Fright night just got scarier. A neural network trained on 5000 images of Halloween masks has dreamt up its own versions, and they鈥檙e more terrifying than the real thing.
鈥淲hat鈥檚 so scary or unsettling about it is that it鈥檚 not so detailed that it shows you everything. It leaves just enough open for your imagination to connect the dots,鈥� says Matt Reed at Red Pepper, a creative agency.
To create these freaky masks, Reed used a type of artificial intelligence called a generative adversarial network (GAN), which learns to recognise certain traits in training data 鈥� in this case, photos of Halloween masks 鈥� and then creates images that might fit into the original set. In theory, after many iterations, the network gets better at producing new images.
Advertisement
鈥淎 lot of them were like abstract art pieces,鈥� says Reed. 鈥淪ome are just black blobs with hollowed out eyes. One of them looks like a Cheeto.鈥�
But some would indeed fit a horror inducing Halloween costume. Reed says he started by downloading thousands of images by searching for Halloween masks, scary masks, and funny masks. He then left his computer running overnight to dream up these horrors.
鈥淚鈥檇 wake up in the morning and the results were terrible. I tested it all throughout September and nearly gave up last week, but I finally got some interesting results,鈥� he says.
You can see the resemblance in some of the GAN images to masks of US Presidents sometimes worn by bank robbers in movies. One of the AI-generated masks looks like a sock puppet smoking a cigarette. Many look like traditionally scary masks 鈥� zombies or monsters 鈥� that have been put through a microwave. They all look like a Rorschach test had a nightmare.
