The more specific the prompt, the more uniform the output; the less specific the prompt, the less uniform the output. The models are large enough that it's little different than offering the same prompt to human artists. The idea was to mimic the input--output flow of human interpretation using massive back catalogs of interpretations made by humans. Humans are, essentially, stable diffusion engines which take everything they've read, seen, and learned, and apply it to the task at hand. Some are better than others. Some - rather few - have the abiliy to create exceptionally refined and nuanced versions - we get the classics. Even fewer can extrapolate from the existing human data set, or set it aside to produce results in unexpected ways, and we get the avante-garde...which then gets folded into the "data-set" for future humans.

To quote Mel Brooks (or his writing team):

Dole Office Clerk : Occupation?

Comicus : Stand-up philosopher.

Dole Office Clerk : What?

Comicus : Stand-up philosopher. I coalesce the vapors of human experience into a viable and meaningful comprehension.

Dole Office Clerk : Oh, a bullshit artist!