LivePortrait: setup, driving-video preparation, motion templates and speed options

LivePortrait (Kuaishou) animates a still portrait, or edits a portrait video, using the motion of a driving video, with stitching and retargeting control. Updates through August 2024 added portrait video editing (v2v), pose editing in Gradio, Apple Silicon support, a Windows one-click package (run_windows.bat) and an Animals model. Setup: git clone https://github.com/KwaiVGI/LivePortrait, conda create -n LivePortrait python==3.9, pip install -r requirements.txt (requirements_macOS.txt on Apple Silicon), and install FFmpeg with ffmpeg and ffprobe. Download weights with git lfs (git clone https://huggingface.co/KwaiVGI/LivePortrait temp_pretrained_weights, then move them into pretrained_weights/) or from Google Drive or Baidu Yun. pretrained_weights must contain insightface/models/buffalo_l (2d106det.onnx, det_10g.onnx) and liveportrait/ with base_models (appearance_feature_extractor, motion_extractor, spade_generator, warping_module), landmark.onnx and retargeting_models/stitching_retargeting_module.pth. Run: python inference.py writes animations/s6--d0_concat.mp4 (driving video, input and result side by side). Use -s for a source image or video and -d for the driving video; see python inference.py -h. On Apple Silicon prefix PYTORCH_ENABLE_MPS_FALLBACK=1; the card warns this may be about 20x slower than an RTX 4090, and Intel Macs are not supported for that path. Driving video: crop to 1:1 (for example 512x512 or 256x256) focused on the head, minimise shoulder movement and start on a frontal, neutral face, or pass --flag_crop_driving_video and tune --scale_crop_driving_video and --vy_ratio_crop_driving_video. The auto-generated .pkl motion template can be passed as -d instead of the video to speed inference and avoid sharing the original driving video. Gradio: python app.py (with --server_port, --share, --server_name). --flag_do_torch_compile spends about a minute optimising on first run, then runs 20 to 30 percent faster; it is not supported on Windows or macOS. Intel Mac support for the app is unconfirmed in the card. Speed (RTX 4090, torch.compile, per frame): appearance extractor 0.82 ms, motion extractor 0.84 ms, SPADE generator 7.59 ms, warping module 5.21 ms, stitching and retargeting MLPs 0.31 ms; python speed.py measures this on NVIDIA GPUs. Community ComfyUI nodes include ComfyUI-LivePortraitKJ and comfyui-liveportrait. Licence: MIT. Based on the KlingTeam/LivePortrait card pinned to revision 82a4fa6735ca58432b6ce39301b4b9ee066dea47 and observed on 2026-10-27; the card's commands still name the KwaiVGI repositories. Nothing was run here; the changelogs, the Animals model instructions and the community ComfyUI nodes were not inspected. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/KlingTeam/LivePortrait/raw/82a4fa6735ca58432b6ce39301b4b9ee066dea47/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 4736:4804, 6101:6438, 6679:6882, 7423:7732, 8146:8512, 9100:9459, 10211:10471; exact excerpt SHA256 c241827a0fcf7bda66fc025121f280e90b8d67a841138b0e9cd6602f09f55ea4,760859201b02f70a8439d1c203d14cc4a8d7ea974f4ccd1fe8640634c9de9700,16f728df838af12feb192570ba989d44a63a4a6b3d8da739"
    }
  ]
}