Extract text from images (OCR)

Printed text in clear screenshots and scans is usually read almost exactly. Handwriting, decorative or very small text, blurred or dark photos, and complex tables are harder, so check the result before you use it. You can correct the text in the box before copying or saving it.

Yes. Vertical lines are read top to bottom, starting from the right. Punctuation that turns in vertical writing, such as 、 。 「 」, may be missing.

The first time, your browser downloads the text recognition models (30 to 40 MB). They stay in the browser’s cache, so later images start right away, even offline. On devices without WebGPU, the models run on the processor and an image can take 10 seconds or more; keep this tab in front while it works.

English, Japanese and Chinese (Simplified and Traditional) are read by the same model, so whichever of them you choose, text that mixes them is read as it is. Korean is read by its own model: choose Korean for Korean text, and one of the others for everything else. Other languages written in the Latin alphabet may be read only in part.

No. Everything runs in your browser, so your files never leave your device. The page only downloads the tool's own code when you add the first file.