real-time AI lip-sync, fully open source, from Tencent Music's research lab it's called MuseTalk - takes any video + any audio and syncs the mouth to match, live → 30fps+ real-time inference on an NVIDIA V100 → works in Chinese, English, Japanese → operates in latent space (not a diffusion model - single-step inpainting, so it's fast) → inference, training code, AND weights are all released - v1.5 added GAN + perceptual + sync loss for a real jump in quality over v1.0 → pairs with MuseV (their video-generation model) for a full text/image-to-talking-avatar pipeline 6.4k stars, MIT licensed, weights on Hugging Face with a live Gradio demo honest limitations, straight from their own README: some jitter since it generates frame-by-frame, and fine details like mustaches or exact lip color don't always carry over perfectly