With fake images, we can customize the exact properties of the objects in the image. That way, we can see if the algorithms we’re training can uncover those properties correctly.