Audio-driven talking avatars from a photo and a voice, by Meituan LongCat. Model · Code · Project page
Longer, descriptive prompts (appearance, action, setting) give the most consistent results.
First ~3.7 s, +3.2 s per extra segment.