What it does LivePortrait takes a source image or video of a face (human, cat, or dog) and reanimates it to match the expressions and head pose from a second “driving” video. The result is a new video where the subject from the first clip moves like the second. It runs as a PyTorch inference pipeline with a Gradio interface, and there’s a one-click Windows installer for the command-line averse.

The interesting bit The project isn’t just a research dump—it’s been adopted by Kuaishou, Douyin, Jianying, and WeChat Channels, which is a decent signal that the efficiency claims hold up. The “stitching and retargeting control” in the title refers to how it keeps the generated head attached to the original body and lets you dial specific expressions up or down, rather than just copy-pasting the whole driving face.

Key highlights

• Humans, cats, and dogs supported (animals need a separate X-Pose dependency and NVIDIA GPU)
• Image-driven mode and regional control for fine-tuned edits
• macOS Apple Silicon supported for humans mode, though ~20x slower than an RTX 4090
• Windows one-click installer with auto-updates
• HuggingFace Space available if you don’t want to install anything

Caveats

• Animals mode is Linux/Windows NVIDIA only; X-Pose doesn’t build on macOS
• Windows users with CUDA 12.4+ may hit “unknown issues” and need to downgrade to 11.8
• Intel Macs are untested

Verdict Worth a look if you need programmatic portrait animation or want to prototype video content without touching After Effects. Skip it if you need real-time performance or production guarantees—the README is honest about platform quirks and speed tradeoffs.