woctezuma / steam-stylegan2-ada-pytorch

Train a StyleGAN2-ADA (PyTorch) model on Colab to generate Steam banners.

Steam StyleGAN2-ADA (PyTorch)

The goal of this repository is to capture the distribution of Steam banners with a StyleGAN2-ADA model (PyTorch).

_{Progression videos of the projection of Steam banners (Dog Trainer and My UnReal pet) with a network pre-trained by Nvidia on the LSUN DOG dataset.}

Data

The Steam-OneFace dataset, obtained with the retinaface face detection module:

• consists of 2.5k Steam banners (256x256 resolution) which should all feature exactly one face,
• can be downloaded (74 MB) with the following commands:

!gdown --id 1-0Nk7H6Cn3Nt60EdHG_NWSA8ohi2oBqr
!tar xf steam-oneface-lr_with_retinaface.tar.gz

Usage

• Run training.ipynb to train a model from scratch,
• Run image_sampling.ipynb to generate images with a trained model,
• To automatically resume training from the latest checkpoint, use my fork of StyleGAN2-ADA (PyTorch).

References

• Karras, Tero, et al. Training generative adversarial networks with limited data. NeurIPS 2020,
• Official implementation (PyTorch),
• Application to Steam banners using the TensorFlow implementation.