What an image prompt generator puts in the line

One line, and the split inside it is lopsided. Across forty real pictures, the median output line is 29 words: 6 of them the shape writes itself, 23 of them measured off your file. The 6 are scaffolding and never change. The 23 come out of twelve slots — and on these forty pictures only nine of those twelve ever took a second value. Framing and the aspect-ratio parameter never moved once; the palette changed in 99.2% of picture swaps.

What these numbers are. Forty real files — 39 photographs at 640 × 360 and one code screenshot at 800 × 450 — were decoded on this machine, box-downsampled to the 192 × 192 grid the live tool measures on (36,864 cells, whatever the file size), and composed into four output shapes at three lengths: 480 lines, twelve per file. No model was called and no image was generated; every figure below comes from re-running that pipeline. The raw output is downloadable as line-anatomy.csv, 480 rows with a word count and a template/filled split for each one.

Here is the shortest line in the whole set, at 14 words and 90 characters, and where each word came from:

A widescreen 16:9 image, well-lit, balanced exposure, with a palette of tan, brown, olive.

Nothing in that line is a guess about what the picture shows. Every one of the eight picture-side words is a threshold comparison against a number the measurement produced — and that is the whole trick an image prompt generator is doing, whether or not it says so. What follows is the same accounting across all 480 lines.

Twelve slots, and how many of them ever move

A line is not assembled word by word. It is assembled from twelve fixed positions, and each position is filled from one measurement or left out entirely depending on the shape and the length you asked for. The table below is the census: how many of the 480 lines carry each slot, how many distinct values that slot took across the forty pictures, how much of the forty the commonest single value covers, and how many words it typically contributes.

Two slots never moved. Framing came out as widescreen 16:9 on all forty pictures and the aspect-ratio parameter as 16:9 on all forty — not because the generator is stubborn, but because all forty files happen to be 16:9. A portrait file would move both in one step. That is worth knowing before you read a fixed slot as a limitation: it is fixed for these forty files, not by construction.

At the other end, the palette is the only slot that behaves like a fingerprint: 37 different strings across forty pictures, and the most common one is shared by just 10% of them. It is also the slot that decides the length of the line, because it contributes a median of 4 words but can run to 6. Everything else sits between: exposure resolves to one of 4 phrases, saturation to 4, colour temperature to 5, and the three slots with the narrowest vocabulary — contrast, light distribution and vertical light — to 3, 3 and 2.

The lines column is where the shapes disagree with each other. Contrast appears in only 160 of 480 lines because two of the four shapes never write it at all; vertical light appears in 56 because only two shapes carry it, and only at the longest length; the aspect-ratio parameter appears in 120 because exactly one shape has a place to put it. Choosing a shape is mostly choosing which of these twelve slots you get.

The shape writes eighteen words in total

The scaffolding is smaller than it looks. Across all 480 lines, the four shapes together use 18 distinct template words:

Eighteen words of fixed vocabulary against 85 distinct words on the picture side. Each shape uses a subset, and the subset is smaller than the total:

The gap in the last two rows has one cause. Flux and DALL·E carry four template words — frame, is, dominated, by — that only exist inside the subject sentence, and that sentence only appears on the 4 monochrome pictures. Midjourney and Stable Diffusion have no conditional scaffolding: all of their template words, 5 and 7, appear in all 120 lines.

That is the practical difference between the two halves of the vocabulary. The 18 template words are closed and knowable in advance; the 85 filled word types are drawn from the measurement, and the number a shape can draw is set by which slots it reads — DALL·E reaches only 66 types because it never writes contrast, light distribution or vertical light.

What forty lines of the same shape still share

Set the forty lines of one shape and length side by side and align them word by word. Two things fall out: a common opening, a common ending, and a middle that is different every time. For the alignment below, two lines are compared by the longest run of words they share in the same order, over all 780 unordered pairs of pictures in each group.

The 0 in the Midjourney prefix column is the most informative cell in the table. Midjourney opens with the subject slot when it has one, so on the four monochrome pictures the line starts with black and white photograph instead of the framing phrase — and forty lines then have no common first word at all. Drop those four and the opening is three words, widescreen 16:9 composition, on all thirty-six. Stable Diffusion has the same four lines but hides them behind its quality prefix, so over all forty it still shares 3 opening words, and 6 over the thirty-six. Flux and DALL·E put the subject later, so their openings survive intact: 4 and 3 words over all forty.

Endings are the mirror image. Only Midjourney has a fixed tail on every length — --ar 16:9, two words, plus atmosphere at the long length for three. The other three shapes have a common ending only at the longest length, and it is one word: atmosphere. At short and medium they end on a measured value, which is different for every picture.

The template column carries its own surprise: it is the one part of the line that does not scale with length. Midjourney uses exactly 5 template words at every length while the line grows from 21 to 45 words; Stable Diffusion exactly 7. Flux and DALL·E vary, from 6 to 11 and 6 to 8, and the variation is entirely the conditional subject sentence — with those four words removed, they too are flat.

Swap the picture and ten of the words move

The last column of the previous table says how much survives a change of picture. On the median pair of Midjourney short lines, 13 words stay in place and 10 move, out of a line whose median length is 22 words — so a little under half of a line is picture-specific. Which slots do the moving is not evenly distributed:

Only 6 pairs out of 780 share a palette — the six that do not are counted in the row above. Everything else in the line is comparatively coarse: contrast changed on fewer than half of all pairs, and the subject clause — the one place where the generator makes a statement rather than a measurement — only ever separates pictures that differ in monochrome status, which is 18.5% of pairs.

So the answer to “how much of this line is mine?” is: about half the words, concentrated in one slot. If you want a different line out of the same generator, the palette is the only part of your picture that reliably moves it; the other eleven slots either agree with the previous picture or were never going to change on a batch of similar files.

What the first page of results leaves out

On 2026-10-18 the ten organic results for this phrase were fetched statically, scripts and styles stripped, and each page counted: body word count, whether it states the grid it measures on, whether it offers anything downloadable to check against, and whether it says how many words its output has. The raw fetch is in prep/_serp/gap-ai-prompt-generator-2026-10-18.json.

Four counts out of that table: 0 of the ten state the grid they measure on, 0 of the ten offer anything downloadable to check the output against, 2 of the ten say how many words the output has, and 3 of the ten carry fewer than 400 words of body text — two of them 9 words, which is a page that renders its content in JavaScript and has nothing for a crawler to read. The median body length across all ten is 1,413 words.

None of that makes those pages wrong. It means the one thing a reader cannot do after any of them is check the tool: there is no stated grid, no file to diff, and no stated line length. This page exists to be the opposite of that, which is why every number above has a row in the CSV.

Forty pictures, 480 lines, and the file behind every number

The short version. A line out of an image prompt generator has a median of 29 words and ranges from 14 to 47. 6 of those words come from a closed vocabulary of 18 template words and are the same in every line the shape produces; the rest are drawn from 85 word types across 12 slots. Two of those slots did not vary once on these forty files and one varied on 99.2% of picture pairs; the remaining nine run from 18.5% to 79.6%. Length buys you more slots, not more scaffolding: Midjourney adds 24 words between its shortest and longest setting without adding a single template word.

That is a description of one pipeline, built so that every word traces back to a number. If you want the numbers themselves and where each threshold sits, start with what this tool measures. If you want to know how much of your picture is left after the grid, the thresholds and the vocabulary have each taken their cut, that is measured separately in how much of your picture survives. And if you want to see the same twelve values written four different ways, the Midjourney shape is where the differences are easiest to read.