Does anyone have any context as to how the image understanding works? From what I can gather they are simply using separate text-summarization step to generate

Does anyone have any context as to how the image understanding works? From what I can gather they are simply using separate text-summarization step to generate some text like "and now we have an image of chicken nuggets" that it then feeds to the text-only network, but I wouldn't be surprised if there is some dialog I'm missing between the previous context and the image understanding mechanism.