Can I consistently create AI images in my own visual style?
I recently gave a talk about how AI threatens many creative jobs. I’ve spoken to some designers and illustrators about it, and a couple of them said they were not concerned because they had their own distinctive style that AI couldn’t replicate. I think their lack of concern is misguided — Generative AI is getting more and more capable, and copying styles is the kind of thing that it is getting good at. That got me thinking — can I get it to create images in my own style?
On this website you will see that I have created a fairly consistent style of images. The style is evolving over time, but some things are fairly consistent:
- A limited colour palette.
- Cartoonish images that I like to think of being in the Ligne claire style.
- A somewhat retro feel.
Here’s an example from my 2023 essay The Moral Machine - Could AI Outshine Us in Ethical Decision-Making?.
Another more recent example, from my essay Learning by writing essays:
Could I get generative AI to create images in this style?
Choosing the AI tool for this experiment
Recraft is a London-based company with a Generative AI tool aimed at designers. The tool is interesting to me for a number of reasons:
- Recraft’s latest model is currently in the number one slot on the Artificial Analysis Image Arena Board. This is a ranking of Text to Image performance for AI models.
- One of their boasts is their “brand consistency” — generating images in a specific style.
- Their tool can generate images in vector format, which is quite appropriate for my style (with line work and flat areas of colour) and is also good for easy editing and adjustments.
- It is supposed to be very good at generating text.
So I paid $12 for a month, which gives me 1000 credits on their system. Each image uses various credits, depending on complexity and size.
The online app allows you create your own custom styles, by uploading up to five images (or choosing images you have created with their system). So I uploaded five images similar to the above. I then prompted it to create SVG images in my style, with the theme of a 1950’s style spaceship on someone’s lawn. After a couple of false starts it created this:
Isolated 1950’s style spaceship on a lawn.
Gosh, that’s rather good. It is recognizably my style. Note that I didn’t ask it to include text, but text is part of my style.
A few tweaks to the prompts resulted in other results. Here’s a couple of them:
Whimsical cartoon with simplified linework and muted colors of a spaceship above a 1950’s town.
Not bad! It seems to like to use the red and orange colours a lot, which I tend to use more sparingly. Since these are vector images it would be very easy to change that colours.
Whimsical cartoon with simplified linework and muted colors of a 1950’s style robot on a lawn next to a spaceship .
I’m fairly impressed. I think with a bit more experimenting, and combining this with other techniques, I could probably come up with very consistent images. So I’m thankful that I’m not a professional illustrator. Those guys are going to have a hard time, unfortunately, at least the lower ranking ones.
To be clear, one of the things I am not going to do is to start replacing my hand-drawn images with images generated by AI. To me these are great new tools to experiment with and incorporate into creative processes. But they shouldn’t replace creativity, nor the striving to become a better creative person. They are a tool for creativity, not a replacement for it.