I'm surprised that the results seem much better (more detailed and sometimes closer to the original prompt) than the regular CLIP interrogation (at least based on my limited experimentation).

But as you say, even so, it still has little in common with the original prompt.