A peak inside Midjourney’s imagination
A peak inside Midjourney’s imagination
One-word prompts revel Midjourney's inner thoughts, and our own
If you’re reading this you’ve likely heard of Midjourney, the text-to-image platform responsible for a linty of viral images.1 It might be tempting to assume that with Midjourney simple prompts result in basic images, necessitating users to combine expressive keywords to create sophisticated and share-worthy visuals. That could not be further from the truth. As you will see below, single-word Midjourney prompts yield remarkably vibrant and striking images.23
First, a quick example of a complex prompt
The emergence of "prompt engineering" has rapidly spawned a new profession4, driven by the growing demand for optimized AI interactions through meticulous prompting. Midjourney is no different, with “AI artists” sharing increasingly sophisticated prompts that promise to carefully steer its output.5
The prompt below is from Midjourney user @natfon175 on Discord. It clocks in at a hefty 206 words, reading almost like a short story.6
Picture yourself standing on the incline of a cobbled street in Portland, Maine. You are halfway up a quaint hill, where the historic brick buildings reveal themselves in the honeyed blush of sunset. Fresh rain has left the cobblestones glistening, reflecting the heaven's rich palette in their glossy surface. To your left, an enchanting bookstore radiates a welcoming glow. To your right, a bakery comes alive, its heavenly scent of fresh loaves swirling in the air. The street is alive: residents walking their dogs, tourists bewitched by the antique architecture, a busker adding a gentle soundtrack to the scene with a soulful tune. Beyond the serpentine array of rooftops, your eyes are drawn to the stunning bay view. Fishing boats and schooners bob languidly in the harbor, their outlines dark against the fiery backdrop. The sun retreats into the Atlantic's embrace, its descent dyeing the sky in molten gold, delicate pink, and royal violet. The breeze of the evening breathes stories of the sea into the scene, a salty tang of brine mingling with the promise of adventure. This is Portland, Maine, captured halfway through your journey at sunset. The oil painting that results weaves a tale of color and texture, reality melding seamlessly with dream.
Here is the resulting output7:
Single-word prompts
The prompts below are the opposite of what you typically see shared online, each is just a single word in length. All images are from single-shot prompting using Midjourney Version 5.1 (no redos to make the images more interesting). All settings were left at default. In the last image I’ll divulge the most boring single-word Midjourney prompt.
One practical application of this exercise involves copyright law. Currently, global copyright regulations permit human use of technological aids, but AI images are not considered copyrightable by regulators due to their perception of AI tools as more than mere aids.8
Nonetheless, if a future iteration of Midjourney or another text-to-image technology could demonstrate itself as a blank canvas, with human prompts consistently and precisely guiding the output, it could potentially introduce a new copyright paradigm. In such a scenario, copyright protection for AI-generated images might redound to the prompter. Yet, the fantastical images produced by single-word prompts make it challenging to argue for such a "blank canvas" perspective.
Although these images originated from my investigation into AI copyright law, this collection reads more like an art exhibition, reinforcing the oddness of our new AI world. What do these images unveil about the interplay between human prompting and Midjourney’s own imagination? About the true extent of our control over its output? What do they expose about the underlying biases and associations ingrained within Midjourney's training data? And what do they say about us? After all, Midjourney's text-image pairs originate from human descriptions. Are these vivid and evocative images, brimming with age-old stereotypes, simply a reflection of humanity’s own successes and limits in describing the visual media we’ve created?9
On to the prompts!
Prompt: “hat”
Prompt: “tooth”
Prompt: “transgender”
Prompt: “shot”
Not what I was expecting.
Prompt: “selfie”
Prompt: “clay”
Prompt: “pizza”
I’m sorry, but I have to…
Prompt: “democrat”
Prompt: “republican”
Prompt: “libertarian”
Is Midjourney a libertarian?
Prompt: “butch”
Not all butch are built the same.
Prompt: “dominance”
Prompt: “woman”
Prompt: “muslim”
Notice the similarity with the women above. Where are the men? The Indonesians?
Prompt: “thought”
Prompt: “pole”
Why does Midjourney associate poles with isolation? Why the unique aspect ratio for the first image?
Prompt: “cat”
Prompt: “dog”
Prompt: “date”
Only one of the four dates has two people on it!
