Efficient face puppetry, now with pets

What it does LivePortrait takes a source portrait—an image or video—and a driving video, then retargets the driver’s facial motion onto the source. The output keeps the subject’s identity while adopting the driver’s expressions and pose, and the system supports humans, cats, and dogs. A built-in stitching step attaches the animated face back to the original body or frame.

The interesting bit The project treats portrait animation as a retargeting and stitching problem rather than open-ended generation, emphasizing precise control over pose and expression. That efficiency-first framing appears to have paid off: the authors note it is already deployed inside Kuaishou, Douyin, Jianying, and WeChat Channels.

Key highlights

• Supports humans, cats, and dogs via separate model modes
• Includes regional control, precise portrait editing, and video-to-video (v2v) synthesis
• Ships with a Gradio interface, HuggingFace Space, and a Windows one-click installer
• Runs on Apple Silicon via MPS fallback, though the README notes this is roughly 20× slower than an RTX 4090

Caveats

• Animals mode requires building a custom CUDA operator (MultiScaleDeformableAttention) and is only tested on Linux and Windows with NVIDIA GPUs
• The README warns that newer CUDA versions on Windows (12.4, 12.6) can cause unspecified issues and recommends downgrading to 11.8
• Intel-based Macs are untested, and animals mode is unsupported on macOS entirely

Verdict A practical pick for developers and creators who need fast, identity-preserving face animation without generating frames from scratch. Skip it if you need full-body motion or a turnkey, cross-platform pipeline for animals.