As mentioned earlier , I’ve declared that one of my creative goals for 2024 is to create at least 52 AI assisted/inspired artworks this year, where I’ll use AI-generated images as a prompt for actual physical creative pieces. This latest project is a real departure from my usual work. Instead of featuring lasercutting or polymer clay, this one involves creating a rose-and-skull fabric design. I didn’t think it would be horribly difficult. I was wrong.
For this project I had a very clear image in my mind of the end result that I wanted. I was aiming for a design made up of large white roses (specifically Yorkshire roses, if possible) and small white skulls. However, my first problem (learned from experience) was that most AIs didn’t seem to know what a Yorkshire rose was. (Have a look on the Wikipedia page for more info, but basically it’s a symbol with particular historical and geographical links.) Yorkshire is a region in the North of England, and its heraldic symbol is a white rose like this:
I’m from Yorkshire so that’s why I wanted a white rose. As for the skull; that was firstly because I often like to have memento mori reminders of mortality around me, to remind me to enjoy (or at least appreciate) life while I can. Secondly it was because my sister used to have a cool scarf that looked like a flowery design from a distance, but up close it was full of skulls. So I wanted to play with that surprise-skull idea.
This was my first attempt at a design, using Adobe Firefly and the prompt of: “Seamless repeating pattern of large Tudor Roses and small human skulls”:
(I used the phrase “Tudor rose” rather than “Yorkshire Rose” because I thought the AI might be more familiar with the term, but obviously it wasn’t.) Some parts of this design were amazing, such as the repeating pattern, and the overall look of the skulls and the roses). Some parts were disappointing, though – e.g. instead of big roses and little skulls it had big skulls and litte roses. Not bad for a first effort, though.
But when I tried to refine the image to get it closer to what I wanted, Firefly stopped cooperating. I went through about 40 iterations of attempting to get it to produce an image with large roses and small skulls, where it was continually ignoring the part about skulls. Example below shows that I even gave it a reference image in addition to the prompt, “Seamless repeating pattern of large Yorkshire Roses and human skulls, on a sage-coloured background. (A Yorkshire Rose is a traditional emblem of the county of Yorkshire in the UK. It typically features five petals, with small, green leaves visible between each petal. It is similar to a Tudor Rose, but all of its petals are white)”:
This is yet another example of when AIs go rogue. Firefly was simply refusing to create any skulls. So I had to give up with Firefly and try Kittl instead.
Kittl created some pretty good roses and skulls, but couldn’t create seamless repeating patterns:
…so I tried ChatGPT/DALL-E. That image generator produced tons of fantastic roses and skulls, but again was completely incapable of producing a seamless repeating design:
By this time I’d spent several hours trying to force the various AIs to come up with a seamless repeating design of large white roses and small white skulls. The different image generators could EITHER create a repeating pattern OR create a pattern of roses and skulls. So in the end I had to give up bashing my head against a brick wall and come to a compromise. I had to take my favourite image and ‘manually’ manipulate the size of the roses and skulls so that they were the right proportions. Then I had to manually adjust the elements of the image so that they formed a seamless repeating pattern.
One day I may create another post showing in detail how to create a repeating pattern, but briefly I used Linearity Curve to create ‘masks’ of the individual elements of the design, so that I could later change their sizes and positions:
I felt like this manual alteration was cheating, because the AIs hadn’t been able to do exactly what I wanted. But as with so many of my other projects there came a time when I had to recognise how much time I’d already spent trying to get the perfect image, and to just relax and tell myself that the technology wasn’t quite ready for me yet. It was so frustrating, though! I knew that AIs could create seamless repeating patterns, because Firefly had been creating loads of them for me. And I knew that AIs could create roses and skulls together. So why couldn’t they just do what I asked?!
Anyway, I had to just suck it up and do some final bits of image manipulation myself, to get the image I’d been hoping for. This was what I ended up with:
Creating the physical object
Having finally created my repeating pattern of large roses and small skulls, I could move onto the next stage of the project. That was to create a physical object from the design.
I thought I might try to use Redbubble to create an actual physical object. Here is their mockup of what my design would look like if it was turned into a backpack:
I’m still tempted to do that, actually. Maybe when my current backpack dies.
…but eventually I chose to have it printed on plain cotton fabric. I used Maake.com. I was so excited when the fabric arrived:
Here’s more of a close-up view:
The roses and skulls are exactly how I wanted them: you just notice the roses from a distance, but then from closer up you can see the little skulls. The only disappointing thing was that there wasn’t enough difference between the shades of blue. Never mind though – I’m sure I can find a use for this fabric. Maybe that will be another whole project. In the meantime – mission accomplished! An AI assisted physical object of rose-and-skull fabric.
How well did the AI(s) follow my prompts?
As explained above, the three AIs all managed some great individual designs but either couldn’t (in the case of creating repeating patterns) or wouldn’t (in the case of creating skulls) do what I asked. At the time of writing this post (April 2024) Kittl and DALL-E still can’t do repeating patterns. And Firefly still won’t mix roses with skulls. I’d score them all about 6/10 in their attempts to follow the prompts. Some of the images produced along the way were great, but just didn’t fully fit the brief.
Lessons learned
- As usual, the technology isn’t quite ready for me yet. I thought my requirements were fairly straightforward, and yet none of the AIs were capable of doing what I wanted.
- How to create repeating patterns. That was a real bonus, actually. I think it’s a skill I can use in future projects.
- There are some great companies around (e.g. Redbubble, Maake) who can turn digital images into physical objects.
Things I still need to learn or improve
- It would be good if I could come up with a way to tell when I’m wasting time versus spending time on the projects. Failing that, maybe I should just set a timer for working on the AI design phase of the project. ChatGPT/DALL-E helps with that, because after a while it stops and says I need to come back in X hours. Kittl and Firefly have ‘token’ systems instead, where you can only ask for a certain number of images.
- Apart from that, in general I just need to keep my eyes and ears open for updates in the AI text-to-image generators, to know when I’ll finally be able to ask the AI to create a seamless repeating pattern of large white roses and small human skulls on a green or blue background.
Homework:
- Now I need to make something out of the fabric!
Summary:
This unexpectedly turned out to be a monster of a project. I had to fight the AIs all the way to try to get them to create repeating rose-and-skull patterns. In the end I just had to compromise by doing some of the design work myself. However I was really pleased with the end result of the rose and skull fabric.
Thanks for reading this post. If you’ve got any suggestions of what I can actually make with the fabric – or if you’ve been experimenting yourself with AI-generated images – I’d love to hear about them. You can either comment below or send me a message via the Contact form.