This code builds on StyleGAN2-ADA and was inspired by Apocalyptify. It blends the orininal StyleGAN2 and the same network but fine-tuned on another dataset (e.g. aliens and zombies). You can train the network on your own dataset.
Go to my website and project yourself into the zombie/alien space. Don't abuse it: the backend runs on my modest GPU.
For quick start, run the Colab notebook hassle-free. If you have serious GPU compute, use it. Clone this repo, make sure you have the right version dependencies, download the data (see the top of the Colab notebook). Open an issue if something doesn't work.
You can project a real target image (left) to the latent space and then back-project that latent represenation to pixel space using a different version of StyleGAN2, e.g. the original (middle) or fine-tuned on Pixar characters (right).
You can also fine-tune StyleGAN2 on a dataset of aliens and zombies, transplant some of its layers to the original StyleGAN2 trained on FFHQ and get some creepy pictures!
Watch on YouTube:
Thanks to Doron Adler.