video-retalking

Why we included this project

VideoReTalking is a research codebase for one narrow job: taking a talking-head video and a new speech track, then re-rendering the speaker's mouth so the lips match the words. It runs expression normalization and a face-enhancement pass rather than just pasting lips on, so results hold up on ordinary footage without manual landmark alignment. You can set a neutral or smiling baseline and nudge the upper-face expression through command-line flags. As a SIGGRAPH Asia 2022 reference implementation, expect to download models and set up CUDA yourself, and keep extreme head poses out of your shots. For dubbing and synthetic presenter work where you'd rather not depend on a cloud API, it's a solid self-hosted option.