GoCorral

Generative AI tools have started to become popular and I’m reflecting on my own uses of them now. I first learned about them as art tools in 2022. The first art piece I created was Stevie the Wonder Wagon because I couldn’t find anything close to what he looked like with a Google search. Even the final image isn’t quite what I imagined, but its close. And I still had to do some photoshop because the AI refused to include the arms I’d requested in the prompt.

Is it good? No. Does it get the idea across? Yes. And its far better than any pencil sketch that I could’ve done.

Stealing

The main issue with AI art is that its effectively micro stealing. The AI doesn’t really know how to draw a metal wagon with a face racing through a grass field. Instead it takes bits and pieces from other images to make this one. The field is a few different ones it mashed together. The wagon shape was another one. The metal texture is another one. The moustached face is another image put on top. Maybe a few more metal textures for that face. The tree and sky in the background is a few more images layered in the back. And then I layered the golem arms on top of all of it myself. So the image is composed of art from probably a dozen different artists, none of which consented to that process or to me displaying the art on my website.

So is this really that different than when I take any random image off the internet that looks like what I want and display it here? One of the more popular images I’ve uploaded is this one of Zeus. I never bothered to look who the original artist was until now. Apparently its an Israeli artist named Yonni Aroussi who made the image for a Greek Pantheon themed digital slot machine called Age of Gods. So the image probably belongs to the Age of Gods company and I’m effectively stealing from them by displaying it here without permission.

How is that different than the conglomerate image of Stevie? Well, it has a single source. If you like the picture of Zeus and wanted to pay for more stuff like it, you could go contact Yonni on his ArtStation profile: https://www.artstation.com/yondon. If you liked the Stevie image, there isn’t a way for you to get in contact with any of the people the AI took from to manufacture the image. Except for the arms which I took from the Mithral Golem in D&D fifth editions Monster Manual III which was illustrated by Bad Moon Art Studio. All the other bits are a mystery. The only way to reproduce that style is to use a similar AI generator with a similar prompt so that it’ll steal similar things from artists again.

AI art ruins the referral and exposure process that small time art theft helps to promote. It also ruins the ability of artists, like Yonni, to contact me and tell me to take their art off my website if they wanted. Maybe someone could recognize Stevie’s face from the art they produced, but its certainly harder to find than the distinctive Zeus from Age of Gods.

The other thing about AI art is that it really sucks at a few different things that are very noticeable for RPG art. The main things are hands, faces, fantasy creatures, weapons, and sensible backgrounds.

Hands

The hands thing really held up the pictures I used for Dwarves. I kept getting weird hands and odd setups for a prompt of “A group of Dwarves around a stone table.”

One six-fingered hand and all the other hands have noodle fingers. Plus, they look like they’re posing for a picture instead of talking to each other. This is “behind a stone table” not “around a stone table.”

Same problem. Six fingered hands, noodle fingers, posing for a photo. And a lot of these guys are cross-eyed if you look closely.

And here’s two more with funky looking fingers and faces.

Here’s the one I decided was good enough. There’s some hand problems. The left hand of the guy on the left is a little off. But they all have normal faces and hands for the most part. Just don’t look too closely at the candles…

So AI just sucks at drawing hands and fingers. It kind of knows what they are like, but it makes odd mistakes very consistently.

Faces

There’s two problems with AI faces. First, a lot of them just suck.

All of these guys look a little off. And this is one of the better sets of faces I’ve seen.

And the second problem with faces, it seems like AI has a set pool that it draws from so there tend to be a lot of repeats. See how many I got of this same guy over and over again with different prompts.

The last old dude is a bit of a stretch, but I still think its the same face. None of the prompts here said anything about facial hair or haircut, but they all ended up being the same style on a Jude Law look-alike. Even the dwarf on the back-left middle of this next one might be the same face.

It seems like generative AI is using a Markov model to adjust its faces to find ones that are “correct.” And once it lands on a correct one, it gets stuck in a Markov trap. Thus it has a limited set of correct faces based off different paths the Markov model took from different starting points.

Fantasy Creatures

AI seems to understand the basics of a few fantasy creatures, but can’t put the pieces together in the right places.

