Minyang Li
I am an undergraduate student in Artificial Intelligence at The Hong Kong University of Science and Technology (Guangzhou), supervised by Prof. Ying-Cong Chen. I have spent wonderful times collabrating with Zhen Yang in Prof. Chen's lab. I was also privileged to be supervised by Dr. Qichun Yang at the begining of my research journey. My research interests focus on how reward or preference signals, test-time feedback, and trajectory-level guidance can be used to improve the controllability, reliability, and alignment of diffusion-based generative models.
Education
B.E. in Artificial Intelligence.
Exchange student.
Publications
Zhen Yang*, Guibao Shen*, Minyang Li*, Liang Hou, Mushui Liu, Luozhou Wang, Xin Tao, and Ying-Cong Chen
CVPR Findings, 2026
Haoyuan Yu, Qichun Yang, Liuqian Yu, Xia Li, Minyang Li, and Yingxian Yang
Water, 2026
Internship
Mentor: Dr. Tianyi Zhang
Knowin AI, Nanshan, Shenzhen, China
Selected Projects
Independent, Course Project in Deep Learning
Proposed VGFE, an inversion-free flow-based image editing framework that uses VQA-guided editing strength search to automatically select optimal edit strength while reducing computational overhead. Extended the method to multi-object editing by designing a reassembly and re-editing pipeline tailored for inversion-free diffusion/flow models. (Report)
Collaborator: Jian Yang, Course Project in CompTec for Sketch-based Creativity
Developed CalliGen, a sketch-based Chinese calligraphy generation system that transforms rough handwritten skeletons into stylized calligraphic characters while preserving users' structural intent. Designed an efficient pipeline to construct a large-scale paired dataset of Chinese character images and skeletons for training. Co-designed an interactive prototype supporting sketch editing, style selection, and real-time generation. (Report) (Code) (Demo video)
Collaborator: Zhiyi Chen, Zhiling Li, Course Project in Intro2NLP
To address the limitations of global embedding similarity in detecting entity, date, number, and span-level errors, we proposed FS-BGE, a BGE-M3 ColBERT-based similarity score with factual token reweighting and answer-span mismatch penalties, and I further explored bidirectional NLI-based entailment scoring and scoring in hyperbolic space. (Report) (Code)
Homepage Template
This homepage is adapted from the open-source m-niemeyer/m-niemeyer.github.io academic homepage template.