Friends, this is the week! Dog Legs & Feet's “Bottom (A Live Collection)” goes live this Friday, April 21. Did you pre-save it? Now's the moment!
Last week I was talking about my quest for new artwork for the remastered album. Something fresh. Something retro. Something cheap or free. And most importantly, something featuring the Shakespeare character Nick Bottom.
Specifically, the image I really wanted was this: A person (or people) with a donkey's head, playing a guitar and singing, on stage.
There was just one problem. I have no visual art skills of any kind. And (just like the first time) no budget.
Robots to the Rescue
Fortunately, this quest was taking place in 2021. Generative text-to-image artificial intelligence was starting to make big leaps in quality, and tools to access it were starting to proliferate on the web. So I started playing with it. Specifically, I was using the tools at NightCafe Creator Studio.
What happens when you ask an AI to draw a donkey-man playing guitar on stage?
First efforts were… interesting. They were kind of impressive, in the sense of “wow, I guess you really can type a sentence and have your computer spit out an image.” But the images didn't come anywhere close to what I wanted. Here are a few examples:



It's like, well, yes… but no. I think these were done using the VQGAN+CLIP algorithm. In fact, I think the prompt may have been something like “Donkey as a rock star” which would explain why all the images have rocks in them.
After a lot of messing around, trying different prompts, and possibly getting a new evolution of the algorithm, I arrived at a couple of images that were better but still not quite there.



I didn't hate these, and I could sort of imagine using them as album covers, but it also wasn't quite what I had in mind. I found the David Bowie-looking one particularly striking and I even posted it on instagram asking for fans' feedback… but it also looks like a different kind of sound than what is on the Bottom album. It looks more arena-rock than taco-patio-bootleg.
The third one has a sort of Shakespeare look to it, and it got closer than anything to having a “Nick Bottom” look. But it also isn't quite there. I tried to go further down the Shakespeare road:


But these also didn't quite land at the right airport.
Some time passed, during which I was working on other stuff, and waiting for the other band members to sign some papers to move ahead with the remaster project. Then, in 2022, a couple of new algorithms came out (Stable Diffusion and DALL-E 2) that really took the results to a new level.
Less Confusion, More Diffusion
When I asked Stable Diffusion to make images of a donkey playing guitar in a variety of different artistic styles, I got things like this:




These are actually recognizable images of donkey-people playing guitars! If the robo-brains can do that, then it's just a question of finding the right image that conveys the right emotion, and that fits with the music. And yeah, maybe some tweaking to make sure the guitars have necks and so forth.
This part was pretty fun, just coming up with images over and over, throwing away the ones I didn't like and lingering on the ones that seemed promising.
A lot of times the algorithm gave me something too grandiose:

Or forgot to include a donkey:

Or got very close but made nightmarish anatomical decisions:

And then… somewhere in the middle of a late-night batch of trial images, I got one that felt like THE THING.

It wasn't perfect. Is it a two-headed donkey? What's going on with the ears on head #2? What's going on with his legs? Why does he have one hoof and one hand?
But also, the expression on his face is the right one. He looks earnest, and goofy, and kind of retro and folksy. He looks like a person who's been cursed to become a donkey, but is just going to go on singing anyway. He looks like the music on our album.
Inpainting the Final Touches
I tried a few variations of the image. I ended up going with one where the expression on his face is slightly more assured and less worried-looking:

When I saw this, I knew we were super-close. In fact, for several months, I thought I would just add a text overlay and use this image as-is for the album cover. During the whole time I was working on the remaster, that was my plan.
But when it came time to push the button, the image didn't quite pass the gut check. Reason: there are no strings on the guitar fretboard. As a guitarist, that bugged me. And I knew it would be easy to fix.
So I went back to Stable Diffusion and created what's called an inpainting mask. That's where you limit the algorithm to only working on one defined area of the image. I told it to add guitar strings, and it did. They were kind of messy looking, but fine, it scratched the itch.
And then I thought, while I'm in here, why not replace that second donkey head? I didn't mind it - after all, there was a lot of close-harmony singing on the album, so the two-headed image was appropriate. But it also looked just a little weird (especially the ears); and strictly speaking, Nick Bottom is the only one who gets a donkey head in the play, so it's a little off to have two donkeys singing together. Wouldn't it be better if that was a piano instead? Et voila:

So this is the final image! All that was left was to crop out the questionable anatomy (legs, fingers) and add the album name.
And you'll see the final version when it comes out THIS FRIDAY!
