I Fed an AI 5000 Halloween Masks and All I Got Was Nightmares.

Over the past few years neural networks have started to achieve things once thought unimaginable. Just last week a neural network generated a piece of art that sold for $432k. Yes, the future is weird. Neural networks come in all flavors but one in particular has recently fascinated me, Generative Adversarial Networks. GANs are special because they go beyond predicting things and actually create things.

GANs work by pitting two networks against each other resulting in an automated, self-improving feedback loop. It basically teaches itself how to make itself better. The longer it runs, the “smarter” it gets 🤯.

After collecting thousands of Halloween masks, we fed them to a GANpowered by Google’s Tensorflow and set it loose to see if it could create some never-before-seen Halloween masks. The generated masks start off as noise but slowly become more and more realistic over hours and hours of training.

The generator network tries to trick the discriminator network by pawning off fake Halloween masks it came up with as real ones. The discriminator will take a look at them and tell the generator how well it did while providing some helpful tips on where to improve. The generator goes back and makes some tweaks and tries again. This process repeats over and over, and the results usually get better and better.

After leaving the network running overnight things started to get more interesting. The following morning I opened up the directory containing the generator’s creations and then immediately closed it, went home, trembled, cried, and retired from computers forever. After 24 hours of training the AI had learned to ruin a perfectly good night of sleep.

Enter at Your Own Peril

Curious to see everything the AI generated? Feel free to browse the raw output here. Some of my favorites are shown above but there are even more if you are brave enough. Between now and Halloween I’ll be updating both of those directories with the latest generated masks so keep checking back between your sleepless nights.

Putting GANs to Work

So GANs are fun and all but they have some pretty unique qualities that are already changing how we think about and understand creativity. While only a technical constraint, the relatively low resolution of the generated images can be a good thing. Just like seeing things in clouds, your unique interpretation of the machine’s generation can result in sparking ideas you may have never considered.

Fashion designers could generate a nearly infinite set of never-before-seen patterns. Product designers could generate ideas for new form factors. Advertisers could generate one-off custom targeted Ads based on someone’s Instagram favorites.

I think GANs will become a part of many creatives’ toolsets over the coming years. Think of them as automated thought starters that can knock you out of a creative rut. A properly trained GAN paired with a human will usher in a new era of creativity, which shouldn’t be too scary…

Do you wanna see more fun tech stuff like this? Head on over to our Garage or just look down 👇.

https://www.youtube.com/watch?list=PL2LmbqRU6aP-HYGKhFkin5Hjwd21vFYF9&v=-i7HMPpxB-Y