Screenshot to Text (OCR)

Recognition runs on Tesseract compiled to WebAssembly, executing in a worker thread inside your browser. The first run downloads the trained data for the language you picked and caches it, so the initial recognition is noticeably slower than every one after it. No image is uploaded — the engine comes to the file rather than the file going to a server.

Screenshots are close to ideal input. The text was rendered by a computer rather than photographed, so it is perfectly sharp, perfectly aligned, and evenly lit — none of the problems that come with a phone photo of a page apply.

The exception is scale. A screenshot taken on a standard-density display captures interface text at roughly the pixel size it was drawn, which can be small enough to leave few pixels per character. Capturing on a high-density screen, or zooming the page in before capturing, gives the recogniser materially more to work with. Dark mode is fine; the engine handles light text on dark backgrounds.

The language setting matters more than people expect, because it is not a hint. Tesseract loads a different trained model per language, and each one encodes that language's character set and letter-sequence probabilities. Running Portuguese text through the English model does not degrade gracefully; accented characters have no good match and the engine's guesses about likely letter sequences are actively wrong.