Image to Prompt Across Models: Four Lines, One Measurement

Put one picture through the four model shapes and the surprising thing is how little they disagree: across 1,040 slot groups — twelve slots, three lengths, forty files — there is not one case where two shapes write different values into the same slot. The maximum number of distinct values any slot takes on any picture is 1, at every length. What differs is which slots a shape bothers to write and the words it wraps around them. At the longest setting the four lines for one picture share a median of 26 words out of a 38-word median line, and only 3 words appear in all four lines on all forty files: widescreen, 16:9 and atmosphere. Switching shape costs you measurements; it never hands you a new one.

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, then composed into four shapes at three lengths: 480 lines. A slot is one of the twelve fields the composer fills (subject, framing, exposure, saturation, temperature, contrast, detail, light distribution, vertical light, mood, palette, ratio). Each slot's value was located inside the finished line by searching for the measured string itself, so nothing here is a paraphrase of the source. No image model was called and no image was generated — every figure describes the words this tool writes, not what any generator would draw.

Where the four lines agree: every measurement, every time

The four shapes read the same measurements off the same grid. That is easy to assert and worth testing, so it was tested the blunt way: for every picture, every length and every slot, collect the value each shape actually wrote into that slot and count how many distinct values appear. Twelve slots × three lengths × forty files gives 1,440 groups; 1,040 of them have a value at all, because shapes drop slots and two slots are conditional. In all 1,040, the largest number of distinct values is 1.

That result is stronger than it looks. It means the four shapes are not four opinions about your picture — they are one opinion, written four ways. When a Midjourney line and a DALL·E line differ, the difference is never "this picture is bright" versus "this picture is dim". It is that one of them left the brightness out, or said it inside a sentence instead of after a comma. Two consequences follow, and both are measurable: the difference can be listed slot by slot, and it always runs in one direction.

What each pair of shapes differs on

Take the two lines for the same picture at the same length and compare the sets of slots they carry. Six pairs of shapes, three lengths, forty pictures each. The table reports the median number of slots that one line has and the other does not, together with the most common list — and how many of the forty pictures that exact list covers.

Three things stand out. The two list-shaped outputs differ by exactly one slot at every length, on all forty pictures: Midjourney writes the ratio as --ar and Stable Diffusion writes no ratio at all, and apart from its two quality tags at the front they carry the same fields. The two sentence-shaped outputs are even closer — identical slot sets at short and medium, and a single slot apart at long, because DALL·E never writes light distribution. The whole gap between the list family and the sentence family is thus 2 to 4 slots.

The 28 / 40 covers are the honest part of the table. Vertical light is conditional: it is written on 28 of the 40 files, so on the other 12 there is nothing for the sentence shapes to lose and the count drops by one. Long length is the only setting with two different lists per pair, and every one of those variations traces back to that single conditional slot.

The difference is subtraction, never addition

So the shapes drop slots. Do they ever add one? The comparison was run in both directions: for each pair, the characters of measurement that appear in the first line and not the second, and the characters that appear in the second and not the first. Across all eighteen pair-and-length cells, the second figure is 0 — eighteen times out of eighteen. No shape writes a measurement another shape omits. There is no shape that sees something the others miss.

Read the two ends of that table together and the practical answer appears. Moving from Midjourney to DALL·E removes a median of 57.5 characters of measurement — up to 91 on the worst file — from a line that was 300.5 characters long, and gives back nothing; the DALL·E line is 222 characters. Moving from Midjourney to Stable Diffusion costs 6 characters on every single file, all of them the ratio, and the 17-character gap between their median line lengths is scaffolding on both sides: a ratio parameter on one, two quality tags on the other. At the shortest length the same comparison gives 63 characters lost going to Flux or DALL·E and 57 going from Stable Diffusion to either sentence shape.

How much of the line survives the switch

A word-level view says it more plainly. For each picture, take the four lines at one length and count the words all four share, as a multiset, so a word appearing twice in every line counts twice. Compare that with the median line length.

Now put the two axes against each other. Two lines for the same picture in different shapes share a median of 0.674 of their words at the longest length, over 240 such pairs. Two lines for different pictures in the same shape share only 0.414, over 3,120 pairs. Change both at once and it falls to 0.286 over 4,680 pairs. In other words, swapping the model preserves more of the line than swapping the picture does — the picture decides more of your prompt than the target model does.

