Coloring page experiment
Winkletter • 5 May 2025 •
I’m letting myself fall behind on my writing again, but this time I’m busy running experiments. I’m pausing and just going to write about what I’m doing at the moment–AI-generated coloring pages (again). This is my benchmark for the state of the art because it has such specific requirements. It’s easy to generate something that looks like a coloring page, but on closer inspection it could have features that interfere with the coloring process: fills, textures, broken lines, and unenclosed spaces.
For this experiment I started with Stable Diffusion, generating tons of images that I could choose from (upper left). These are low quality, so I use Ideogram to improve the quality (lower left). Then I asked ChatGPT to create a coloring page image based on the new image. I even asked it to change the orientation to a 2:3 aspect ratio.
It’s not entirely perfect, but very close. I would probably edit this a bit with a bitmap editor to erase unnecessary details, and close off any incomplete shapes. Then it would go into Inkscape to turn it into a vector image so it can be upscaled and sharpened.
Comments
With ChatGPT you can always upload another image as a style guide. Ask it, “How would you describe this style?” Then upload the face photo and ask for the image of the person in the described style.
COOL! I’ve seen custom children’s stories with the name and features of the hero changed to represent the child being gifted the story. Maybe you can add a face to match a picture of a person. I did this recently for a book I’m helping a friend publish. He’s an old academic who has translated the Bible and the Quran in rhyming verse, with commentary. I uploaded a picture of him to ChatGPT and asked it to draw him as a dancing dervish. The first drawing was too cartoony, but I asked for an outline and it kinda works well.