AI Generated Images

Someone wrote a scraper that converts Worm and Ward to ebooks. I used it to read them offline.

https://github.com/domenic/worm-scraper

It'd be kind of cool if there were another project that incorporated illustrations to the generated ebook.

EDIT: And maybe somewhere that could incorporate patches to fix typos and the like. I remember seeing some of those when reading them.

EDIT2: Looks like the scraper already has a collection of typo fixes, so the functionality is probably already present; just needs PRs for the issues I noticed. Damn, wish I'd made note of them when reading through:

This project makes a lot of fixups to the original text, mostly around typos, punctuation, capitalization, and consistency.

EDIT3: One problem with using generative AI to do illustrations -- especially across multiple LLM models -- is that it'd be hard to get a consistent appearance for a character.

• Some of that can, in my experience, be masked, by using some artist's style, because as long as an LLM is reasonably trained on that artist's style, it creates some level of consistency in general appearance. That doesn't help with specific attributes.

• Won't work for the proprietary LLM models, but maybe it's possible to generate multiple images of a character and use something like Stable Diffusion's clip-interrogator to "converge" on prompt terms that reasonably-consistently generate a character. Maybe train a LoRA on a corpus of "bad" and "good" examples of a character as consistency goes, attach terms to those, like "badworm" and "goodworm" or something.

• It might also be possible to use some sort of "sketch generator" that generates a rough outline of the character, pose that in a 3d modeler, and then hand it off to LLMs; for Stable Diffusion, at least, that's existing functionality, though I don't know about all of the other LLMs. Then one just has to use a modeler to really-quickly pose a premade model for the characters, feed it in as input to the LLM. I suspect that artists are, in general, going to develop techniques like this further for working with LLMs. Here's someone using existing Stable Diffusion tools, though I have no idea how well that would work as a general solution.

"A shifting mass of hooks, blades and other weapons in the shape a a dire wolf. it is evil. it is positioned in a nighttime city street facing down a bearded superhero in sleak dark blue power armor and helmet wielding a large long electric halberd."