Image to Text — Free Online OCR

OCR stands for optical character recognition. The engine finds the shapes in your picture, matches them against a trained model for the language you picked and writes out the characters it recognizes. All of that happens inside this page. LeanImg uses Tesseract, the open source OCR engine, compiled to WebAssembly so it runs on your own hardware instead of a server.

No. The only requests the extractor makes are for the OCR engine and the language model, both served from this site, and they travel in one direction. Your image is read from disk into the page, handed to the engine in memory and turned into text there. There is no upload endpoint in the codebase to receive it. That matters when the picture is an ID card, a payslip or a signed contract. Open your network tab during a run: you will see model files arriving and nothing carrying your image away.

JPG, PNG, WebP, GIF and BMP go straight to the engine. AVIF and HEIC are decoded in the page before the engine sees them, so a photo straight off an iPhone does not need converting first. Whatever goes in, plain text comes out.

English, Spanish, Romanian, Arabic, Greek, Turkish, Indonesian, Russian and Japanese. Pick the language before you extract, because the engine matches shapes against the model you choose and the wrong model reads badly. The picker starts on the language of the page you are on. There is no automatic language detection here yet.

Because the recognition happens on your device, so the model has to get there first. The first time you use a language its data file downloads and the browser caches it, and later runs in that language start immediately. That is the trade for privacy. A hosted OCR service uploads your image and keeps the model on its own machine. This one keeps your image and fetches the model instead. Switching to a language you have not used yet downloads that model too.

Yes, free with no account and no usage counter. There is no sign-up, no watermark on the text, no credits and no paid tier holding back the good version. The work runs on your hardware, so there is no per-image server cost for us to pass on to you.

It depends on the input, and the difference is large. Screenshots and clean scans read very well, because the characters are sharp, evenly lit and sitting straight. Photographs of paper are less predictable, and they get worse with blur, glare, shadow or a tilted angle. There are cases we will not pretend to handle: handwriting and text in a scene such as a shop sign or a label on a shelf. This engine is built for printed characters. Other sites advertise perfect accuracy on all of it. That claim is not true of any OCR engine, including this one, so read the output before you rely on it.

You can select a batch, and the dropzone shows the limit for a single run. Images are processed in sequence rather than in parallel, and each result appears as soon as it is finished, so you can start reading and copying while the rest are still working. Sequential processing also keeps memory use predictable on phones.

The engine reports how sure it is about each word it reads, and the result card shows the average for the image. When that average falls below the threshold we set, you get a warning instead of a clean-looking result. It is not a verdict that the text is wrong. It is a signal to check it. In practice it fires on blurry captures, low contrast, very small text and the wrong language selection.

Give the engine bigger, sharper characters. Screenshot the source rather than photograph it whenever that is possible. If you have to use a camera, shoot straight on rather than at an angle, fill the frame with the text and keep the lighting even so no glare or shadow crosses the page. Dark text on a light background reads best. Pick the right language before you extract, since that choice changes the result more than anything else on this list.

Yes, and a screenshot is the best case for this tool. It is rendered text: crisp edges, flat background, no lens and no lighting to fight. That is exactly what a printed-text OCR engine is built to read. Screenshots of error messages, chat threads, app dialogs and terminal output all come through well. Table layout is not reconstructed, so you get the words, not the grid.

Copy it to your clipboard, or download it as a .txt file. When you process a batch you can also download everything as a single .txt. There is no searchable PDF output here, and we would rather say that plainly than imply it: this tool gives you plain text, not a rebuilt document with the original layout. If you need the words in a spreadsheet, a document or a translator, paste them there.

Yes. Drop a PDF into the same box you would drop an image into. A digital PDF, the kind software produced, already carries a text layer. That layer is read directly, so the text comes out exactly as it was written with no OCR involved. A scanned PDF has no text layer, only pictures of pages, so each page is rendered and run through OCR the way a screenshot is. Either kind is read up to 25 pages per PDF, and nothing is uploaded: the file is opened inside this page and stays on your device.

Because the file mixes two kinds of page. A digital report with a photographed signature page inside it, or a scan with a generated cover sheet on the front, carries a real text layer on some pages and nothing but an image on the rest. Pages with a usable text layer are read verbatim. Pages without one are rendered and sent through OCR, which reads well on clean scans and less well on crooked or blurry ones. The summary line on the result says which pages came from the text layer and which went through OCR, so you know where to check the output.