One thing I wish this article had tried: adding _restrictions_ like "no cheese" to the input.
That kind of input confuses the image models pretty badly - for instance, asking Stable Diffusion for a picture with no giraffes will usually return a picture with giraffes...
Lots of Stable Diffusion web UIs have "Negative Prompts", which attempts to remove that stuff from the image. It does actually work phenomenally well, and watching 20 steps with updates each step is mesmerizing. You can really see the random colorful blobs deciding what they want to be, as some err too close to a negative prompt, then flip into a less-blurry thing that looks more like your positive prompt.
The trouble is, any pixels/objects that are interpreted as "negative" according to this prompt are reduced or removed. So, overzealous Negative-Prompting can Negatively impact images. Ha ha!
For Text AI, chatGPT is more than capable of even removing "length" from it's responses, if you spend 3-4 messages telling it "I'm dyslexic", "Please be concise",, "Please do not repeat prior messages unless they are relevant".... it finally went from "Essays" to "Tweets", for that session
That kind of input confuses the image models pretty badly - for instance, asking Stable Diffusion for a picture with no giraffes will usually return a picture with giraffes...