I have been lowkey obsessed with snapchat for a while now and decided to scrape some images from instagram to see what a neural network made of them. I expected the Generative Adversarial Network to meld together the Augmented Reality components with the faces to create something nightmarish. What I didn’t expect was the GAN to detect the commonality in the images and erase the components all together!

Generative Adversarial Network grid at mode collapse

Working first with a DCGAN I trained until it reached mode collapse. I really enjoy watching neural networks collapse and the converging grids made for some really interesting works. While training I also used the face picker in snapchat to see if it would recognize any of the faces and was shocked when it separated them all out and used them all, with horrifying and hilarious results! I documented this progress on Twitter in a lengthy thread as I explored everything snapchat.

A grid of faces from early GAN snapchat training

From there I used style transfer on some images to upsample them. For others I used resizing GANs and upsampling algorithms to bring the works up to print size from 128px. All of this ensures large enough prints for the art that will be on display. When the exhibition goes live I will publish the art and snapcodes on my website so that everyone can join in the fun even remotely!

If you have fun with this project please share it around and if you use the lenses tag me in or use #snapchatgan so I can see your wonderful works!

A grid of style transferred faces