Human Faces Created By Neural Networks Are A Little Too Realistic

Collaborating with a neural network to draw pictures of humans can be a lot of fun–one drawing tool, featured recently on Co.Design, has had readers submitting their masterworks all week. But while the results of those experiments often look more like fleshy, hairy blobs than actual people, artist Mike Tyka is working on neural net portraiture that is almost too lifelike.

While Tyka’s so-called “portraits of imaginary people” wouldn’t necessarily be mistaken for photographs, they achieve something unusual: They’re in focus. That’s something that neural net art isn’t typically known for. Tyka’s goal for the project, which is still ongoing, is high resolution texture.

Tyka is using generative adversarial networks, or GANs, a type of neural network that combines two different networks working together to create more realistic images–with one network “judging” the realism of the images produced by the other (a more thorough explanation is here). Resolution is a major challenge with these images; Tyka writes that for him to increase the resolution and make the images good enough for printing, he would need datasets of 4K images on which to train the system.

So far though, his results show great potential. In one, a teenage girl sneers at the viewer, her eyes looking in slightly different directions. In another, a pale woman smiles calmly as she looks into the distance amidst an eerily white backdrop. In a third, a woman looks at the viewer, her gaze particularly human. One man’s face has different textures in different places, making him look like a cyborg–which, one some sense, he is. One is a dead ringer for Michael Jackson.

While areas of the portraits look glitchy when you zoom in, others almost look like oil paintings. Perhaps Tyka will one day be able generate images that are so detailed, they look like centuries-old masterpieces.KS