But the second point here is also wrong: the whole reason these models are interesting is because they can generate things they haven't seen before - the corpus of knowledge represents some type of abstract understanding of how words relate to things that theoretically does encode a little bit of the mechanisms behind it.
For example, it theoretically should be able to reconstruct human like poses it has never seen before provided it has examples of what humans look like and something which transposes to an approximate value - an obvious example in the context of the original question would be building photorealistic versions of a sketched concept (since somewhere in it's model is an axis which traces from "artistic depiction of a human" to "photograph of a human" in terms of style content).
Of course, most people aren't very good at drawing realistic human poses - it's a learned skill. But the magic of deep learning is really that eventually it doesn't need to be - we would hopefully be able to train a model which can be easily copied and distributed which represents the skill, and SD is a big step in that direction (whether it's a local maxima remains to be seen - it's dramatic, but is it versatile?)