The one place that reverses is the shortest length, and it is close enough to call a tie: 0.446 for a different shape on the same picture against 0.455 for a different picture in the same shape. Short lines drop five of the twelve slots outright — contrast, detail, light distribution, vertical light and mood are written by no shape at that length — so most of what is left in them is scaffolding that every picture and every shape shares.

One picture, four lines

Here is one file — a monochrome photograph, one of the four that also gets a scene statement — at its longest setting, exactly as each shape writes it:

Midjourney        black and white photograph, widescreen 16:9 composition, dominant palette of tan, slate, charcoal, silver, moderate, slightly dim lighting, monochrome, black and white, neutral colour temperature, medium contrast, moderate detail and texture, light falling on the centre with darker edges, balanced, natural atmosphere --ar 16:9
Stable Diffusion  masterpiece, best quality, black and white photograph, widescreen 16:9 composition, dominant palette of tan, slate, charcoal, silver, moderate, slightly dim lighting, monochrome, black and white, neutral colour temperature, medium contrast, moderate detail and texture, light falling on the centre with darker edges, balanced, natural atmosphere
Flux              A widescreen 16:9 image. The frame is dominated by black and white photograph. Moderate, slightly dim lighting. The dominant colours are tan, slate, charcoal, silver. Monochrome, black and white, neutral colour temperature. Moderate detail and texture. Light falling on the centre with darker edges. Balanced, natural atmosphere.
DALL·E            A widescreen 16:9 image dominated by black and white photograph, moderate, slightly dim lighting, with a palette of tan, slate, charcoal, silver. Monochrome, black and white, neutral colour temperature, moderate detail and texture. Balanced, natural atmosphere.

Every measurement in those four lines is the same: the same palette in the same order, the same exposure, the same lighting distribution. What changes is the wrapper. The palette arrives as dominant palette of on the two list shapes, The dominant colours are on Flux and with a palette of on DALL·E. Counting distinct wrappers per slot across the four shapes, on each picture, gives the last table — a per-slot measure of how much of the difference is phrasing rather than content.

Temperature, contrast and vertical light are the three slots all four shapes phrase identically — a comma and a space, in every case. Framing, exposure and palette are the three with the widest spread: the same value reaches you three different ways depending on the shape you picked. On the four files that get a scene statement, all four shapes wrap it differently, which is the highest figure in the set at 4.

That is the whole of it. Choosing a shape changes which of the twelve measurements reach you and how they are introduced; it does not change any measurement, and no shape adds one the others lack. If you need the line to carry contrast, light distribution and vertical light, the list shapes are the ones that write them, and the sentence shapes will quietly leave them out.

What the 1,040 slot groups and 240 line pairs were compared on

Forty files, twelve lines each: 480 lines from the same pipeline the live tool runs — decode, box-downsample to 192 × 192, measure, describe, compose. Each slot's value was found inside the finished line by searching for the measured string, and the wrapper was read as the characters between the end of the previous value and the start of this one; slot order was taken from the line, not assumed. The 1,040 groups are the (slot, length, file) combinations where at least one shape writes that slot, out of 1,440 possible. Pairwise slot-set differences were computed for all six shape pairs at all three lengths on all forty files; character losses are the summed length of the values one shape writes and the other does not, plus two joining characters each. Shared-word counts are four-way multiset intersections; the overlap figures are multiset Jaccard over 240 same-picture pairs, 3,120 same-shape pairs and 4,680 mixed pairs at each length.

• No model was called and no image was produced. Every statement is about the words this tool writes, 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 takes one value everywhere here and why widescreen and 16:9 are invariant; a corpus with square or portrait files would give that slot more readings and lower every shared-word figure.
• Vertical light is conditional and appears on 28 of the 40 files, which is the source of every range and every sub-40 cover in the tables above.

For the slot-by-slot rewrite between these same four shapes, see what carries over when you translate a line. The colour word inside each of those lines is generated by a separate lookup: how the tool names a colour from thirty-five anchors explains why a red object in a grey frame loses its name. For the measurements behind every phrase, see what an image to prompt tool measures, and for the values at which each phrase flips, the twenty-nine cut points.