I’ve heard similar things from the same individual about the progress of results. Indeed, they said the things I mentioned in my first comment a few months ago.

I’ve heard similar things from the same individual about the progress of results. Indeed, they said the things I mentioned in my first comment a few months ago. Still, it sounds like a current problem is that it’s difficult to reliably get visually nearly-identical output with incremental changes. That’s one area where LLM prompt templating can be useful. This lack of reliability is often mentioned about these generative technologies. It seems to me that “how do I get the same T-Rex with lasers shooting out of its eyes but different color lasers?” and “how do I get the same JSON structure in my output and always only the JSON with no other text?” are roughly similar problems in “prompt engineering”.