Yeah, I agree, I think specificity is key. If you leave an LLM, or any AI model, to its own devices, it gravitates to the figurative "mean of its training data" which tends to be pretty sterile.

I actually have some examples of my own comics that I used to put gpt-image-2 through its paces, and the results were surprisingly good as well:

https://mordenstar.com/other/gpt-2-comics