Yaohui Wang, Piotr Bilinski, Francois Bremond and Antitza Dantcheva

• Python 3.6
• cuda 9.2
• cudnn 7.1
• PyTorch 1.4+
• scikit-video
• tensoboard
• moviepy
• PyAV

You can download the original MUG datest from https://mug.ee.auth.gr/fed/ and use https://github.com/1adrianb/face-alignment to crop face regions. You can also download our preprocessed version from here and save it under $DATA_PATH.

Download the pretrained model on MUG from here and put it under ./pretrained.

Generate videos and save them under ./demos/mug

python demo.py --dataset mug --model_path ./pretrained/mug.pth

python train.py --data_path $DATA_PATH

To compute FID, please ref to G3AN.

If you find this code useful for your research, please consider citing our paper:

@InProceedings{WANG_2020_WACV,
  author = {WANG, Yaohui and Bilinski, Piotr and Bremond, Francois and Dantcheva, Antitza},
  title = {ImaGINator: Conditional Spatio-Temporal GAN for Video Generation},
  booktitle = {The IEEE Winter Conference on Applications of Computer Vision (WACV)},
  month = {March},
  year = {2020}
}