 2022.
FOSTER introduces a novel approach to Class-Incremental Learning by combining feature boosting and compression strategies. This method effectively mitigates catastrophic forgetting while promoting the learning of new classes, showcasing robust performance in dynamic learning environments.
Github  •   arXiv  •