Using Human Performances to Guide AI-Generated Video
At the 1024 Modularity Community Day AI event, 奇点川越 founder Cheng Jiawei described his team’s method of recording human performances as references for AI-generated microexpressions and movement. He said this produced more stable and controllable results than relying only on text descriptions, though the report provides no independent validation of that comparison. The team first settles scripts, tests and fixes prompt templates, and then uses Agents to generate material in batches. He also noted that maintaining a character’s appearance solves only part of the problem: scene and spatial continuity between shots remain difficult. Story planning, material authorization, careful treatment of direct imitation, and records of the creative process remain necessary.
The talk presents a concrete, practitioner-reported workflow for controlling expressive AI-generated video while also identifying unresolved constraints in spatial continuity, story planning, and rights management.

在 1024 模力工场 · AI社区日"主论坛上,复旦大学中文系硕士、奇点川越创始人程佳伟以《影视制作工具发展史:从胶片机械到生成式AI,生成方式如何影响审美》为题作主题分享。

程佳伟指出,想让 AI 演出细腻的情绪,与其反复修改提示词,不如自己先演一遍。他分享了团队的一个实用方法:录下真人表演作为参考,让 AI 跟着处理微表情和动作,比单靠文字描述更稳定、更可控。