Art on a Tiny Budget Part 1: Generative AI
I used to think the biggest barrier to making games was coding.
It only took a few small hobby projects for me realize the truth: most games have a lot of art. Someone has to create every little tree, painstakingly draw every animation frame for every character. If you want your game to look professional, all of your art needs to match in style and quality. And if you want your game to stand out, your style needs to be unique.
For me, an increasingly experienced coder with minimal art skills, it’s become clear that art is a much bigger barrier than coding.
So what is an indie dev to do?
Generative AI
Generative AI is pretty freaking crazy.
Now that it’s here we all have to grapple with a bunch of ethical and legal questions around ownership.
Who owns the art created by AI? In the US the answer is likely “no one.”
Is it fair for AI to train by looking at art that other people own? Some artists seem to think not, if we can judge by their lawsuits.
Can people use AI to create images of me without my permission? As I write this, actors are on strike in part because of this exact concern.
Then there are the larger societal questions around lost jobs and the role that human creativity should play going forward.
There are a lot of articles out there discussing these difficult topics, and I have my own viewpoints on many of them. But for the purposes of this article I want to ask a different question: in its current form, can generative AI even create useful art for games?
The answer, I’ve found, is mostly no. But let’s get into some examples and details, using OpenAI’s DALL·E tool.
2D Game Sprites
I found DALL·E to be basically useless at making game sprites. To start, it only creates fixed images, so it can’t be used to create anything that needs to be animated.
When I tried to get it to make top-down spaceship sprites, all of my results looked like this:
DALL·E seems to roughly understand the shapes and colors it should use, but the details are sloppy, with weird lines and asymmetries. It does far better with drawing a single large asset than a set of smaller ones, but here “better” still isn’t good enough.
Character Portraits
Next I tried generating character portraits for JRPG-style dialogue sequences.
After doing some research I decided to add “trending in artstation, digital illustration” to all of my prompts, in an attempt to get a similar art style across images.
Here are some of the best results:
These look a lot better than the spaceships. Some details still look off and the art styles don’t quite match, but they’re at least close.
I tried using a couple of GIMPs effects (an oil-painting effect to smooth details, followed by a cartoon effect to get bold border lines) to standardize the style a bit more. Here were the results:
These look worse than the originals, but I believe they’re also a bit more coherent as part of the same product.
Other Categories
The keycard in my studio logo was AI-generated. At some point I’ll want to change that, as it’s probably a good idea to be sure I actually own my own logo.
But ignoring that, the prompt “sci-fi keycard on a white background, minimalistic, geometric” was largely successful at generating options I liked:
The last interesting thing I tried to generate was concept art for a Flaregate. Flaregates are star-powered technological marvels that connect different star systems. They play an important role in the developing lore of my game.
Here’s the result of my efforts:
The above image is a combination of elements from two different attempts by DALL·E:
Final Thoughts
After all of the above experiments, I’d mostly recommend using generative AI to create internal concept art. Internal concept art keeps the artists and designers on the same page as they independently create game assets.
In my case (a single developer who will eventually try to find one or two artists to work with) I can use AI-generated concept art to more easily explain to artists what various characters and objects should look like.
Secondarily, if you’re not worried about owning your own game art, you may be able to generate some specific category of useable non-animated images. This depends a lot on your game’s needs and your own tolerance for imperfect work. I was able to generate character portraits that I felt were good enough, but if possible I still want to replace these with the work of a real artist.
My original plan for this post was to cover a whole set of strategies I’ve been using to deal with this whole art issue, but this post turned out to be long enough to stand on its own!
If you’re interested in more content like this, you can follow me on Twitter (or X or whatever you want to call it now)!
The game also has a supportive and active Discord community following its development.