check out c/botart@dbzer0.com, which does a lot of stress-testing generative AI, either its capacity to draw stuff that isn't NSFL (it likes scary hands and blobs of people and misshapen animals) or its ability to parse out individual elements and their relationships to each other. Or even literal objects from cultural references and figures of speech.

So in order to get desired results of specific, intended output, you need not only to run the same prompt multiple times and curate but also focus on specific elements and run those through with additional prompts. Lather, rinse, repeat.

So it's skilled work, not merely a few words into a black box