SAGAN

When is a face a face? How much do we need to see before we can recognize who it is? And when does an image become a portrait?

SAGAN is a series of portraits generated by the Machine Learning algorithm Self-Attention Generative Adversarial Networks. For each portrait, 800 photographs are taken of the subject. The photographs are fed into a neural network, and the network is tasked with producing a new “photograph” that is indistinguishable from any of the original 800.

To accomplish this task, the neural network employs an iterative process where the image evolves over the course of approximately 80,000 training cycles. In the course of this training, the neural network generates 8,000 attempts at a photograph. Of those 8,000 attempts, 16 are selected by the artist to form the final portrait.

Due to Covid restrictions, the artist was not able to get close enough to photograph the actual subjects. Instead, the subjects were given a remote control for the camera and they took the 800 photographs themselves. Thus the portraits in SAGAN are actually self-portraits.