I am a huge fan of AI generated art as a way to bang out concepts. I realize it isn’t art, it steals inspiration from real human artists, and that the models that allow it to do its stuff were harvested without artist consent. I don’t know how to right those wrongs, or if abstaining from using tools based on this truth would rectify the situation. But as a vehicle for generating zany scenarios that a real artist would go “uh.. what?”, an AI will generate its best efforts despite the fact that even its own model building screams “this concept does not compute!”

One of the final game segments of Book 1 is a racing game, inspired by Outrun with a generous sprinkling of “dodge this!” hazard thrown at you as you streak towards the finish line. Only instead of a sports car, you’re driving a giant warbarge. Crewed by orcs. Who are intent on slamming headlong into the castle in the distance. Which is protected by battle mages throwing spells at you and trying to blow you off course, pick off your crew and wreck your plans. At some point, I’ll want to show this concept to a crew of game designers who will undoubtedly say “this concept does not compute.” Hence the need to tell Stable Diffusion about my plan to create a hit point based hazard racer with a orc raft bearing down on a castle. AI art may staple extra fingers on knees and armpits, but one thing it does not do is judge you. In any obvious way, that is.

And… it pulled it off. Yes, it had its biases; boats go on the water, so without rolling up a newspaper and hitting it, every picture I asked of it kept putting boats into the water and not on paved roads. Which, in 99.999% of requests, is a safe bet for it. The value of using Stable Diffusion is that you can indeed bop it with a magazine and say “no. No water. Bad!” and it will get it. Some percentage of the time, and the more successful images you can re-feed back into it, the higher that percentage goes. Eventually, most of the outputs were boats on land, which is exactly what I was hoping to show.

Is it high art? Hell it isn’t even art. But its concept art, and that is what was needed at this point.

The other aspect I really admire about asking for AI help is its ability to come up with something I didn’t think about. If you have an exact look for a character in mind, then it is a very iffy tool. If, on the other hand, you are ok with something inspired by a look, or will allow for some novel variation from your prompt, then it can generate some surprising results. I had two act villains I needed fleshing out, and asked StableDiffusion for some guidance on how to make the look, so I could write a little more flowingly about their characters.

Director Hunfried Stelen is the casino boss of the Lapis Doily. He also happens to look an awful lot like Edward Snowden. That immediately clicked with the AI, who clearly has source material to draw from, and then just created a weird casino pit boss scene for him to sit on. Because I said casino, very often he was sitting at a green felted poker table; I had to reject a lot of pretty decent looking character arts because the backdrop was lousy. StableDiffusion (and probably others; I am just mostly satisfied with my current tool) lets you specify the classifier-free guidance scale (CFG) you want. Which in practical terms is a scale between “do whatever, just make sure its something like what I said” to “listen to me, you calamitous creative contraption!” I mention this because it was utterly unneeded in this particular case; the only thing I had to do was generate enough images that I got one where the desk wasn’t something straight out of an MC Escher painting.

Madam Geirtrud the kitchen witch, on the other hand, was hard. Getting a witch is easy; getting a witch that doesn’t look like she’s wearing a Spirit Halloween costume to a sorority party is hard. Asking it for a cauldron full of gingerbread dough was surprisingly simple despite that not actually being a thing there are many pictures to steal for source material. The hard part, honestly, was the spiders. Or more specifically the lack thereof.

It seems there are things that AI likes to see together. If you have a witch, she most likely is a sexy witch. If you have spider webs, you most likely have spiders on them. I didn’t actually want spiders in this picture; her familiars are indeed spiders and they are important to the story, but I really didn’t want them in this scene. Unfortunately they kept creeping back into the picture until I spelled out (((spiders))) on the negative prompt. Eventually it backed off and just made webbing on her, which is not something I intended — but I am ok with that. She does have a bit of a vacant stare going on, which maybe is the AI interpretation of my pointing out that she is blind, or maybe its because AI art is a soulless endeavor and she just looks like that. Either way, not 100% what I had in mind, but was a very good interpretation of what I started with.

As a writer who has no claim to artistic talent outside of the written word and cartooning, I know that leaning on AI art is a crutch for me. Similarly, somebody using ChatGPT is leaning on that as a crutch for stringing together words. Neither is what a human would do, and it does not replace the work of talented writers (potentially including me). Buuuuut, its pretty good for roughing out a concept to show off to an audience, and given how visually oriented we are, it really helps writers communicate.