I’ve been using Midjourney, an AI text-to-image generator, to make the images for this website for a couple of months now. For a person such as myself with zero artistic talent, Midjourney provides the opportunity to have custom art without paying out the nose. Naturally, many artists are upset about this and I’ll write about the economic and social impact of AI in the future, but today I just want to have a lighthearted look at the technology itself and how it has evolved.
The first thing to know about Midjourney is that it works better for artists. I recently used it to create art for a website I was developing that required a very specific style and ideas. Stock footage just wasn’t going to work. We could have created the art in-house with an artist, but that would have blown the client’s budget away. I couldn’t get Midjourney to create anything useful until the artist helped me write the prompts. Knowing the language of art and being able to translate a vision into text goes a long way to bringing out the best in Midjourney.
For this website, one of the first things I attempted was a picture of Beastie the BSD daemon holding or eating a raspberry. Several of the articles on this site are about getting a tiny home FreeBSD server running on a Raspberry Pi. I created a cross between the FreeBSD logo and the Raspberry Pi logo for this, and while it turned out decent considering my lack of proficiency, I dreamed bigger. My dreams confused Midjourney, even though I fed it a picture of Beastie as a guide:
Obviously several things went wrong. First, much in the spirit of my own amateurish design, Midjourney decided to merge the beastie with a raspberry. But Midjourney also didn’t seem to understand what Beastie looks like, so I tried to just make Beastie:
Clearly Midjourney is not familiar with Beastie. My next strategy was to become more descriptive.
Things seemed to have gotten worse. Nevertheless, I persist. The images below were one of my favorite abominations and I considered using some variation of them. They’re certainly interesting:
At some point I realized that Beastie is the BSD mascot, not the FreeBSD mascot, so I changed up the language a bit. I also began playing around with Midjourney flags that allow you to modify the prompt in various ways. It just made things worse:
Here’s another twist that had some interesting results:
I decided to switch strategies and go for something specifically using a Raspberry Pi. The results were pretty awesome, although I’m not sure how useful they are:
Version 4 of Midjourney turned out to really be a game changer…except it still doesn’t get Beastie right. Here I didn’t use the Beastie picture as a guide to see if it was aware of him:
However, with version 4 I was able to make much closer approximation of what I was looking for. Midjourney allows you to make variations of pictures you’ve made to try and hone in on exactly what you’re trying to make. I went a little overboard doing this but eventually I settled on something I could manipulate into holding a raspberry:
Here’s the end result. Not exactly what I originally envisioned, but maybe some day I’ll figure out Midjourney well enough to create my vision. I had to add the raspberry using GIMP and I tried to correct his right hand, which looked like a mangled mess. If you look closely, the hand still looks unnatural, but I figured I could get away with it since he’s a weird little demon.