Dragons look correct at a quick glance. Serpentine neck, four legs, magic colors, two wings. But… the wings? They’re both on the same side of the dragon? And what’s that third, tiny, messed-up dragon in the background?

The prompt for this one was “a human transforming into a snake.” Not bad actually, but it appears that part of the transformation involves the human’s penis turning into a snake tail first?

This is supposed to be a centaur. While it might be a historically accurate centaur, it isn’t what I wanted. I ended up having to use a reference image to get the AI to draw a centaur correctly because it kept getting it wrong.

And of course when I fed it a centaur reference image it made the creature purple.

Another interesting error was how the AI had a consistent idea for what githyanki should look like. I tried a few prompts and kept getting this Indian looking horned humanoid. So the AI has a meaning for githyanki even if it isn’t the usual D&D meaning.

The dragon mistake and the snake mistake I can understand in a way. The AI is putting pieces together to create these images and it did those incorrectly. The centaur and githyanki are more confusing. Those are defined terms and the AI drew something similar, but wrong. Why didn’t it do those correctly? I don’t know.

Weapons

Similar to the dragon and snake parts put together in weird ways, AI doesn’t understand weapons very well. The AI starts one portion of a weapon and adds on whatever it thinks should be next, but that sometimes creates weird unnatural creations.

A bow. But its sort of broken in half and floating in midair?

Janky swords. And who’s holding that one in the middle? And what goes into that scabbard the guy on the right has?

The foreground sword looks fine except for having a double hilt. Funky hand holding it. But there’s also a sword with no hilt floating over the demon guy’s shoulder.

Another centaur that got messed up. The legs are wonky. Its butt got turned into a coal-powered steam engine. And its spear has a head made out of moss because its next to a tree, so the AI started thinking the end of it was a tree instead of a spear.

There are some tools to get the AI to redraw specific portions of images. It was an annoyingly iterative process, but I eventually got it to correct the centaur image above. Still quite a pain in the ass.

Sensible Backgrounds

AI makes a weirdly consistent error for backgrounds. It often pulls way too much of the given prompt into the background.

You wanted some goblins? HOW ABOUT ALL THE GOBLINS?!?

Tone down the goblins? What if there’s only a few goblins, but there’s also a building behind them and the building is actually a goblin?!?

Or if you want to have some werewolves howling at the moon. And then the AI puts four moons in the sky plus a night-time sun for some reason. And the werewolf on the left has three back legs.

Or some folks around a bonfire, while they are also on fire. But casual about it. And one of the folks at the bonfire is missing a head while the guy behind them fingers their neck hole.

Kobolds enjoying a nice meal around a campfire in their front yard. While their houses burn down in the background.

Another consistent problem for AI images is the lighting makes no sense. Where is the light coming from in this image? Do all the shadows make sense? There always seems to be this weird diffuse lighting, even with obvious light sources.

One More Thing

I also found that AI just couldn’t follow odd but specific instructions.

The prompt for this image was “tea kettle covered by bananas.” It is not covered by bananas. There are bananas next to the kettle, inside the kettle, and the handle appears to be a banana. But no bananas on top of it. I tried about 50 variations of this prompt and got some weird shit every time, but never bananas on top of a tea kettle. It’s not a complicated concept but the AI just couldn’t figure it out.

Going Forward

I was curious about generative AI for fantasy art. I was excited about it. And now I’m just disappointed. There’s certain concepts it just can’t do without a heavy amount of prodding. And if I’m spending a bunch of time prodding it, I can also use that time to find what I want by doing a Google search for real art.

The ethical issues with AI art have caused an enormous backlash against it. Posting AI art on some parts of the internet is met immediately with the online equivalent of boos and thrown tomatoes. There’s no way for AI art to be involved with any successful creative project unless its VERY well hidden. And with the quality issues I’ve discussed here, I don’t think hiding it is really possible.

In the future I’m planning on abandoning AI art. The ones that I already have on my site will get replaced with real art. And I’m also going to do my best to credit the artists of the stuff I am using. Still theft I suppose, but its more ethical in the long run than AI art. And it looks better.