These Pigeons Don’t Exist

I started to play around with RunwayML recently, and one of the most compelling features is the ability to train your own models. You essentially start with a pre-trained styleGAN model and then upload your own dataset and then set the number of steps the algorithm will take to train your model.

RunwayML user interface

This model is trained on a little over 2,500 photos of pigeons. The quickest way I found to – say – harvest the sample material was to head over to flickr.com and do an image search. In order to download the large number of images, I left Charles Proxy running in the background, and then saved the session to my desktop. Since the session comes right from a CDN all the images are nested in various folders, so to flatten them into one to prepare for upload you can simply run a command such as: find ~/Desktop/MYFOLDER -mindepth 2 -type f -exec mv -i '{}' ~/Desktop/MYFOLDER ';'

The results aren’t that great and are based on two attempts of using StyleGAN 1 at 7500 steps and styleGAN 2 at 3000 steps. Nevertheless, it was an interesting introduction to playing with RunwayML.

Leave a Reply

Your email address will not be published. Required fields are marked *