CoolFace
Apppublic

augustolucasg/audio2face-3d

sourceHugging Faceotherupdated 5mo agoView on Hugging Face
0likes
App README

Audio2Face-3D v3.0

Streaming-stateless ONNX wrapper around nvidia/Audio2Face-3D-v3.0.

Upload an audio clip, pick an actor (Claire/James/Mark) and emotion, and download a .npz containing the predicted vertex animation at 30 fps:

python
import numpy as np
data = np.load("audio2face_prediction.npz")
data["skin"]    # (n_frames, 24002, 3) — face skin vertex positions
data["tongue"]  # (n_frames,  5602, 3) — tongue vertex positions
data["jaw"]     # (n_frames, 15)       — jaw rig channels
data["eyes"]    # (n_frames, 4)        — eye rotation channels
data["fps"]     # [30]

The model card lists 10 named emotions plus an unconditioned neutral. A 52-pose blendshape solver (ARKit-compatible) is not applied here — this Space returns raw vertex predictions.

API

python
from gradio_client import Client
client = Client("devsomosahub/audio2face-3d")
out, summary = client.predict(
    audio="path/to/voice.wav",
    actor="Claire",
    emotion="neutral",
    seed=0,
    api_name="/predict",
)