Image to Prompt Round Trip: What the Line Brings Back

Two answers, and they disagree. Measured in numbers, a generated line gives a lot back: on long lines it recovers brightness to within 3.9% of its range and colour warmth to within 1.5%, and it points at the right file out of forty on 54 of 160 tries — 33.8%, against 2.5% for guessing. Measured in pixels, it gives back close to nothing: hand someone all eight numbers, exactly right, and the nearest picture they can find among thirty-nine candidates is still 61.07 grey levels per pixel from the original — no better than the 62.14 you get by averaging the other thirty-nine files and using no information at all. And the line does not ask for much to begin with: with all twelve lines the tool writes held character-for-character fixed, the median photograph can be moved 56.88 grey levels per pixel away from itself, 74.3% of the 76.55 grey levels that separate two unrelated photographs in the same set. All four of those numbers, and the loop they were measured on, are below.

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 (36,864 cells whatever the file size), and composed into four shapes at three lengths: 480 lines. Every number below comes from re-running that pipeline, not from a description of it. No image generator was run and no image was generated — there is no image model behind this site and none was called — so what follows measures the round trip at the number level, and states plainly where it stops.

What a line hands back, one number at a time

Eight measurements sit behind the words. To ask what a line gives back, invert it: for each measurement, sweep the whole axis and keep every value at which the composed line stays character-for-character identical. What survives is an interval, and the honest reading of that interval is its midpoint. Compare that midpoint against the true value and you have a recovery error; compare it against a blind guess — the median of the same measurement across all forty files, which is what you would say with no line at all — and you know whether the line was worth reading.

Five of the eight beat a blind guess on more than half the lines, and brightness is the clearest case: a blind guess is 14.7% of the axis away, and reading dark, low-key lighting or bright, high-key lighting off the line leaves you 3.9% away, with 132 of 160 lines improved. Warmth is tighter still at 1.5%. Three axes lose, and they lose for one reason: their intervals run off the end. monoRatio reaches an end of its axis on 160 of 160 lines, which means vivid, saturated colour is a floor with no ceiling — it tells you the picture is not grey and nothing about how far from grey it is. A midpoint taken across a half-open interval sits in the middle of the whole axis, which is why reading the line can be worse than knowing nothing.

Length buys specificity here too, and only here. On edge, a short line leaves the median error at 35.7% because the detail clause is not in it at all; the same measurement read off a long line is 22.6%. On meanSat the short line gives 13.4% against 7.0% at medium and long. Brightness barely moves, from 5.5% to 3.9%, because the exposure clause is in every line at every length.

Which of the forty it points to

The eight numbers describe a point in an eight-dimensional box. Take the midpoints, normalise each axis to its own span, and ask which of the other thirty-nine files sits nearest: if the line carries the picture, the right file should come out first.

A long line is worth 13.5 times a guess: 33.8% against 2.5%, with the true file in the top three 51.25% of the time and a median rank of 3. It also fails completely on some files — 40 is last place, and some long lines push the recovered point to the far side of the whole set. That failure is not random. It happens where the line is silent: a file whose spread, edge density and both lighting ratios all run off their axes has only brightness, saturation and warmth left to steer by, and those three do not separate forty photographs.

Set against that, the line is almost a fingerprint. Of the 780 pairs of files in this set, exactly one produced the same line as its partner at any length — two short lines shared by p15 and q776 — and at medium and long, none did. No two of these forty files are described identically. They do not have to be, because the description does not reach very far.

Baselines: from no picture to the real one

One more way to put a size on it. Suppose you had to produce the photograph rather than describe it, and the only thing you had was the line. Here is how close each starting point gets, again in median grey levels per pixel over the forty files.

The row to read is the third one. Hand someone every one of the eight numbers, exactly right, and the closest picture they can find among thirty-nine candidates is still 61.07 grey levels per pixel away — statistically indistinguishable from the 62.14 you get by averaging the other thirty-nine files and using no information at all, and better than that average on only 19 files out of 40. Knowing the measurements perfectly is not the same as knowing the picture, and that is the ceiling any round trip has to work under: no line, however long, can carry more than the eight numbers that produced it. A long line read honestly does slightly better than perfect knowledge here, 58.98, but that is retrieval noise, not signal — the recovered point is not the true point, and on 20 files it happened to land nearer.

