What is a picture prompt?
A line of adjectives. A picture prompt is not a description of what is in your picture — it is a line about how the picture looks: its light, its colour, its texture, its framing. We tested 61 common object nouns against the 480 lines this tool wrote from forty real files. 59 of them appear zero times. The only two that appear at all are sky and forest, and both of them only inside a colour name. The word picture never appears; neither does photo. Below: the noun sweep, the one pair of different photographs that got a line identical to the character, what the line reads off and what it never reads, and what to do with a line that has no subject.
What these numbers are. Forty real files — 39 stock photographs at 640 × 360 and one code screenshot at 800 × 450 — were decoded on this machine and box-downsampled to 192 × 192, the grid the live tool measures on, then composed into four shapes at three lengths: 480 lines. Every figure on this page is derived from those 480 published lines, the forty files’ own byte sizes, and the composer source; nothing was re-generated, no model was called, no image was produced, and nothing here is quoted from another site. The 480 lines themselves are downloadable from the page that published them.
The only nouns in 480 lines are colour names
Start with the simplest question you can ask of a line of text: what does it name? We took 61 ordinary object nouns and searched all 480 lines for them as whole words. Fifty-nine returned nothing. The two that returned something both did so for the same reason.
| Noun | Lines containing it | What it is doing there |
|---|---|---|
| sky | 24 | Only ever as sky blue, a palette entry. It never describes a sky. |
| forest | 24 | Only ever as forest green, a palette entry. It never describes a forest. |
| the other 59 | 0 | Absent from all 480 lines. |
That is not a coincidence of this sample. The composer draws its words from a closed set of phrases, and the only place a concrete noun can enter that set is the colour list. There are 20 colour names in it, all 20 of them turned up across the forty files, and they account for 1,908 of the words in the 480 lines — a median of four per line, with 468 of the 480 lines carrying exactly four and the other twelve carrying three.
| Colour name | Colour name | Colour name | Colour name |
|---|---|---|---|
| white | cream | beige | silver |
| grey | slate | charcoal | black |
| navy | blue | steel blue | sky blue |
| cyan | moss green | forest green | olive |
| gold | coral | brown | tan |
Two more words are worth counting, because they are the near-misses. image appears 240 times — it is the opening of the two sentence shapes, A widescreen 16:9 image. photograph appears 48 times, and every one of those is the monochrome clause, written only for a file the tool measured as black and white. picture appears 0 times. photo appears 0 times. So the tool will say the word image about your file 240 times over and never once call it a picture — which is a fair summary of what it is doing: it is reading a file, not looking at a scene.
Two different photographs, one identical line
If a picture prompt were a description of a picture, two different pictures could not produce the same description. They can. Among the forty files, one pair came out identical to the character in two of the four shapes at the short setting:
flux/short A widescreen 16:9 image. Moderate, slightly dim lighting. The dominant colours are charcoal, slate, grey, silver. dalle/short A widescreen 16:9 image, moderate, slightly dim lighting, with a palette of charcoal, slate, grey, silver.
Those two lines were written for p15.png and q776.png. Whatever the two pictures actually show, the tool measured the same twelve numbers in both — or at least the same values for the slots each shape writes — and so it wrote the same sentence twice. Across the twelve shape-and-length cells, 478 of the 480 lines were unique to their file; both of the exceptions were this pair, and both were at the short setting, where the fewest slots are written.
That is the cleanest statement of what a picture prompt is and is not. It is a readout of measurements, and a readout can collide. It is not a fingerprint of your picture, and it is not a caption. When it happens to be unique, that is because the colour list is doing the identifying — and the colour list is the only part of the line that survives a change of picture essentially never.
What the line reads off, and what it never reads
Everything the composer writes comes from twelve measured numbers: brightness, spread, edge density, saturation, monochrome share, warmth, two lighting-placement ratios, the aspect ratio, and the colour histogram behind the palette. Those are the only inputs. There is no field for an object, a place, an action, a time of day, or a count of anything.
