Where it works well
- Screenshots of applications, dialogs, error messages and terminals
- Printed documents scanned straight on
- Slides and diagrams with real text
- Tables, though the layout will not survive
Where it does not
Handwriting, stylised or decorative type, text over a busy photograph, low contrast, and anything photographed at an angle. These are limits of the approach rather than of this particular tool — the engine here is the same one most browser-based OCR uses.
Choose the language first
The language model is what turns shapes into characters, so a Cyrillic page read as English produces plausible-looking nonsense rather than an error. For a document mixing Latin and Cyrillic, pick the combined option — it is slower and substantially more accurate than either alone.
Always re-read the output
OCR does not know when it is wrong; it returns its best guess with a confidence figure attached. The recognised text is editable here for exactly that reason — fix it while the image is still in front of you rather than after it is pasted somewhere else.
Questions
Does my image get uploaded?+
No. Recognition runs on WebAssembly in this tab. The one network request is for the engine and the language model, which download once and are then cached — worth saying plainly rather than claiming everything is local.
How accurate is it?+
Good on clean screenshots and printed text, poor on handwriting, low contrast and photographs taken at an angle. Always re-read the output: OCR confuses similar glyphs, and the classic substitutions — 0 for O, 1 for l, rn for m — are the ones that survive a quick glance.
Which languages work?+
English, Russian and Kyrgyz, plus English and Russian together for mixed documents. Picking the wrong language is the most common cause of nonsense output — a Cyrillic page read as English produces confident gibberish.
How can I improve the result?+
Give it more pixels and more contrast. A screenshot at native resolution beats a photograph of a screen; a straight scan beats an angled phone shot. Cropping to just the text block before recognising also helps, because there is less for the layout analysis to get wrong.