This library implements the end-to-end facial synthesis model described in this paper.
The models provided are checked out using git LFS. You can install git LFS by following these instructions.
To install the library do:
$ pip install .
To create the animations you will need to instantiate the VideoAnimator class. Then you provide an image and audio clip (or the paths to the files) and a video will be produced.
The model has been trained on the GRID, TCD-TIMIT, CREMA-D and LRW datasets. The default model is GRID. To load another pretrained model simply instantiate the VideoAnimator with the following arguments:
import sda
va = sda.VideoAnimator(gpu=0, model_path="crema")# Instantiate the animator
The models that are currently uploaded are:
- grid
- timit
- crema
- lrw
import sda
va = sda.VideoAnimator(gpu=0)# Instantiate the animator
vid, aud = va("example/image.bmp", "example/audio.wav")
import sda
import scipy.io.wavfile as wav
from PIL import Image
va = sda.VideoAnimator(gpu=0)# Instantiate the animator
fs, audio_clip = wav.read("example/audio.wav")
still_frame = Image.open("example/image.bmp")
vid, aud = va(frame, audio_clip, fs=fs)
va.save_video(vid, aud, "generated.mp4")
The encoders for audio and video are made available so that they can be used to produce features for classification tasks.
The Audio Encoder (which is made of Audio-Frame encoder and RNN) is provided along with a dictionary which has information such as the feature length (in seconds) required by the Audio Frame encoder and the overlap between audio frames.
import sda
encoder, info = sda.get_audio_feature_extractor(gpu=0)