Can you spot which of these faces are real? Well, it’s a trick question, as absolutely none of these faces belong to real human beings. Every single one in the header photo is generated by a neural network, without human input or refinement.
Researchers at AI company Nvidia have modified Generative Adversarial Network (GAN) technology, which can create fake human faces so photorealistic that actual humans are unable to distinguish them from a genuine photograph.
The team trained their algorithm to learn patterns from over 70,000 photos of real human faces taken from Flickr. The AI used these images to learn about different aspects of people’s faces, e.g. hair color, skin tone, pose, and even subtle variations such as freckles, skin pores, stubble, etc.
It also learned to distinguish aspects like the color scheme and style of the photos, important for making the generated photos look realistic.
After it had learned the characteristics of faces, and was able to separate them out, it was then taught to combine facial features to create entirely new faces, virtually indistinguishable from normal photographs, except for the fact that the people in the photos don’t actually exist.
The faces can be combined from specific inputs, as shown in these photos generated by the study.
“We came up with a new generator that automatically learns to separate different aspects of the images without any human supervision,” the researchers said in a video.
“The new architecture leads to an automatically learned, unsupervised separation of high-level attributes.”
This video shows how freakily smooth the code is, with seemingly human faces morphing into other completely different (also not real) faces without ever looking even briefly unrealistic.
Earlier this year, Nvidia created another freaky bit of software that will recreate parts of photographs that have been destroyed or damaged (or, for the study, erased on purpose).