Unofficial implementation of StyleGAN 2 using TensorFlow 2.0.
Original paper: Analyzing and Improving the Image Quality of StyleGAN
Arxiv: https://arxiv.org/abs/1912.04958
This implementation includes all improvements from StyleGAN to StyleGAN2, including:
Modulated/Demodulated Convolution, Skip block Generator, ResNet Discriminator, No Growth,
Lazy Regularization, Path Length Regularization, and can include larger networks (by adjusting the cha variable).
Trained on Landscapes for 3.48 million images (290k steps, batch size 12, channel coefficient 24): To clarify, 3.48 million images were shown to the Discriminator, but the dataset consists of only ~14k images. Thus, of those 3.48 million images, most are repeats of already seen images.
Mixing Styles:
Please ensure you have created the following folders:
- /Models/
- /Results/
- /data/
Additionally, please ensure that your folder with images is in /data/ and changed at the top of stylegan.py.
For pretrained models, download the pretrained models from here
Then, adjust the main code at the bottom to use model.load(model_num), where model_num = the number from the downloaded model.