It's designed to use Tensorflow 2.X on python (3.7), using cuda 10.1 and cudnn 7.6.5.
Run conda create -n environment.yml to create a conda environment that has the needed dependencies.
Tested with Tensorflow 2.0.0, Python 3.7.9, Ubuntu 14.04.
Our method relies on several pretrained networks. Some are needed only for training and some also for inference. Download according to your intention.
Put all downloaded files/directories under a single directory, which will be the baseline path for all pretrained networks.
| Name | Training | Inference | Description |
|---|---|---|---|
| FFHQ StyleGAN 256x256 | ✔️ | ✔️ | StyleGAN model pretrained on FFHQ with 256x256 resolution. Converted using StyleGAN-Tensorflow2 |
| FFHQ StyleGAN 1024x1024 | ✔️ | ✔️ | StyleGAN model pretrained on FFHQ with 1024x1024 resolution. Converted using StyleGAN-Tensorflow2 |
| VGGFace2 | ✔️ | ✔️ | Pretrained VGGFace2 model taken from WeidiXie. |
| dlib landmarks model | ✔️ | dlib landmarks model, used to align images. | |
| ArcFace | ✔️ | Pretrained ArcFace model taken from dmonterom. | |
| Face & Landmarks Detection | ✔️ | Pretrained face detection and differentiable facial landmarks detection from 610265158. |
To try out our method with other checkpoints of StyleGAN, first obtain a trained StyleGAN pkl file using the original StyleGAN repository
Next, convert it to Tensorflow-2.0 using this repository.