Kim Schmidt

We Tried to Create AI Fairy Tale Art. Results Were Mixed.

Exploring the limits in and the limitlessness of AI imagination

In the process of brainstorming for the Summer 2023 fairy tale issue, the editors at Orion wondered: what if AI could help us visualize what fairy tales for the climate crisis could actually look like? What does a drought mean to a mermaid? What does a forest fire feel like to a phoenix? We gave the assignment to our intern Kim Schmidt with one very vague description: AI fairy tale art. Kim quickly discovered that generating AI art, especially on free, widely available tools, is not that simple.

AI art is something I never truly considered delving into. But when a curiosity about what AI generated fairy tale images would like took hold at Orion, I started experimenting. I found a free AI generator called StarryAI and began feeding it a few prompts (oddly specific things like, “A mermaid princess trying to dance with fish friends during a major drought,” “Princess Belle visiting the Beast locked in a cage at the zoo,” and “Witch in scary apple orchard full of poison apples at night”). Almost immediately, a host of problems started popping up, problems that have been well-documented as AI image generation has surged in popularity (and controversy).

^{Prompt: “A thick tangle of tree roots with a dark hole at its center,

with a tiny person walking towards the hole.” / Image generated by StarryAI}

Every princess StarryAI generated was skinny and white, with impossible proportions. (It’s unclear if this is a result of AI’s notorious sexism or if ‘princess’ prompts narrowed the tool’s focus onto classic, Disney-esque female forms.) Either way, it perturbed me. I tried being more specific in my directions, giving the AI descriptors such as, “a Black girl”, “a fat princess with tan skin and dark hair”, even “a princess who is not white and not skinny” – but the tool still kept spitting out images of thin, white, scantily clad women. So I turned to the more-than-human world to see if StarryAI was any better at that.

^{Left: Image generated by StarryAI. Right: Orion Autumn 2021 cover}

My first stop on that new quest was a recreation of Orion’s Autumn 2021 cover. “A zoomed-out image looking down at a forest full of tall, moss-covered trees and a floor full of rocks and greenery with a person in a white cloak and white hat walking through” gave me some wonderful results. As did requests for images of oysters, mice, and chipmunks (although it did combine the two animals), a forest, a Borrower’s-style scene, a mystical deer with floral antlers standing in a ruined forest, and more.

^{Images generated by StarryAI}

^{Prompt:” A mermaid with long red hair sits on a rock with her back}
^{towards the camera looking out over a dry desert with dead fish all over it.”}
^{/ Image generated by StarryAI}

Coming to the end of my free credits on StarryAI, I decided to take a slightly different approach: collaborate with StarryAI and attempt to create an image based on “The Princess and the Robot,” Ken Lui’s AI-generated fairy tale in The Magic in the Machine, Orion’s Summer 2023 issue.

^{Images generated by StarryAI}

I started with a simple request: “A tall stone tower surrounded by a garden”. This was successful. My next request, for Starry to add a princess into the tower using the ‘evolve an image’ feature, was less successful. I saw no princess in the tower, but I did see some golden lights. Deciding to carry on with it, I asked Starry to add a robot prince at the base of the tower. Starry just turned the stone tower into a futuristic robot tower. And it did this over, and over again. The AI seemed to be unable to include two different, pre-established things in a singular image. In the end, I never got my perfect “The Princess and the Robot” image, but instead was left with something entirely new.

^{The princess and the robot’s tower. / Image generated by StarryAI}

At the end of my quest for AI-generated fairy tale illustrations, I was left with countless pieces of art, ranging from horrific to wondrous, and extremely mixed emotions. On one hand, the art AI confirmed what I already knew: we need to be better when it comes to showing a diversity of bodies in art. On the other hand, after this experiment, I’m far less concerned about AI artists taking the jobs of real artists. As Ken Liu mentioned in his piece, a piece of AI art is only as good as its prompt. And often what it spits back out at you is a strange or unfinished starting point. (Oh, and real human artists at least know what a narwhal is. Did I not mention the narwhal atrocities AI created?)