StyleGAN2

This article explores changes made in StyleGAN2 such as weight demodulation, path length regularization and removing progressive growing!

The first version of the StyleGAN architecture yielded incredibly impressive results on the facial image dataset known as Flicker-Faces-HQ (FFHQ). The most impressive characteristic of these results, compared to early iterations of GANs such as Conditional GANs or DCGANs, is the high resolution (1024²) of the generated images. In addition to resolution, GANs are compared along…