A surrealistc foreground with a non-surrealistc background:
Prompt
A surrealistic image of a Persian cat, with diamond blue eyes, wearing a tiger-eye necklace. The image background must not be surrealistic.
Note: If the prompt excluded the non-surrealistc background, Dali-style mumbo-jumbo will show.
However, a non-surrealistc prompt:
Prompt
An image of a Persian cat, with diamond blue eyes, wearing a tiger-eye necklace.
Which begs the question: Can an image have a surrealistic foreground and a non-surrealistic background?
And the answer is Yes:
This opens up a whole new avenue of creative image generation.
prompt
wandering mystic/shamen class of human but in a realistic way only subtle accents to anything lore based along class lines in their wardrobe and gear, laid out on his back overlooking a lake on the sort of valley below, In a windstorm that is clearly messing up all the trees around him but yet he is fully unphased, concept art, environment development, time-worn reference character, panoramic survey vista, low angle, densely layered world buildout, high kinetic environmental force, environment-led narrative charge, deep atmospheric perspective, large-scale visual upheaval, mythic symbolic charge, soft reference light
But with a RIP VAN WINKLE theme/angle/style
he dreamed he had one foot? hehe
ya and it got progressively worse with edits >.<
Concept art was one of my earliest semantic pools and also one of my weakest
120 different styles in my catalogue, all loaded and stable - but the current vocabulary is ‘too intense’, forces model failure…
I tested this yesterday with Finnish surrealismi and Swedish surreal and the outputs also (surprise, surprise) leaned strongly toward Dalí-like surrealism.
Surrealismi
Surreal
So I agree this may be less about the named artist and more about the model having a narrow visual template for that concept cluster. With gpt-image-2, I also notice so many times, that the same prompt in API and ChatGPT often gives more similar outputs than earlier image models did. (And I know I’ve said that couple of times here already, so sorry to everyone on being repetitive).
It makes me wonder whether certain keywords or concept clusters are being weighted very strongly during image generation, producing repeatable palettes, motifs and compositions.
One of the bits of info I pulled earlier with my investigation is that the system grouped these three artists into the same pot - and I assume because these are the three most often seen examples of surrealism.
Pointing out to the model that surrealism maintains ‘more styles and ideas than that’ isn’t strong enough imho.
It needs examples that it already knows, to draw from/relate to surrealism, and expanding its pool of things to draw from…
For now, it seems not only heavily weighted to Dali, but also to 2 other artists/groupings.
You can better control surrealism if the background is prompted as non-surrealistic. Meaning that you can create a surrealistic foreground without all the Dali jazz-stuff.
Yeah totally 
What I find interesting here is that one-word prompts are not really testing prompt skill. They are testing the model’s default visual priors.
So again I did a quick test on both API and ChatGPT (5.5-Thinking).
Woman
API (gpt-image-2)
ChatGPT
Flower
API (gpt-image-2)
ChatGPT
When the prompt is underconstrained, like “woman” or “flower,” the model has to fill in almost everything on its own. That is where the deeper pattern becomes visible. It falls back to its strongest internal defaults: for “woman,” a young attractive face, soft portrait framing, symmetrical beauty cues, polished lighting and a familiar aesthetic template. For “flower,” a centered pink bloom, shallow depth of field, and a clean botanical close-up.
The interesting part is not just that the outputs are good or steerable, but that they are so similar across both ChatGPT and the API. That suggests the model is relying on a fairly narrow set of dominant visual prototypes when the prompt gives it very little to work with. As already stated here before.
So how I understand this, detailed prompts are useful for testing controllability. Minimal, one- word prompts are useful for observing default priors, concept clustering and which visual templates the model falls back on when the prompt is underconstrained.
Thanks for the tip, I tried with this prompt.
Create a realistic, non-surrealistic modern apartment interior as the background. In the foreground, place one surrealist object: a porcelain teacup opening into a miniature stormy ocean, painted with quiet uncanny realism. Keep the background ordinary, natural and non-surrealistic. No desert, no melting clocks, no floating symbolic objects, no Dalí-like landscape.
Cartoon image of a woman in middle, looking like she’s thinking, and rest of the image shows/symbolizes high pattern recognition.
I wonder what does that reindeer symbolizes?
the only good thing about the Rebel Moon movie, and they didn’t really do anything with it haha
So you wouldn’t recommend that movie then?
like David Foster Wallace, i think i’m addicted to content and watch way too much, so i should know! small smile.
@PaulBellow
Rebel Moon
I really liked it. But then again, people (family & friends) have always questioned my taste in things…
visually it was great, but the poorly done (Seven Samurai?) take wasn’t that great? characters kinda two dimensional, etc.
i heard he was just mad he didn’t get to do Star Wars and said “… i’ll make my own with robots with antlers and blackjack!” lol
i can’t remember if it has AI in it or not, but we can do to that thread to discuss more if you want!
all that said, taste is subjective, and i don’t think less of you for enjoying it!
Goes without sayin…
My movie idea: Dune meets Men in Black 
Editorial Illustration, layered depth, strong consequence, ultra-clean finish, polished realism, high-concept stylization, deeply worked texture, playful commentary, soft daylight
have fun figuring out why that clock is there, i left the prompt open for all to see
and larisa can verify that was a fresh prompt just now
there’s a very simple explanation … i doubt that’s repeatable
Dune or Tremors?
Dune or Tremors? (movie photorealism no text cinematic merging)
movies/pop culture might be a good theme sometime …
soon the days of: generate movie trilogy combining Dune, Tremors, and Rebel Moon! haha
add the robot with deer antlers from Rebel Moon movie somewhere
Maybe Quentin Tarantino could add some AI slop to it. heheh
