Face puppetry in ComfyUI, minus the ONNX tax

What it does

Wraps KwaiVGI’s LivePortrait model in ComfyUI nodes so you can animate a static portrait from a webcam or video feed. The author converted the original pickle weights to safetensors, replaced the ONNX landmark runner with a PyTorch .pth model, and added MediaPipe and face-alignment as alternatives to the default Insightface detector. It handles image-to-video, video-to-video, and near-realtime webcam pipelines, auto-skipping frames when no face is found.

The interesting bit

The rework is aggressively pragmatic: it eliminates the ONNX runtime dependency entirely, sidesteps Insightface’s non-commercial license trap, and still hits ~80–100 ms delay inside ComfyUI. The author even admits you “just have to trust me” on the converted landmark model—a refreshingly honest note in a space full of opaque binary blobs.

Key highlights

• Near-realtime webcam face reenactment inside ComfyUI with ~80–100 ms latency.
• All weights converted from pickle to safetensors; models auto-download to ComfyUI/models/liveportrait.
• Three face detector options: Insightface (default, non-commercial only), Google MediaPipe (CPU-only on Windows, faster on CPU than Insightface), and face-alignment with BlazeFace back-camera model (better for small faces, long first-run warmup).
• Proper video-to-video mode with a smoothing algorithm.
• Improved Apple Silicon (MPS) support and auto-skipping of frames with no face detected.

Caveats

• The recent main-branch rework broke legacy workflows; old projects need the legacy branch.
• Insightface remains strictly non-commercial and is notoriously finicky to install, while MediaPipe is weaker and cannot use the GPU on Windows.
• The converted PyTorch landmark model comes with no ONNX verification; the author admits you just have to trust it.

Verdict

Worth a look if you want LivePortrait inside a ComfyUI pipeline without wrestling with pickle files or license ambiguity. Skip it if you need production-grade Windows GPU face detection out of the box, or if you are still maintaining pre-rework workflows.