There are 2 AIs at work here: one that has been trained to understand language and the other that has been trained with visual datasets. It starts with perlin noise and the images are not created line by line but they gradually emerge, almost through a fog