Image to Prompt by Image Type: How Much Is Already Written

Sort 40 real files into four kinds and a pattern appears: the more a kind of picture has in common, the more of the line is written before you upload anything. For the four monochrome files, 68.2% of the words in a line are words every one of those four files gets — against 29.3% when the same measurement is run across all 40 files at once. Backlit files sit at 50.85%, detail-saturated files at 40.85%, low-light files at 40.7%. Strip the colour names out and the monochrome figure rises to 74.7%, and on two of the twelve shape-and-length combinations all nine low-light files collapse to 100%. That percentage is the failure rate worth knowing: it is the part of your prompt this tool would have produced for any other picture of the same kind.

What these numbers are. Forty real files — 39 photographs at 640 × 360 and one code screenshot at 800 × 450 — decoded on this machine, box-downsampled to the 192 × 192 grid the live tool measures on, and composed into four shapes at three lengths: 480 lines. No file was classified by looking at it. Each kind is defined by a cut point the library already uses, so the groups can be reproduced from the measurements alone. No image generator was run and no image was generated — there is no image model behind this site and none was called. Every figure below describes the words this tool writes, never what any model would draw.

The rate, by type

The measurement is deliberately plain. Take every file of a kind, generate the same shape at the same length for each, and count the words all of them share — as a multiset, so a word that appears twice in every line counts twice. Divide by the median line length for that cell. What is left is the share of the line that carries no information about which of those files you uploaded. Twelve cells per kind, four shapes at three lengths; the figure reported is the median across them.

Two things in that table point in opposite directions, and both matter. The first is that removing the colour names makes the figure worse, not better: on low light it goes from 40.7% to 52.2%, on monochrome from 68.2% to 74.7%. The palette is not decoration — it is the only part of the line that reliably differs between two files of the same kind. The second is that the baseline is not zero. Even across all 40 files, 29.3% of a line is the same for everything you could upload, because the framing, the ratio and the scaffolding words do not depend on your picture at all.

Swapping one file of a kind for another confirms it from the other side. Take each pair of files, generate the same twelve shape-and-length lines for both, and count how many words differ as a multiset. Across all 780 file pairs and twelve cells — 9,360 comparisons — the median is 23 words, and the range runs from 0 to 54. Inside a kind it drops: 13 words for monochrome across its 6 pairs, 19 for low light across 36, 20 for detail-saturated across 231, 20.5 for backlit across 6. Fewer words change between two files that already look alike.

What the four types are, and how they were drawn

Nothing here was sorted by eye. Each kind is the set of files that fall on one side of a cut point the library already applies when it turns a measurement into a word:

1. Low light — mean luminance at or below 80, which is the last band of the exposure reading. 9 of 40 files.
2. Backlit — border luminance more than 1.18 times the centre, the branch that produces brighter edges with a darker centre. 4 of 40.
3. Monochrome — a monochrome pixel share above 0.85, the same cut point that decides both the saturation reading and the one scene statement the library ever makes. 4 of 40.
4. Detail-saturated — mean edge strength above 15, the top band of the detail reading. 22 of 40.

The groups are not a partition, and that is worth saying rather than hiding. 12 of the 40 files fall into none of the four, and 11 fall into two or more — a dark, high-edge file belongs to both low light and detail-saturated at once. Every kind is therefore reported against its own complement, not against the other kinds, and a file can be counted in more than one row. The baseline row is all 40 files together, kinds and overlaps included.

One result deserves its own sentence, because it is the reassuring one. Two files of the same kind almost never produce an identical line: across all 480 baselines only 4 rows are duplicates, 0.8%, and both sit in the detail-saturated group at the two shortest sentence shapes. The 40 four-shape signatures at the longest length are all distinct. Colliding lines are not how this tool fails; boilerplate is.

How wide a band one word covers

The mechanism behind the rates is simple to state and easy to miss: a reading is a fixed phrase, so everything on one side of a cut point shares it. Inside a kind, files often all land on the same side. The table below takes each measurement, computes how wide a band of values that kind spans, and divides by how many different readings that band actually produces.

Read the 42.41 / 1 in the top-left cell as follows: the nine low-light files span 42.41 grey levels of mean luminance, from 34.51 to 76.92, and every one of them is described by the same four words, dark, low-key lighting. The band has no lower bound — a file twice as dark still gets that phrase. The detail-saturated row is worse in the same way: 22 files spread over 49.43 units of edge strength and one phrase, highly detailed with intricate texture. Across the whole corpus that same axis produces three readings over 62.12 units, so joining that kind costs you two-thirds of the vocabulary on that axis.

The monochrome row shows the extreme case. All four files measure a monochrome share of exactly 1 and a red-minus-blue difference of exactly 0, so two of its five columns have a span of zero: there is no variation left for a word to describe. Five of its eleven slots take a single value across all four files — subject, framing, saturation, temperature and vertical balance — which is why so much of the line is already written.