That is why the noun sweep comes out the way it does. The line can say dark, low-key lighting; it cannot say a street at night. It can say moderate detail and texture; it cannot say a face. It can say brighter upper half; it cannot say sky above a field. The one time a noun of any kind reaches the line as a statement rather than a colour, it is the word photograph, and only for a file measured as near-monochrome — which is itself a measurement, not a recognition.
That one noun can be priced. We pushed all forty files down the saturation axis — brightness held where it was, colour taken out — and measured how far each one had to travel before the monochrome clause fired and the line finally said photograph: how much of a colour photograph you have to remove before the tool will call it one, and what that purchase costs the rest of the line.
Put the other way round: a picture prompt is a set of adjectives, and the adjectives are answers to twelve questions about pixels. If you want the picture prompt to mention a dog, you have to write the dog. Nothing in the tool can put one there, and nothing in the 480 lines did.
A twenty-one-times-larger file, the same-sized line
The forty files ranged from 22,834 bytes to 489,354 bytes — a factor of 21.4. If the line were a compression of the file, its length would track the file’s. It does not. Here is the correlation between a file’s byte size and the length in characters of the line written from it, for each of the twelve cells:
| Shape | Short | Medium | Long |
|---|---|---|---|
| Midjourney | −0.004 | 0.041 | 0.125 |
| Stable Diffusion | −0.004 | 0.041 | 0.125 |
| Flux | −0.063 | 0.095 | −0.007 |
| DALL·E | −0.032 | 0.158 | 0.139 |
Across all 480 lines the length runs from 90 to 357 characters, and the setting that decides where in that range you land is the shape and the length you picked, not the file you dropped. Inside a single cell the spread is still 48 to 107 characters, but that spread comes from which phrases the twelve numbers happened to select — a file that lands on well-lit, balanced exposure produces a longer line than one that lands on dark, low-key lighting — and from how many characters the four colour names take, which runs from tan to forest green. None of it is about how much file you gave it.
What to do when the line has no subject
Once you accept the definition, the working rules follow from it.
- Write the subject yourself. It is the one thing the tool provably cannot supply: 59 of 61 object nouns at zero, and the two that appear are colour names. Your noun is the only part of the finished prompt that says what you want rather than what the file is. The five-step page puts it at step 4, after the crop and the shape, which is where it belongs — there is no point naming a subject before you have decided the frame. Where in the finished line that noun can sit, and how to attach it so the parser does not read it as a weight, is measured on how to write a picture prompt.
- Do not trust the framing clause in someone else’s prompt. Framing matched on all 18,720 borrowings in our portability test, but only because every file in the set was 16:9. It is the clause that travels furthest and says least; if your picture is square, it is wrong on the first word.
- Check the colour list first. It is the only clause that identifies a photograph, and it survived 144 of 18,720 borrowings — 0.8%. If the four names in a prompt are not the four your picture actually has, the prompt is about a different picture however good the rest of it reads.
- Check warmth second. It is the narrowest band the tool produces, pinned to 4.2% of its range, and it is closed at both ends on 21 of the 40 files, so it is both the easiest clause to get wrong and the easiest to verify.
Forty files, 480 lines, one collision, no model called
Forty real files — 39 stock photographs at 640 × 360 and one code screenshot at 800 × 450 — were decoded on this machine, box-downsampled to 192 × 192, the grid the live tool measures on, and composed into four shapes at three lengths: 480 lines. Everything on this page is derived from that published set, from the files’ own byte sizes, and from the composer source: the noun sweep is a whole-word search of 61 nouns over the 480 lines; the colour counts come from matching the palette clause against the 20 names the composer can emit; the collision is the only pair of lines in the set that are equal as strings; the correlations are Pearson coefficients over 40 points per cell. The band figures and the borrowing test are the same ones published on the examples page, and the twelve measurements behind them are set out on what an image to prompt tool measures. No model was called and no image was generated, so nothing here says what any generator would draw — only what this tool will write about a file, and what it will never write. For the vocabulary the description is drawn from, see what an image describer actually writes; for the resize that decides every number before the words exist, see where your image goes; for the shape that carries the ratio, see the Midjourney shape from an image; and for adding the one part measurement cannot produce, see where your own words land in the line.