I've gone back and forth on AI art being cool and scary, but I can't deny that these Overwatch hero fusion renders I saw on Reddit last week are pretty sweet. It is fascinating that we can now ask an AI's opinion of what Soldier 76 and Mei would look like smooshed together, and it can spit out a shockingly cool, arguably original character design.
That's what author Drew Harrison did when he became one of the few thousand with access to Stable Diffusion, an AI image model similar to other popular tools like DALL-E and Midjourney. Harrison started by feeding the AI simple instructions like "make a fusion portrait of Mercy from Overwatch and Junkrat from Overwatch" and in some cases took more direct control over artistic influences with tags like 4K, highly detailed, cinematic lighting. Having only messed around with the DALL-E Mini online tool for a bit, I wasn't aware you can get so specific with instructions.
The results are striking:
I'm particularly intrigued/unnerved by the glare of Mercy x Junkrat (Mercrat) and the arm-mounted belly on Moira x Roadhog's (Moidhog) arm. You often can't expect an AI to spit out a fridge-worthy sketch on its first try. Harrison told me that AI art is often a process of trial and error.
"You try a prompt, see what comes back, revise, rerun, and repeat until you get the desired image," Harrison said. Speed is why Harrison currently favors Stable Diffusion over other AI art tools. He describes Stable Diffusion as "lightning quick and lightweight," two major advantages when your only creative bottleneck is render time and VRAM.
Stable Diffusion is also the biggest AI art tool in the process of going open-source, though it sounds like you'll need a pretty beefy rig to run it.
AI art still has some clear limitations, though. As you can see from Mercrat's weird octo-hand, AI models still don't seem to know what to do about fingers and palms. Sometimes there are too many fingers, or not enough, or what's there is a blurry mess. AI also tends to go weird on proportions when mimicking humans, though Moidhog is an extreme example. Check out the full gallery of Overwatch fusions here.
While I do hope Mei 76 becomes an actual Overwatch hero someday, the rapid proliferation of AI art makes me a little uneasy. Stable Diffusion, like every other AI, is trained on a dataset of internet images pulled from everywhere and anywhere. As Harrison demonstrated to me with "a fusion of Joe Biden and Barack Obama painted by Van Gogh," these tools are capable of seeking out and mimicking a specific person's style.
You could argue this is no different than a person drawing inspiration from another artist, but when we're talking about a computer brain storing millions of copies of other people's art, it's hard to know where inspiration ends and copying begins.