I Trained AI to Make Art

I’ve become fascinated with AI in case my previous article wasn’t clear enough. Oddly enough, I’ve been curious with OpenAI’s engine and defining what it thinks art is by telling it what I want. OpenAI will search a database of pictures with tags and create what I write based on what it “thinks” I’m asking for.

It’s a simulation of logic using pre-existing data to create something out of nothing. It’s basically assessing various objects from the input and placing them based on where those objects are normally placed in other pictures. This is what the first artwork came up as when I told it to make “a floating city in clouds”:

a floating city in clouds

Cool. Then I thought about asking it to create an image of a band I like. However, many bands I like have rather generic names like Tesseract, Contortionist, Dark Fortress, Enslaved, etc. I would imagine asking AI to build that would just create a literal version. Then I remembered a band I used to listen to as a kid named Cradle of Filth. This is what it came up with:

cradle of filth

It literally made a picture of the band. I was honestly hoping for some cradle used for babies with expiring foods inside, but that, unfortunately, did not occur. I know it looks weird, but that’s because I stopped its simulation at 2,000 iterations after I realized it was re-creating a picture of the band. I imagine by 10,000 iterations or more it would look like a picture if you typed it in your favorite search engine (please say it’s Duckduckgo). Not interesting at all.

Then I wondered if I made a request for something that I can’t imagine a picture existing for. Something that could be taken many different ways, and go into many directions. Like, “death falling into hell.” Hey, I got to keep the dark theme going here.

death falling into hell

Wow. I honestly love this and tried searching for it again and could find nothing that made me think, “this looks like a copy of x.” It looks completely original. I asked a friend to define something very specific without much context as to why. This one is called “a nice clown helping a mean old lady”

a nice clown helping a mean old lady

Weird, but I like it. It kind of reminds me of some Russian art style. This one had 5,000 iterations. At 1,000 iterations it looked very different:

a nice clown helping a mean old lady at 1,000

This was quite an interesting and fun experiment that was created in Portuguese, so it took me a while to translate everything so I could run this AI. Now, I want to show a video of the iterations over time. I used an AI random sentence generator to create “The beauty of the African sunset disguised the danger lurking nearby”. The following video includes 5,000 iterations.

Anecdotally, I don’t see a huge difference between iteration 500 and 5,000. Some minor clarities maybe, but I think anything beyond 1,000 is probably unnecessary until AI can be improved. This video was produced at an interval of 10 iterations.

One reply on “I Trained AI to Make Art”

Comments are closed.