August 12, 2019 — ongoing

A Generative Adversarial Network, or GAN, is a model trained on image data — in this case, pictures of people pulled from By looking at millions of these images, the software can create near-perfect portraits of human faces for people that don’t exist. (See for yourself at the appropriately titled

The archive collects examples of a specific aesthetic: a GAN-generated glitch that results from the model “learning” from images in which a companion has been cropped out of the frame. Because the GAN is trained to focus on the faces that appear in the center of the image, it doesn’t pay as much attention to these cropped-out digital companions.

So they appear here, photo bombs by uncanny demons, rendered as if they were nothing but background noise. The GAN is creating terrifying distortions of human beings who are literally deformed composites of every human being who has ever been cropped out of a photo. I imagine them, transformed by their removal, struggling to make their way back into the center of the frame, like scorned and hungry ghosts desperate to become complete at the center of the camera’s eye.