Click here to skip the write-up and go to the Stable Diffusion imagery.
I was playing around with Stable Diffusion, with the idea in my head of trying to visualize dreams. I haven’t searched, but I suspect it is a fairly common thing to do. Dreams are, after all, the original generative AI. An actual image generating AI that can visualize anything you can dream up is no doubt a dream in itself for some artists and researchers. I decided to take a scrap of a writing exercise I had done and run it through Stable Diffusion. The exercise was pretty on-point: “write a short-short story based on a nightmare.” I tried to capture the feel of a real recurring nightmare I used to have (no longer, thankfully).
I settled on feeding it in one sentence at a time. Without providing the text yet, I want to talk about it as a generative prompt: To wit, it was immediately obvious to me that it was a poor prompt. For all the words, there are few adjectives or sight-related words for Stable Diffusion to latch on to. Lots of vagueness: somewhere, something, etc. I mistakenly thought I was a poet in one or two sentences, but other than that, not much to go on. Not even any colors, only suggestions of shadows and darkness. Some sentences are very short, some run on. I was a little dismayed. With visions in my head of some miniature graphic novel, looking at the text, I could see it would be more of a chaotic jumble of randomness.
Ultimately, I resolved that I should embrace the randomness. Let the AI dream its dream, not mine. Further, I wouldn’t game my concept by re-generating endlessly in hopes of finding cooler imagery from each sentence. Instead I simply picked one of the four variations Stable Diffusion generates at its default settings. Below, the result.
Welcome to my nightmare
I think you’re gonna like it
I think you’re gonna feel … you belong-Alice Cooper, Welcome To My Nightmare