mlx-community/MuseTalk-1.5-q4

MuseTalk 1.5 — MLX (q4)

Apple-MLX port of MuseTalk 1.5 (TMElyralab / Tencent Music) — realtime, high-quality lip-sync via single-step latent-space inpainting (not diffusion). Runs natively on Apple Silicon. MIT-licensed, commercial use OK.

This variant: UNet Linears quantized to int4 (group_size 64); VAE + Whisper encoder kept fp16. Per-pass cosine vs fp16 = 0.99985. Decoded-face error vs the PyTorch reference: mean |Δ| ≈ 2.74/255.

Components (all in this repo, self-contained, torch-free)

File	What
unet.safetensors	SD1.x UNet2DConditionModel (in=8, out=4, cross_attn=384), single-step t=0
vae.safetensors	sd-vae-ft-mse AutoencoderKL (fp16)
whisper_encoder.safetensors	whisper-tiny audio encoder (fp16)
config.json	dtype / quantization / scaling factor

Performance

Realtime on an M-series GPU: ~34 generated 256² faces/sec at batch 8 (>25 fps video rate), ~7 GB peak. fp16 inference.

Usage

from musetalk_mlx.pipeline_mlx import MuseTalkPipeline
pipe = MuseTalkPipeline.from_pretrained_mlx("MuseTalk-1.5-MLX-q4")
# crop_bgr: a 256x256 face crop; chunks: (N,50,384) whisper audio features
latents = pipe.get_latents_for_unet(crop_bgr)
faces = pipe.generate_faces(latents, audio_chunks)   # BGR uint8 lip-synced faces

Face detection / cropping / paste-back blending use the upstream (MuseTalk) CPU preprocessing.

Parity (vs PyTorch, cpu fp32)

VAE encode 1.7e-5 · decode 3.4e-5 · UNet forward 1.4e-6 · whisper encoder 1.6e-5 · face-level e2e recon ≤ 2/255.

License

MIT (mirrors upstream MuseTalk). Dependency models keep their own permissive licenses. Port by MVS Collective (xocialize-code).