I’ve been playing at the edges of AI art for a few years, from early dalliances with running the Electric Sheep Screensaver to creating terrible book covers in Deep Dreamer. It’s an interesting intersection of my enjoyment of art, even if I can’t draw so much as a stick figure comic without failing, and my fascination with computer programming.
It all fell to the wayside for a while until the explosion, and subsequent ongoing implosion, of NFT art last year. I’ve mostly stayed away from crypto and generative art throughout COVID, other than riding the Dogecoin wave enough to pay off my car, then bail out and watch it all burn. Something about the attitude of NFT hawkers has always irritated me, with their declarations of superiority over old art, stupendously inefficient systems for “ensuring authenticity”, and bland, repetitive art styles. Toss in some echoes of the early days of video game DLC and the whole situation is reminiscent of my badass Dragon Age Blood Armor getting lost because my Electronic Arts account somehow got screwed up.
But there’s something alluring about the concept of teaching a computer to create art. The machine doesn’t know what it is doing. It really doesn’t know anything insofar as we can judge intelligence at this point. What the machine does have is the ability to evaluate a massive dataset of images, break them into colors, patterns, and tone maps, associate each of these fragments with keywords, and then reverse that process to generate something that the human AI wranglers look at and do a double take. This is of course vastly oversimplified description of how AI researchers are creating complex neural networks, which they only understand insomuch as they established the rules upon which the networks are based.
And much of the code that is used to train and run artistic AIs is open source. It isn’t always easy to get running and the hardware requirements can be bonkers, but in theory one could train their own private AI Artist Bot for a couple thousand dollars in consumer hardware or cloud computing tokens. The copyright status of the datasets used to train the AI, as well as the images which result, is in question and likely will be for several more years. And that’s before you get into the ramifications of using AI to generate illustrations, rather than paying a human artist.
Sometimes the results using artistic neural networks are genuinely stunning. The Verge reported on Wombo Dream a while back and the system has only gotten more fascinating in the months since. Just in the last two months Wombo has added the ability to use a source image to direct the AI into creating a particular shape and, of course, you can upload “your” NFT to create a remix. Of course, for every image that looks like the magical result of Vincent Van Gough and Timothy Leary sharing a weekend bender, there are just as many hilariously insane mistakes, such as the machine not understanding spatial relationships like the difference between “One orange in a bowl of apples” and “One apple in a bowl of oranges.”
I have an acquaintance who has been using Midguard to generate portraits for characters in her novels and roleplaying games. My experiments with the free version of Midguard has been less successful thus far, but I’m curious to keep playing with it as I get back into writing on a more regular basis. Learning to tease good art out of an AI is similar to learning to all good writing out of yourself.
We are just scratching the surface of this… revolution… experiment… distraction on the road to AI perdition… Whatever we call it, generative neural networks are in their infancy. It’s been just over a decade since Google taught an early neural network to recognize cats on YouTube and we have already come a long way.