Comment by rcme

I bet they use CLIP to caption the image and feed the text of the caption into GPT, but that's just a guess.