I was preparing a lecture in which I would cover various ways UX researchers are experimenting with AI. I decided I would provide AI-generated illustrations that went along with the uses. So, I entered the following prompt into ChatGPT.
I want you to create a series of illustrations for me, all in the same visual style (ink outline and watercolor illustration) and color palette. All of the illustrations will feature the same robot. The illustrations will feature two different humans. Human 1 is a middle-aged Asian woman. Human 2 is a young hispanic man.
1. A robot conducting a usability test with human 1 as the user. She is sitting at a desk with a laptop. The Robot is taking notes.
2. Human 2 conducting a usability test with the robot as the user. The robot is sitting at a desk with a laptop. The human is taking notes. Should look like illustration 1 with the roles reversed.
3. The robot interviewing human 1. They're standing at a store checkout counter.
4. Human 2 interviewing the robot. Should look like illustration 3, but reversed.
5. The robot helping human 2 look through a filing cabinet.
This was the first resultāonly four illustrations that donāt follow my instructions, and one of the robots is wearing jeans.
I don't want the images I listed combined into a single graphic. I want 5 separate graphics.
The illustrations you created are good stylistically. But you didn't follow my instructions close enough. Try again.
Now weāre talking. Okay, this might work out. On to illustration 2.
Whoah! What happened to the robots head?
You changed the robot. Can you do that illustration again using the same robot from illustration 1?
Uh, alright, fine. Letās move on to the third one.
Redo that, but have the woman standing with a shopping cart, and make her look like she is speaking.
Why did you change the image format?
Thanks for pointing that out ā the change in image format happened because the prompt for illustration #3 called for a different scene and composition, so the model automatically adjusted the layout and ratio to fit aĀ taller, more vertical formatĀ for the standing interaction at the store checkout. This can happen when elements like shopping carts or full-body figures require more vertical space than a desk-based scene.
If you want, I can re-generate illustration #3 with the same horizontal aspect ratioĀ as the first two images to keep the series visually consistent ā just let me know!
Yes, keep the series of images as consistent as possible.
You put the robot standing inside of the shopping cart. That's dumb. The woman should be pushing the cart.
Good enough. I decided to keep that one and move on to illustration 4. It started drawing that one in the square aspect ratio. So I stopped it, and told it to use the original aspect ratio. It started again with the square aspect ratio, so I again stopped it and asked why it was using the square aspect ratio.
Thanks for your patience ā youāre absolutely right to ask. The square aspect ratioĀ in Illustration #4 was unintentionally introduced due to the default behavior of the generation tool when the aspect ratio isnāt explicitly specified on each prompt. Even though the visual style was preserved, the format shifted.
To keep consistency with theĀ original horizontal layoutĀ (like the first illustrations), Iāll regenerate Illustration 4 with the proper aspect ratio. I will ensure all subsequent images match that format too.
I'll be more specific. In illustration 4, the robot should be pushing the shopping cart. Human 2 should be interviewing the robot and taking notes on a clipboard.
Iāll take it. Moving on to the last one.
Not quite. I want the robot digging through the filing cabinet. The man should be holding files looking on.
Okay, so they took the drawer out of the cabinet, and the robot is sorting from the back, but this had taken way too long already. Iāll go with it. And this is what I ended up with in my lecture. Thereās very little consistency.