FasterLivePortrait-MLX-weights

FasterLivePortrait-MLX Weights

Converted MLX .npz runtime weights for FasterLivePortrait-MLX.

These files are converted from permissively licensed source checkpoints:

• KlingTeam/LivePortrait, MIT
• warmshao/FasterLivePortrait, MIT
• jdh-algo/JoyVASA, MIT
• TencentGameMate/chinese-hubert-base, MIT

Included

• Human LivePortrait core MLX weights
• Human landmark MLX weights
• Human stitching / eye / lip retargeting MLX weights
• Animal LivePortrait v1.1 core MLX weights
• JoyVASA MLX audio-to-motion weights for the configured Chinese HuBERT path
• JoyVASA motion template runtime asset

Not Included

This repository intentionally does not include XPose. XPose is used only for animal landmark detection in FasterLivePortrait-MLX, and its upstream license is restricted to non-commercial research use.

This repository also does not include the original JoyVASA PyTorch checkpoint or the original Transformers HuBERT directory. Those are conversion inputs only.

This repository also does not include the MediaPipe Face Landmarker task model; download it from Google's MediaPipe model URL as documented in the project README.

Use

uv run python scripts/download_mlx_weights.py --repo-id ivanfioravanti/FasterLivePortrait-MLX-weights

Conversion

The weights were produced with:

uv run --group convert python scripts/export_mlx_weights.py --include-animal

JoyVASA weights were produced with:

uv run --group convert python scripts/export_mlx_weights.py --include-joyvasa

Converted tensors are derivative model weights and inherit the obligations of the original model licenses.