Continuing with my ongoing project of “An AI made me do it”, this latest mission was a great excuse to use AI to generate an image that I could then turn into a physical object. I needed a birthday card for my dad, and I wondered if AI would be able to help me. Normally I like to do papercut cards for family and friends, so this project became “Papercut birthday card for my dad”.
I logged into ChatGPT/DALL-E, and gave it the prompt:
Birthday card papercut saying “Happy Birthday Dad”.
As a first attempt, I thought this was amazing. DALL-E even added an envelope! I was impressed that most of the design could actually work as a papercut. There were very few isolated elements that would fall out if they were cut exactly as shown. However, there was a confusing section near the middle of the design where two elements were layered together (the snail or spiral on top of a flower shape).
At the same time, DALL-E also created this card…:
…which I would have immediately tried to cut out – if only it didn’t very clearly say “THUD” right in the middle. What on earth?!
So having discovered that DALL-E knew what a papercut birthday card looked like, I gave it another prompt:
“Great! Please could you include some little light aircraft in the image?”
At first sight these were amazing again, but:
- they both had non-standard spellings of “Birthday”,
- The left card was almost impossible to turn into a card without a lot of work,
- Most of the planes didn’t look feasible as 2D paper objects.
Try, try again…
As usual, just when I thought that the AI was going to produce a fantastic result with the next image, we ended up having a fight instead. It either mis-spelled words (including “Dad”!), or presented the image at a sloping distorted angle, or came up with a design that would be impractical as a papercut.
Also as usual, the designs were nearly right, not really right. Here are some of the rejects:
In the end I just had to pick one that looked possible to cut without many adjustments:
I printed it out on a normal piece of printer paper, then used that printout as a cutting template.
Here is the hand-cut version that I produced first:
…but then I continued to work on a laser-cut version. To do that I used Linearity Curve, which has an ‘auto trace’ function. Auto trace converts JPG files (i.e. photos) into vector graphics (i.e. collections of flat shapes). Auto trace sometimes makes little errors with the shapes (but is still way faster than tracing around the images by hand), so afterwards I edited the nodes of the vector shapes to try to get them to match the original image a bit better:
…and here is the physical laser-cut version (on the right), next to the original AI design:
How well did the AI follow my prompts?
I would say about 8/10, for this project. Very successful! Most of the designs looked like papercut birthday cards. And all of them (when requested) contained light aircraft. However – as mentioned several times before in this ongoing series of projects – creative spelling was the opposite of useful or helpful.
How close is the physical version to the original AI design?
I’m going to say it’s about 85%. The differences are:
- I didn’t cut out all of the sections that were cut out in the original design – I just etched them instead.
- I changed the tail section because if I’d cut it out then lots of the pieces would just have dropped out. (Also, the stripe extends forward from the tail, which doesn’t look quite right.)
- Some of the swirly bits have been simplified.
- I couldn’t find any dark green card or paper to mount the plane on, so I chose a speckled dark blue piece instead.
- The letters in the “Happy Birthday” banner had to be converted into stencil-type letters (otherwise the central sections of the As, Ps, B, D and R would have dropped out).
The elements where I just left the AI design and didn’t make any changes:
- The outline of the plane
- The design of the banners.
Lessons learned
ChatGPT is great at papercut birthday card designs.
Things I still need to learn or improve
Actually I’m pretty pleased with this project. Even the ‘duds’ had lots of elements in them that I felt could be used in future projects.
Homework:
- Convert at least three of the AI’s original designs into actual cards.
Thanks for reading this post. If you’ve got any suggestions of prompts – 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.
