Turning my Dog into a Person with TensorFlow and StyleGAN2-ADA on Windows

StyleGAN: The Gateway Drug

Earlier this year, I started looking into something called StyleGAN which is a machine-learning project that was made public by Nvidia in early 2019.

stylegan-teaser These people are not real – they were generated by StyleGAN

This naturally led me to find stylegan2 which added improvements to image quality, and then to the even newer stylegan2-ada which adds even more improvements when working with smaller data-sets.

Getting Started (on Windows)

One of the coolest things about ML projects is that many are hosted online using google colab, meaning that you don't need to install or configure anything on your computer. And that's fun and all, but I wanted to get my hands dirty and make use of some idle GPUs.


First, a couple things need to be installed:

Experiment: Turn my dog into a person

The first thing I thought to do, mainly because so many data sets are based on human faces, was to see what would happen if I provided an image of my dog as the input:

Dog to Person (stylegan2-ada + ffhq)

Ok, that was actually cooler than I expected with very little work, so I had to repeat the process using more data sets:

Dog to MET Painting (stylegan2-ada + met)

Dog to Dog (stylegan2-ada + dog data set)

Dog to Cat (stylegan2-ada + cat data set)

Dog to Cat (stylegan2-ada + wild cat data set)