Weblog:
I’ve been seeing all the online debates about how AI-generated art will steal the jobs of artists, so I decided to do an experiment. I used Midjouney–one of the most popular AI art generators out there (the same one where someone used it to win an art contest, causing a scandal), and gave it the following prompt:
“Baek Yerin with blond hair and in short black dress with black ankle boots, facing us, sitting with legs splayed and holding a black Gibson Les Paul guitar, in a dimly lit room with orange wall, boxes and crates, and a large suitcase with many stickers, photorealistic.”
These are the results:
I picked the one of them to refine:
And then created more variations:
As you can see, the mood of the lighting is not bad, but the actual subject isn’t even fully human, with missing body parts.
What I actually wanted was to get something that’s like an extension of this photo:
It’s the reference I’m using for a new painting, and the prompt I used is how I imagined the finished painting. None of the AI-generated results came close.
So next, I used that photo as the source image with heavy weighting towards it, to see if I can get the AI to generate what I wanted. I used the same prompt, but with –iw 5 (image weighting of 5) at the end of the prompt, to place heavy emphasis on the source photo, and I got these:
Again, not impressed. It tried to source Yerin’s face, and probably whatever photos of girls in black dresses it could find online, but completely ignored the rest of the prompt. So I tried one more time with a simpler prompt:
“blond girl sitting down with yellow guitar on lap, wearing short black dress and black ankle boots, legs splayed –iw 10 ” (I don’t know how high the image weighting scale goes up to, but since it starts with .25, I assumed 10 would be very high).
These are the results:
Still disappointing. At this point I decided it wasn’t capable of doing what I wanted. Maybe there is a way to achieve the result I want, but I was done spending time trying, because I can take my own additional photo references to supplement the original photo to get what I wanted. Or, I can just paint what I wanted out of my head (with some basic references like the Gibson Les Paul’s missing headstock and control knobs) by extrapolating from my decades of experience as a professional artist. The point is, the painting I can do on my own will be far superior than what the AI could generate.
But I wanted to give it one more chance, so I fed it this prompt, which describes a scene from the first chapter of my new novel:
“Father and teenage daughter being chased by vampires with glowing eyes on a dark street at night, lit by only a flashlight, Dutch Tilt low camera angle with 3-point perspective, photorealistic”
The AI gave me these:
Where’s the perspective I asked for? Where are the vampires? The figures don’t even look convincing as real people.
So, my conclusion is we artists don’t have to worry just yet. Maybe artificial intelligence will take over artists’ jobs one day in the future, but we’re not there yet. But what about all those incredible results you have seen on the Internet? Well, if you analyze them, they are all compositionally very simple images with straight-on camera angles with 1-point, or maybe 2-point perspective. And the ones with very realistic and detailed renderings? Those happened to contain components that already have detailed images on the Internet for the AI to source (I’m not sure if the AI can generate totally new components using tools like procedural 3D modeling, texturing, lighting, and effects). So, if you gave it a prompt that contained components the AI could find detailed source images online, you might get great results (though it will have very simple perspective), but if you want something that doesn’t have many high-quality sources for the AI to source (or the AI has no idea how to utilize the existing source images already out there effectively), you will get unusable messes like I did.
If you happen to be an artist who does the kind of work that AI can now compete against, then maybe it’s time to up your game and do the kind of work that’s beyond the capability of AIs. It shouldn’t be that hard, right? I mean, all you gotta do is make it 3-point perspective, with people that have all their body parts intact, and follow the description given by the client/boss closely, and you’d still have a job.
