AI Diversity Is Hard

The Updated Section Image

This is the last of the 36 rules-specific illustrations that needed to be updated. Which took all of 2 weeks. There will be more for the faction-specific fluff, although some of those are done already. But the only thing left to update in the rules is the photographs of gameplay situations. For which I'll be using the 3rd-Party miniatures and terrain that I painted up for Gencon a few years back. That's all ready to go.

But the reason I'm so happy with this particular illustration has to do with its difficulty. Given the inherent biases that Stable Diffusion (my AI model of choice) has.

First, a last look at the previous image:

This was the old section illustration, and it couldn't be saved with a quick AI pass.

Now for a closeup on that map table...

There's 4 races of Republic Laansguard depicted here, and each is nicely different from each other. But that didn't happen automatically, and even if I'd instructed the AI to do it, it wouldn't have.

You see, the AI likes to make all of the characters in an image look the same. It might gender-swap a character or two, but it will still be the same look, face, hair, etc. in a differently shaped body.

Here, despite a multi-racial prompt, we have all white guys. The two on the left have the same hair style, and so do the two in the middle.

These were prompted as "East Indians" because prompting "Indians" literally gave me bare-chested Native Americans in headdresses. Ick!

Even so, they're all the same guy! They're also looking at jewelry instead of the map I prompted, because... WTF?                                                

The AI requires that you use "Africans" or you'll only get slightly tanned White guys instead. Again, all the same guy.

And the Chinese version.

So I had to break out my Photoshop skills, and pluck my favorite racial representatives from each group. Placing all of them in the same image looking at a proper futuristic map. Which I cut out of yet another image that I only liked that particular part of.

This turned into a 4.5 hour slog as I made modifications in Photoshop, ran them through the AI again to make the subjects match the new background, and then made more changes. Rinse, repeat.

Still, that's only about half of the time that a professional painter would take to do a similar scene with just 1-2 characters with less detail and finish, so we're still talking about a substantial net productivity gain.

Thoughts?