Neural nets have been trained on vast databases of all kinds of images–dogs and cats, cities, Pokemon, penises. When you try to train one on pictures of people, things get creepy–and now you can play around with such a neural net yourself, transforming your idle doodles into fleshy, leering blobs.
To demonstrate how machine learning works, one Dutch radio station has trained a neural network on hundreds of drawings and images of one of its reporters. When you draw a face, the program translates your sketch into what’s supposed to be a photorealistic image, based on its database of drawings of reporter Lara Rense. But the results are generally horrifying, mixing fleshy shapes with dark hair to create monstrous images that resemble a distorted human face. Or balls (see image eight in our slideshow).
The radio station used the Pix2Pix neural network. It can be trained on thousands of photos to generate a photorealistic image from a sketch.
Fair warning, the program is intensively addictive. The Co.Design team nearly broke Slack uploading sketches of everything from Trump to a bicycle to Magic Mike. Try it yourself here–and send your best drawings to CoDTips@fastcompany.com. We’ll publish the best ones.