There exist systems where you can describe not only initial scenes but successive textual modifications to existing images and furthermore variations aren't random. Successive selections are a way to zero in on a concept.

You are about a year behind the state of the art.

AKA tell me you haven't spent time with diffusion models, without telling it :)

I actually did figure out what works and what doesn't in real artistic use. Which is the entire point of the article in OP which nobody seem to have read - text doesn't work well beyond the basic use or amateur play, regardless of it being the initial prompt or editing; you need sketching and references (and actual skill) to do real work. I don't think anybody's using available methods of textual modifications for anything complex - they are cumbersome and unreliable, even worse than textual prompts. In fact, I haven't seen anyone using them at all.

Besides the implementation details, natural language just doesn't have enough semantic density and precision to give artistic directions, even for a human or AGI. That's a fundamental limitation. Higher order guidance, style transfer, and compositing is how it's done.