If there's interest, I will surely automate it.
This is humorously wise. I feel like I just read a Chinese proverb on how start a software business. Very few programmers could hold themselves back from going further on their first iteration.
Similar to browser games that give you a new Tower Defense map every 24h, or Wordle and all its variations, or .. "A roguelike a day keeps the Balrog away" (that's the tagline, forgot the actual name).
Context windows cause things to break down very quickly. And overfitting of brief and to the point RLHF responses, as well as summary datasets makes coaxing many of the models into slowly narrating things, as one would in a novel is challenging.
Agent-based approach is one way to tackle this, but in that case maintaining narrative style coherency is also somewhat difficult to get right.
Come to think of it, it's actually pretty hard to automate any sort of advanced artwork generation with LLMs. Once you go beyond your average prooompting, you have to put in a lot of work.
