The seam is measured, not assumed
Every other way of putting scrolling captures together stacks them and hopes. Where the captures overlap — and they always overlap, because you scrolled less than a full screen between them — the result carries the same three lines of text twice, and the only way to find out is to read the finished picture. This one measures the overlap before it stacks anything, tells you what it found, and cuts nothing until you accept it.
- Furniture first. A band that is identical in every capture is not content. A status bar, a sticky header, a tab strip at the bottom — each is recognised as furniture and left out of the matching, because otherwise it matches everywhere and tells you nothing.
- Then the slide. The bottom strip of one capture is slid against the top of the next and scored at every position. The best-scoring position is the overlap, and its score is the number printed beside the seam.
- Then the doubt. If a second position scored almost as well as the winner, the page repeats itself and the seam is flagged rather than guessed. That flag is the thing no stitcher without a detector can show you.
What the numbers beside each seam mean
A seam that reads matched 94% is saying that the rows about to be cut are a 94% correlation with the rows above them. It is evidence, not a promise: a correlation is a measurement of similarity and a page can be similar to itself in more than one place. A seam that says two positions scored almost the same is the detector declining to break a tie it cannot break. Every proposal waits for Accept — one at a time, or all the confident ones at once — and is drawn on the picture as a dashed band until then. Every seam stays nudgeable afterwards, remembers that you moved it, and the Seams step carries a count of the proposals still waiting.
Where it will not help
Stitching works when the captures are the same width and the content sat still. A header that changes as you scroll, a video playing mid-page, an advert that rotates between captures, or two captures taken at different zoom levels all give the detector something that genuinely is not the same content — and it will say so rather than pretend. Nudge the seam and you will beat it in about five seconds.
Questions about captures
How does it know where one capture ends and the next begins?
It measures. Each capture is reduced to a row-by-row brightness profile, the bands that are identical in every capture — a status bar, a navigation strip — are set aside as furniture rather than content, and the bottom of one capture is slid against the top of the next until the profiles line up best. That best position, and how well it scored, is what the seam panel prints.
What does the percentage next to a seam actually mean?
It is the correlation between the two overlapping bands at the position the detector chose: how closely the rows it is about to cut match the rows above them. High is good evidence and it is not a guarantee, which is why every seam can be nudged and why a seam you set by hand is never touched again.
Why does one seam say two positions scored almost the same?
Because the page repeats itself. A chat thread of similar bubbles, a table of similar rows, a list of identical cards — each of those gives the detector a second position that fits nearly as well as the first, and guessing between them is exactly how stitchers silently lose a paragraph. It is flagged instead, and the magnified joint shows you which of the two is right in about five seconds.
One of my captures is a different width and it was refused. Why?
Captures of one page scrolling are all the same width by construction. A different width means a different device, a different zoom level or a different window — and stacking it would either squash it or leave a step down the side of the finished picture. The file is named and left out; the rest carry on.
The finished screenshot is enormous. Is there a limit?
There is, and it is the browser's rather than ours. A canvas stops at about 32,767 pixels on a side on a desktop and about 8,192 on an iPhone, and six phone captures stacked can pass 15,000 without trying. The readout under the picture carries the height and this browser's verdict on it from the moment the second capture lands, so a refusal arrives while you can still do something about it.