The floor is worth stating too. The two most similar files in this corpus are 23.57 grey levels apart, so no method that has to pick one of thirty-nine real photographs can beat that. The gap between 23.57 and 61.07 is what the eight numbers cannot say, and it is most of the picture.

How far the picture can move while every line holds still

So instead of asking what the line gives back, ask what it never asked for. Starting from each real file, we changed the pixels — swapping rectangles, lifting and dimming them, mirroring inside them, moving them elsewhere — and kept every change that left all twelve composed lines character-for-character identical. The distance reached, in mean grey levels per pixel over 36,864 cells, is how much of the photograph the line simply does not mention.

Six files could be moved less than 20 grey levels; 12 landed between 20 and 50; 13 between 50 and 70; and 9 moved 70 or more, the worst of them 87.90 — further apart from its own original than 76.55, the median distance between two unrelated files in the set. The single pair that shared a line, p15 and q776, differs by 68.41 grey levels per pixel with a grid correlation of −0.14: two files the tool described the same way, agreeing on nothing about where the light and dark actually fell.

The measurements moved underneath the words while this happened. Across the forty files, brightness shifted by a median of 6.43 grey levels (most, 22.86), spread by 4.93 (22.29), and edge density by 48.94 (79.38) — on measurements whose true values in this set run from 3.71 to 65.84. Every phrase in every line stayed the same, because those phrases are one-sided comparisons and the numbers ran away in the direction the phrase does not test.

That is the difference between this search and the shuffle on how much of your picture survives, which swaps pixels only inside the six regions the measurement actually distinguishes. There, eight of the nine numbers cannot move at all by construction, the grid still changes by 50.00 grey levels on the 22 files whose lines survive a full scramble, and the only number that moves is edge — which is why that page finds the walk stopping at the detail threshold. Here the regions were not preserved, so brightness, spread and edge all moved and the search was free to climb further; 56.88 is what a line permits when the eight numbers are allowed to drift inside their buckets, not a correction of that other number.

What the sixty thousand proposals were allowed to touch

Forty files, 1,500 proposals each: 60,000 in total, of which 10,373 were accepted — a median of 313 per file. A proposal was one of four moves on the 192 × 192 grid: swap two equal rectangles, add or subtract up to 40 levels inside one, mirror one left-to-right, or move one elsewhere. It was accepted only if all twelve composed lines — four shapes at three lengths — came back character-for-character identical to the lines of the untouched file; otherwise the grid was restored. Moves were kept when they increased the distance from the original, or, on one proposal in ten, regardless, so the search could walk out of a corner it had climbed into. At the end, every modified grid was re-measured and re-composed from scratch: 40 of 40 still produced the original twelve lines.

What this number is, precisely: a lower bound found by search, not the size of the set of pictures a line permits. A better search would move further, so 56.88 is the least the line leaves open, not the most. And it is measured on a grid of 36,864 cells, which is the resolution this tool actually looks at; at full resolution the same line is compatible with more pictures still.

Three limits belong next to these numbers. No generator was run, so nothing here says what any image model would draw from these lines — only what this tool claims about a file and how much of the file that claim fails to pin down. The eight measurements are this tool’s choice, not a description of images in general: a tool with different measurements would draw its line somewhere else. And the corpus is forty stock photographs plus one screenshot, so the distances are distances inside that set; on a set of near-duplicates every number here would shrink. What survives all three caveats is the shape of the result: the words are specific about brightness and warmth, silent about spread and near-monochrome, and compatible with pictures that differ from the original by more than two unrelated photographs differ from each other.

For the twelve measurements behind the phrases, see what an image to prompt tool measures; for the cut points where each phrase flips, the twenty-nine cut points; for how much of a picture survives a change to the file rather than to the description, how much of your picture survives; and for what to do once you have the line, writing the rest of the prompt yourself.