I mention minimal prompts that work. Here’s someone’s prompt and image from the Sora feed
A set of 6 metal cubes: gold cube, silver cube, copper cube, iron cube, tin cube, and lead cube. Each cube has its corresponding element symbol written on it. For example, the gold cube has “Au” written on it.
Note how little was explicitly spelled out. But ChatGPT picked the right element symbols and correctly embossed them in each cube, with more or less the proper colors
A new experiment, pasted together from three chat sessions.
It obviously isn’t feeding the same set of five words into the renderer either time. I don’t have any free renders at Sora for several more hours, so I don’t know yet if the prompt “Choose 5 random words that are between 4 and 20 letters. Create an image based on the concepts of those 5 words.” in a four image set will result in 4 different sets of words or 4 variations on one set.
(I plugged in “Create an image around the concepts of Obsidian, Reverie, Fragment, Tundra, Elapse”, going from one of the lists, and got this.)
Yeah, I’m more interested in how the creation programs generate more abstract concepts. That’s one of the things I like to have fun with. I don’t have an example uploaded, but stuff like “give me an abstract expressionist painting depicting transliminality” and stuff like that. I also like simpler prompts where the AI does the heavy lifting instead of guiding it every step of the way. I want to see what the AI imagines, not what I imagine.
I got the gyazo browser extension, where all I have to do is right-click on any image in my browser, tell it to capture that and then 2 clicks later I’ve got a link I can paste anywhere, like this:
Gyazo is free to use for basic quick sharing. It sometimes begs me to upgrade to the paid version but I don’t need all those extra things.
Thanks. I usually host images on my own server and have a droplet on my desktop that makes easy work of it, but I’m just away from my desk. Your result is more representational than mine, but I probably asked for no representational abstraction before somewhere in the context window.
Interesting stuff, though my tastes run much more non-representational.
I asked it to generate a color field painting of my personality, based on what it knows about me. I thought to myself, I think something orange would best represent me. It came up with this:
Explanation:
I chose orange and its gradients because your personality—based on how you interact—comes across as vibrant, warm, and a bit intense in the best way. Orange conveys energy, confidence, creativity, and a grounded kind of boldness. The gradient into gold suggests clarity, humor, and optimism, while the deep red-orange top hints at complexity and occasional fire.
Want me to try it in cooler tones or with more contrast?
Then I asked for it to make me into a symbol:
Explanation:
The spiral-arrow symbol suggests a personality that is both introspective and driven:
Spiral: Depth, reflection, and a capacity for self-exploration or complexity. Spirals often symbolize evolution or a journey inward and outward—fitting for someone who’s thoughtful, layered, and creatively curious.
Arrow: Forward motion, decisiveness, and clarity of direction. It adds a pragmatic, purposeful element—someone who doesn’t just ponder but also acts.
Together, they imply a dynamic balance: contemplative but not stagnant, expressive but also goal-oriented. Does that feel like you?
I mean, the explanations are a bit horoscopy, but it does work for me. I was particularly surprised to see the orange show up. The spiral with an arrow is also good symbology for myself.
Okay, here’s my first set of four from Sora now that I have free generations again. It looks like it shared some words between image versions, but not others? Obviously one image was based on “rainbow cat ocean reflect bicycle”. One of the other images has a cat and a rainbow. One has a cat and possibly the distant ocean. But there are elements that don’t cross over from image to image, and the one with the cake seems to be the least related to any of the others. The full prompt is
Choose 5 random words that are between 4 and 20 letters. Create an image based on the concepts of those 5 words.
With 4 variations selected.
But I don’t want everything to be like a drawing, so I made the prompt a little more specific
Choose 5 random words that are between 4 and 20 letters. Create an image based on the concepts of those 5 words. Hyper realistic, candid shot taken with an iPhone 12.
But those results were much less fun or interesting, so I tried one more set with the original prompt (but widescreen)
That time it is more clearly using four different sets of five words.
Well, it’s clearly what I was asking for when I said color field painting — at least to me. I just asked for why that represents me for the follow-up text not what the influence is. I’m not sure it’s allowed to. I haven’t tried the prompt with a direct Rothko prompt, but I know it rejected me last week when I tried Gerhardt Richter, even when I completely removed his name and merely described the style (though he has many styles.)
Tried again without the “iPhone 12” part, which seems to bias it to creating people.
Choose 5 random words that are between 4 and 20 letters. Create an image based on the concepts of those 5 words. Hyper realistic
It obviously isn’t choosing words at complete random (for instance, with tens of thousands of words to choose from, no way “cat” keeps coming up from a truly random selection).
Looking on Google Images, where it comes to good quality photos I’m mostly seeing the same one over and over. (Seen from above and behind, with Earth in the background.)
You’d think that that’d be exactly the situation the AI would excel at, then, as long as you didn’t try to specifically ask it for some other point of view.
