Cosmopolitan leaked how the process looks like: https://www.cosmopolitan.com/lifestyle/a40314356/dall-e-2-ar...
https://hmg-h-cdn.hearstapps.com/videos/creation-loop-165578...
It takes a lot of editing to get something good and not pure nonsense. Its a fancy Content-Aware Fill on steroids. It drops objects from the prompt on canvas and tries to fill the rest while minimizing error.