What's really striking to me is that the reconstructed images don't have a direct correspondence to the prompts in terms of layouts or other basic visual arrangements but are undoubtedly "about the same thing" (at least in the examples provided)