They allude to it but do not fully explain. I am assuming they are also using the vision model to describe the input photo. I would not have initially thought that you given the text description of an image you could recreate it well enough for fun avatars...it makes sense but I did not initially think about it.