Beyond research
Alongside research, I build products and contribute to open-source software. Here is a closer look at that work, followed by a little of life outside the lab.
LuckyShort
Technical co-founder
Video agents in commercial production
LuckyShort develops AI-native short dramas for international audiences and makes its production tools available to partner studios. The work brings together generative models, a repeatable production process, and an understanding of how stories connect with audiences across cultures.
As a technical co-founder, I lead work on multimodal generation, intelligent model routing, and the end-to-end workflow behind our 天鹏座短剧 Agent. Different scenes call for different strengths: emotional expression, action, visual atmosphere, or transitions. Our routing system interprets the creative intent and selects suitable model workflows, while shared references and production checks help maintain continuity.
The system connects script analysis, shot planning, asset coordination, generation, and review. Generation feedback, viewing data, and regional preferences inform later model and workflow improvements. My focus is on making these capabilities usable in sustained production, where consistency and the cost of revision matter alongside visual quality.
Reported time per person to generate an approximately 30-episode series
Generation attempts for standard genres in the reported workflow
Figures reported by our team in a 扬帆出海 interview on 2 June 2026. Human review and finishing remained part of the process; the interview reported higher generation-attempt counts for more demanding genres.
Read the interview · 中文 ↗Technical overview ↗Example video ↗
ViMax
Open-source collaboration
with HKUDS
Making agentic video creation accessible
I worked with HKUDS to develop and open-source ViMax, a framework that coordinates the roles of screenwriter, director, producer, and video generator. It connects narrative planning, character references, storyboards, shot generation, and final assembly in a workflow that others can inspect and adapt.
The project supports starting from an idea, a screenplay, or long-form fiction. An interactive agent and project workspace let creators revise plans, inspect storyboards, review generated assets, and resume work. This turns a sequence of model calls into a process that can be understood and guided.
For me, open source is a way to share the system behind the result. ViMax gives researchers and developers a concrete foundation for experimenting with model choices, visual consistency, and how people collaborate with creative agents.
12.3k GitHub stars · 1.8k forksMIT licensed · GitHub snapshot, 9 Sep 2026
Outside the lab
I enjoy photography, singing, painting, calligraphy, and badminton. These are a few photographs and pieces from my personal collection.
A few university memories: second prize in the Golden Autumn Photography Competition and runner-up in men's doubles at the Wuhan University Freshmen Badminton Tournament.