Last month I designed the best t-shirt I’ve ever made. It was for Manifest, a conference whose shirts I’ve designed for 3 years now. I went a bit insane for this one – once I saw the glimmer of a really good shirt, I jettisoned the rest of my life for a month to chase it down1. I ideated with Midjourney, narrowed it down to 2-3 designs, refined those by hand + Photoshop + more Midjourney, narrowed it down to 1, and re-did everything by hand for stylistic and chromatic unity.
There are two aspects of this process I want to talk about: internet popularity, and AI.
Turning attention into beauty
Throughout this whole process I made decisions by running polls on Discord, twitter, and tumblr so I wasn’t anchoring on personal taste. I’ve been surprised a few times when the choice that seemed obviously best to me was unpopular, and surprised many times when people had a strong preference between two choices that seemed equal to me.
I’ve been doing this 2-3 years, and I think that while I am not an amazing artist, I have an amazing process for making the right/crowdpleasing choice at every major junction. I ran about 16 polls for the main shirt, 7 for the volunteer shirt, and 7 for the stickers.
A non-obvious part of all this is channel management. Most of these polls were in a 70-person Discord channel where I talk about my projects. Maybe half of those 70 read the channel. Some of those people do not vote on anything. If I just use the channel for A/B testing a design for a conference they are not going to, they will stop reading. So… I have to be funny. Crack jokes about my day, gossip, engage with their ideas, talk about other projects both speculative and ongoing. Now, I do this anyway, but this year more than ever I felt the transience of human attention. I would open up Discord and have the uncommon experience of consciously thinking, I haven’t been funny for a while. I have to be funny now. I’m not done with this design and I need people to look at this channel.
I have rarely found being popular online to be good for anything other than making me happy. Attention is pretty hard to turn into non-social value unless you’re pretty famous. A/B testing my way to beautiful designs is the main exception I’ve found. A leak in reality where I can turn attention into beauty.
Midjourney
I asked the organizer I could use AI this year because I have less free time with a 15 month old than with a 3 month old, and he enthusiastically said yes. And what actually happened was that I spent ~50% more time on making something that’s ~30% better than what I could have done without AI.
What did AI add to the end product? Mainly two things:
- ideas for stylization that I elaborated on (to a very different-looking end result, in part due to personal choice and in part because I’m not technically skilled enough to mimic some of the best parts)
- correct-looking anatomy for animals I am not used to drawing, in situations and poses I cannot find photographic references of them in
But the process was significantly more collaborative and organic than the above implies.
- At the start, I used Midjourney to generate six drafts that seemed about equally good to me. When I polled, the fox and hedgehog design tied with another design.
- Last year, I started with three candidates. Midjourney helped broaden the starting option pool. But some of the gains of ideating faster were eaten up by having to ask my poll audience to ignore the heterogeneous stylization and focus on the concept. This is a big ask!
- I pursued the other design for a bit, filed it away as the backup option, and came back to the fox and hedgehog.
- I repositioned the animals in a week-long spiral of futility, generating hundreds of Midjourney foxes and hedgehogs in the process. I was stuck for a long time because it is bad at getting animals into specific positions.
- Someone on the organizing team suggested adding a crystal ball, and I came up with a new composition where the fox curled to frame the hedgehog. I sketched out the desired fox shape on a post-it note and fed it to Midjourney to elaborate on. Sadly, their “elaborate on user-provided input” feature is weirdly terrible. After seeing the resulting foxes (best one featured below), I accepted I would be hand-drawing the design.
- At this point I had an AI-generated fox, hedgehog, and crystal ball stitched together into the final composition I would work from. After this I did not use AI.
I drew over this frankensketch in Procreate. I didn’t have the skills or experience to mimic the original Midjourney generation, whose polygonal style I had liked a lot, but if I hadn’t been trying to riff off of it I wouldn’t have ended up these interesting textures. (The hedgehog has one style and the fox has many!)
Midjourney did well generating digital assets or badge designs that could be used as-is, but t-shirts needed a lot of intervention because they have all sorts of problems with transparency and color and texture. You want to limit colors if the t-shirt is being screen printed, and AI doesn’t reliably do that yet. It’s nice to have an image file with lots of transparent gaps where the color of the shirt itself provides the color, because you don’t want a big contiguous area of slightly stiff ink, but Midjourney can’t work with transparency yet. These are not inherent limitations – I think people could easily tweak Midjourney to do this. They just haven’t yet.
So with that caveat, my take right now is that using AI made for a better end product, but I will not be doing this again because it was so much more work for me.
Footnotes
-
It didn’t have to take a month. Half the time was consumed by trying to get information from CustomInk about their printing limitations, failing, and putting in more iteration time when they told me it wasn’t viable. I also would have worked faster if my work period didn’t start at 10pm, when I felt recovered enough from childcare to do creative work. ↩