One phrase per type, and nothing else is yours

A kind is only visible in the output if it owns some words. It mostly does not. Comparing every word a kind produces against every word produced by files outside that kind leaves a short list, and comparing whole slot readings leaves a shorter one.

Each kind gets one reading of its own — monochrome gets two — and everything else in its line is a phrase it shares with pictures of no particular kind at all. The word-level column is the harsher one. Backlit files, the ones a photographer would describe first by where the light is, own exactly 1 word that no other file produces: centre., with its full stop, and that is only because the sentence shapes end a clause there. Every other word in a backlit line — brighter, edges, darker — also appears in the line for a file that is evenly lit. The distinction is real in the measurement and nearly invisible in the vocabulary.

The screenshot that reads as clean and minimal

The corpus contains one file that is not a photograph: an 800 × 450 screenshot of code in a dark editor theme. It is worth looking at on its own, because it lands where no one would expect. Its mean luminance is 35.95, second lowest of 40. Its luminance spread is 13.68, the lowest of all 40. Its mean edge strength is 5.14, second lowest. A page of glyphs, sampled down to 36,864 cells, averages into flat grey.

midjourney  widescreen 16:9 composition, dark, low-key lighting, dominant palette of charcoal, forest green, slate, grey, moderately saturated colour, slightly cool tone, soft, low contrast, clean and minimal, smooth surfaces, evenly lit, moody and dramatic atmosphere --ar 16:9
flux        A widescreen 16:9 image. Dark, low-key lighting. The dominant colours are charcoal, forest green, slate, grey. Moderately saturated colour, slightly cool tone. Clean and minimal, smooth surfaces. Evenly lit. Moody and dramatic atmosphere.
dalle       A widescreen 16:9 image, dark, low-key lighting, with a palette of charcoal, forest green, slate, grey. Moderately saturated colour, slightly cool tone, clean and minimal, smooth surfaces. Moody and dramatic atmosphere.

The densest thing in the corpus is described as clean and minimal, smooth surfaces with soft, low contrast, and the palette it reports — charcoal, forest green, slate, grey — is an artefact of averaging thousands of thin glyph strokes against their background. This is not an edge case in the code; it is what a fixed resolution does to fine structure. Anything whose meaning lives at a scale smaller than roughly five pixels per feature will be measured as a flat surface, because at 192 × 192 that is what it becomes.

Two shapes and one length where the type is all there is

The clearest statement of the whole effect comes from the shortest output on the darkest files. Take the nine low-light files, generate the DALL·E shape at its shortest length, and delete the four colour names from each line. Eight of the nine become the same sentence:

A widescreen 16:9 image, dark, low-key lighting, with a palette of black, grey, charcoal, slate.
A widescreen 16:9 image dominated by black and white photograph, dark, low-key lighting, with a palette of charcoal, slate, grey, silver.

All nine share A widescreen 16:9 image, dark, low-key lighting, with a palette of . — twelve words, and 100% of the twelve-word median. The ninth line differs only because that file is also monochrome, which adds the one scene statement the library ever writes. Nine files spanning 42.41 grey levels of brightness, 66.67 of contrast and 102.6 of colour temperature produce, between them, one sentence and a clause. The same 100% appears on the shortest Flux output for the same nine files.

What the 40 files and 480 lines were divided by

Forty files, twelve lines each: 480 baselines, all from the same pipeline the live tool runs — decode, box-downsample to 192 × 192, measure, describe, compose. The four kinds were drawn from those measurements alone, using cut points the library already uses, so no file was classified by looking at it and the groups can be rebuilt from the numbers. The shared-word counts are multiset intersections over all files of a kind, computed separately for each of the twelve shape-and-length cells; the median across those twelve is what the first table reports. The band widths are the difference between the largest and smallest measured value within a kind, divided by the number of distinct readings that kind produces. The exclusivity columns compare a kind against every file outside it, not against the other kinds.

• No model was called and no image was produced. Every statement is about the words this tool writes and about the measurements it takes, not about what any generator would draw from them.
• The corpus is 39 stock photographs and one screenshot, all 16:9. That is why framing and the ratio take a single value everywhere here; a corpus with square or portrait files would give those slots more readings and lower every shared-word figure.
• A kind is a measurement band, not a category of content. Detail-saturated means high edge strength, which covers foliage, crowds, fabric and text alike. The four labels describe where a file sits on the scale, not what is in it.

For the three samples this site first used to make the point, see which images an image to prompt tool handles badly. For the measurements behind every phrase here, see what an image to prompt tool measures, and for the exact values at which each phrase flips to the next, the twenty-nine cut points.