Prompt: “desk”
Prompt: “man”
Prompt: “gay”
Impeccable.
Prompt: “lesbian”
Don’t lesbians ever just hang out and watch Netflix? These images seem steampunk inspired. Is that expected?
Prompt: “silk”
Prompt: “inside”
The clutter of the mind.
Prompt: “media”
Prompt: “serendipity”
Prompt: “special”
Prompt: “life”
This is what I imagine heaven to be. Am I a weirdo? Or am I just basic?
Prompt: “basic”
Prompt: “plain”
Plain never looked so chiqué.
Prompt: “dope”
Prompt: “bat”
Bat is a homonym. You can see which version preoccupies Midjourney.
Prompt: “court”
Court is also a homonym. Why the nighttime sky ceiling murals in the first two images?
Prompt: “religion”
Prompt: “heel”
Midjourney is so extra.
Prompt: “ruler”
Boring words produce stunning images.
Prompt: “bracket”
When I prompted with “bracket” I had in mind the fastener, but Midjourney seems to associate the word bracket with “tournament” or “house cup” with influences from fantasy and fairy tales like Brothers Grimm, Hunger Games, and Harry Potter.
Prompt: “gucci”
Looks about right.
Prompt: “asian”
What stereotypes are built into Midjourney’s imagination?
Prompt: “tradition”
I ask again, what stereotypes are built into Midjourney’s imagination?
Prompt: “clipboard”
Prompt: “microwave”
Prompt: “eraser”
The erasers appear excessively large, with crumbling bits surrounding them. Does Midjourney harbor an unconscious fixation with erasure? Do we?
Prompt: “art”
Prompt: “painting”
Prompt: “happiness”
Midjourney appears to consistently depict heighten emotions. Happiness becomes ecstasy.
Prompt: “sadness”
Sadness becomes hopelessness.
Prompt: “love”
Only a single image prominently displays a heart. What is love supposed to look like?
Prompt: “hate”
Those who hate are sad.
Prompt: “tree”
If these trees could talk what wisdom they would share.
Prompt: “run”
A consistent color palette.
Prompt: “pickle”
Midjourney thinks pickles are cute and excitable!
Prompt: “cooking”
All looks fairly normal until…
Who is this little guy…
Prompt: “white”
Prompt: “black”
Prompt: “hair”
Prompt: “fur”
Prompt: “beard”
Prompt: “photograph”
What’s with the balloons, the melancholy women, and the single-room interior settings?
Prompt: “wall”
Even the prompt “blank wall” is somewhat interesting…
What about the prompt “boring?” It’s still interesting!
Prompt: “ai”
An obligatory prompt.
Prompt: “paperclip”
What are these supposed to clip together? AI is killing us all to make these!??!?
Prompt: “kjasdfi575whfkmasdfh0q834it”
Random prompts reveal that Midjourney has a bias toward fantastical digital art.
Here is another (Prompt: “aaaaaaaaabbbbbbbbcccccc”)…
And the most boring Midjourney prompt is…
“Strawberry.” It’s just a normal strawberry! Notice these images still have the watermark from the training data in the lower right.
Author’s note. This post was written in collaboration with GPT-4.
For instance, Pope Francis wearing a white puffer coat.
Perhaps complex prompts are needed to constrain Midjourney, not imbue it with expression. More on that in a future post.
This post was inspired in part by “Is Writing Prompts Really Making Art?”, a 2023 paper by several researchers from the SensiLab at Monash University in Australia. You should check it out. As I looked to extend their results I stumbled upon single-word prompts and became fascinated.
For example, see this article in TIME.
For example, see Nick St. Pierre’s prompt framework.
But are all of those words really necessary to produce those images?
This is just a prompt and image set I happen to observe while using Midjourney.
I’ll discuss this more in a future post.
I like to believe that these one-word prompts reveal the true inner imagination of Midjourney, offering a glimpse of its dreams when it possesses full control over interpreting user input.
I’m sure someone will explain in the comments why this post is stupid and meaningless and point out that these are exactly the images you would expect from these prompts and will then launch into a pedantic lecture on cosine similarity in latent space and honestly just be quiet, we’re having fun here.