see2023/Bert-VITS2-ext

Extends Bert-VITS2 to simultaneously generate facial expressions and body animations synchronized with TTS output by extracting latent variables (z) from the posterior encoder and training auxiliary LSTM+MLP modules for expression mapping. Recent 2025 improvements replaced 1D convolutions with 2D Conv2D operations and removed Transformer modules for efficiency. Integrates with Unreal Engine MetaHuman via Live Link Face protocol, supports MotionGPT for gesture generation, and extends compatibility to CosyVoice and GPT-SoVITS TTS backends.

539 stars. No commits in the last 6 months.