LivePortrait

Give LivePortrait a still photo or short clip of a face plus a driving video, and it transfers head pose, expressions, blinks and lip motion onto the source using a motion extractor, a warping module and a SPADE-style generator instead of a diffusion model; the per-module timings in its speed table add up to about 15 ms per frame on an RTX 4090 with torch.compile. Small stitching and retargeting networks paste the animated face back into the frame and let you set eye and lip openness separately. Extras include an Animals mode for cats and dogs, video-to-video portrait editing, an image-driven mode, regional control, pose editing in a Gradio app, and reusable .pkl motion templates. Setup needs Python 3.10 via Conda, FFmpeg and PyTorch with CUDA; Apple Silicon handles human faces through MPS far more slowly, Animals mode needs an op compiled for X-Pose on NVIDIA hardware, and a one-click Windows package exists. The code is MIT, but the bundled InsightFace detection models are non-commercial, so commercial users must replace them. The authors say Kuaishou, Douyin, Jianying and WeChat Channels use it.