Think about the problem in an end-to-end fashion: the user has an idea of what sort of image they want, they just need an interface to tell the machine. A combi

Think about the problem in an end-to-end fashion: the user has an idea of what sort of image they want, they just need an interface to tell the machine. A combination of natural language plus optional image/video input is probably the most intuitive interface we can provide (at least until we've made far more progress on reading brain signals more directly). How exactly we get there, by adding layers like on top like language models, or adding layers below like what you described, doesn't seem like such a fundamental difference. It's engineering, you try different approaches, vary your parameters and see what works best. And from the onset, natural language does seem like a good candidate for encoding nuances like "make it pink, but not cheesy" or "has the vibes of a 50's Soviet propaganda poster, but with friendlier colors".