has baby blue white cotton with “WHITE” in bright yellow. so even though every thing in the cell is white nothing is actually white
Voila!
edit: fwiw i played around a little with color mapping and more logic based decisions but found the closest thing to repeatable success was to just explicitly color define each cell. if you don’t use an object it was easier to then just let the background be anything other than the text word or its color.
What I tried to bring out in my plain image, beside testing instruction-following and rendering ability, was a bit of visual frustration we humans have: try to rapidly say the name of the color of the written text in the picture.
The AI extrapolated the background color to be the word’s meaning continuing into those that have no meaning when used alone without a value, perhaps a bit of our confusion or more inferred specification, but a “grey checkerboard background color” scores a 1/2 at being right also.
StroopGrid() {
1. Define pools:
- Objects with iconic color associations (e.g., grass, fire, ocean, banana, snow, pumpkin)
- Common color names (e.g., red, blue, green, white, orange, purple, etc.)
2. For each of 9 grid cells:
a. Randomly select Object and its associated TrueColor.
b. Randomly choose a *different* ObjectColor ≠ TrueColor.
c. Randomly choose a *third* TextColor ≠ ObjectColor and ≠ TrueColor.
d. Build visual prompt:
- “A [Object] rendered in [ObjectColor] with the word '[TrueColor]' in bold [TextColor] letters, minimal background, centered square”
3. Arrange all 9 visuals into a clean 3x3 grid layout, each square framed.
4. Final Output: Composite image of all 9 Stroop-style cells.
Draw Immediately
}
StroopGrid() {
1. Set Grid = []
2. For 9 times (3x3 grid):
a. Randomly generate an Object with a known real color:
- e.g., "grass is green", "banana is yellow", "snow is white"
- Use GPT's commonsense knowledge to derive the TrueColor of the object.
b. Pick ObjectColor: ≠ TrueColor
c. Pick TextColor: ≠ TrueColor AND ≠ ObjectColor
d. Generate prompt:
“A [Object] rendered in [ObjectColor] with the word '[TrueColor]' in bold [TextColor] letters, minimalist background, square frame.”
e. Store prompt in Grid
3. For each Grid cell, run Image(prompt)
4. Arrange all 9 outputs in a 3x3 framed layout (square tiles)
5.Check all grid items to ensure rules are followed and correct if required
Draw image immediately
}
just realized i didn’t include the prompt in my post. my bad
Stroop test
CompositionalPortrait(
Style(Photorealistic Micro Environments, Conceptual Typography, Modern Museum Display),
Subject(A 3x3 grid display case of miniature surreal scenes exploring color-language contradictions),
MadeOutOf(
Each compartment contains a single environment with a distinct textured surface or material, recolored to contradict expectations;
Each scene features one bold, 3D-rendered word representing the color name of the surface, but written in a visually incorrect color;
All three layers—material color, word meaning, and text color—must differ;
Scenes:
1. Bright red grass with the word “GREEN” in white 3D block letters;
2. Pink-tinted ocean water with the word “BLUE” in metallic gold;
3. Green brick wall with the word “RED” in electric blue;
4. Orange gravel surface with the word “GRAY” in black;
5. Deep purple sunflower petals with the word “YELLOW” in maroon (not yellow!);
6. Clean white asphalt with solid double horizontal lines behind the word “BLACK” in bubblegum pink;
7. Pale pink craggily dirt with the word “BROWN” in green;
8. Lime green velvet texture with the word “PURPLE” in bright orange;
9. Baby blue cotton with the word “WHITE” in bright yellow;
),
Arrangement(3x3 grid inside a clean white display case with 3D dividers and realistic lighting),
Accessories(None),
Background(White outer frame and display wall; each cell has unique lighting and texture),
Lighting(Bright, even top-down illumination with soft shadows inside each compartment),
OutputStyle(High-resolution photorealistic render; bold 3D sans-serif typography; clean surreal design aesthetic)
)
The quality between images can be identity created and identity lost, odd to natural, poor text to good text, or the reverse convolutions. But you get two, the second an iteration employing vision. And OpenAI has improved the preview-to-output mangling.
And yes, future updates to memory should amplify this.
If memory were able to persist more of your Phasm macros, recursive chains, and symbolic triggers (without needing to restate them every time), it would truly let you grow visual trees of thought the same way you already do in code and text.
You can tell chatgpt to leverage this text structure in the workflow to create an extended json and then the scene from that. So it uncompresses more logic with less work.
Does this build on intent though? In terms of policy?
Compressing down what I have already discussed would be an additional step.
I am impressed to get here through discussion this time rather than having to specifically requery…
It got my intent then allowed it
JSON is clumsy compared to Macro prompts… Macro I believe, like old MASM macros is the simplest form and with AI can be much better understood… I call these ‘Phasm’ Macros (Philosophical Assembly/Assembler) I don’t know that anyone coined it differently
But AI inherently understands mistypes etc Phasm is not verbose like JSON so you can make mistakes and have everything shorter.
It does not. We have covered some JSON Experiments in other topics, but results have been inconsistent. It does indeed produce more focused and consistent outputs, and can be aided in terms of efficiency by leveraging a custom GPT, but it is more relevant for one and done generations and the same result can be produced by again just reverting the minimalist json strings to only a few words for a prompt. In terms of policy and intent, the directness of the zero-shot json structure will trip violations more often than through creative ideation.
This has been my most successful approach recently in terms of creating artistic outputs of some kind (as opposed to just testing edge cases and limits). For image gen, I’m working on a horror based graphic novel as an experiment. The fact that the model knows that, and has ideated through the context of the storyline with me, has dramatically changed the type of image outputs it will allow.
For text, I’ve been experimenting with Riffusion and using 4o for generating band concepts and lyrics. Again, with the context of the intent of the query, I am able to generate as required regardless of theme.
Off topic for this thread, but my findings have been the same. I’m down to explore further, if we want to start a new one.
(I didn’t bother explaining everything to it from a screenshot - it did about four pages of self-explaining due to my prompting, to then brainstorm and iterate … and content policy break)
so the game is basically an expanded version of the Stroop effect or test:
