Which of these faces can you identify as real? Well, that’s a strategy question, just as none of the faces belong to real people. Every single human in the title photo is produced by a neural network without input or correction.
Researchers at AI firm Nvidia have modified the Generalized Advertising Network (GAN) technology, which can make fake human faces so bright that real people are unable to distinguish them from a real image. The team trained their algorithm to learn patterns from more than 70,000 photos of real human faces taken from Flickr. AI used these images to learn about different aspects of the human face, such as hair color, skin tone, posture, and even subtle variations such as freckles, skin pores, and straw.
It has learned to distinguish aspects like photo color scheme and style, important to make the generated photos realistic. It was after learning facial features and being able to distinguish them that they were taught to combine facial features to create completely new faces, practically indistinguishable from ordinary photographs, only the people in the photos do not exclude this fact except actually exist.
As can be seen in these photos generated through research, faces can be assembled from specific inputs. The researchers said in a video, “We’ve come up with a new generator that automatically learns to differentiate between different aspects of an image without human supervision.” “The new architecture leads to an automatically learned, unsupervised separation of high-level attributes.”
This video shows how the code is easily smoothed, morphing into other completely different (not even real) faces, while seemingly human faces look absolutely unrealistic. Earlier this year, Nvidia created another bizarre piece of software that had some parts of the photographs destroyed or damaged (or, for research, deleted on purpose).