Read a page with Tesseract, on your own GPU

Drop a scan, a photo or a PDF. Each text line is flat-fielded and cropped tightly to its ink, and Tesseract's LSTM recognizer, ported to WebGPU, reads every line in one batch, in any of the 320 LSTM models in tessdata_fast and tessdata_best (about 160 languages and scripts, each in both sizes; the legacy-only equ and osd models are not offered). Math is read as plain text. The page is not uploaded anywhere. Models load on first use from this site, jsDelivr and Tesseract's public repository. It needs a browser with WebGPU; without one the page shows a message and does nothing else.

Starting…

fast: Google's 8-bit model, about 1 to 5 MB a language. best: the float model, about 1.5 to 15 MB, slightly more accurate.