Reading the percentage
The figure on each pair is the share of the marks on the page
that differ, not the share of the sheet. A page of text is almost
entirely whitespace, so measured against the sheet two completely
different documents would score under one per cent and look
identical. Measured against ink, that same pair scores in the
nineties — which is the truth, and not a sign that anything
went wrong.
How pages are paired
Two arbitrary files need not agree about page numbers. Each page is
fingerprinted by the structure of its ink, and the two sequences are
aligned in reading order, so an inserted or deleted page shifts the
rest rather than throwing it off. Where that guess is wrong, set a
fixed offset instead.
What is compared
Rendered pixels, not text. Anything that draws is compared, including
figures and scanned pages. Annotations are not drawn: comments and
highlights are marks left on a document rather than part of it.
Nothing is extracted, and nothing is changed — both files are
only ever read.
Privacy
Files are read and rendered entirely in the browser. There is no
server, no upload and no telemetry — the page works offline
once loaded.
This build
- Version
- —
- Rendering
- pdf.js 6.2 · WASM
- Licence
- Not yet chosen