Image to text
Pull the words out of a photo, a scan or a screenshot, then copy them or save them as a text file. Your images never leave your device.
Drop images here or click to choose
JPEG · PNG · WebP · BMP — read on your device, never uploaded
The OCR engine and the English model — about 4 MB — download from shrink.run the first time you use this page, then stay cached in your browser.
How to turn an image into text
- Drop your images into the box above, or click it to pick files.
- Wait while the engine loads. The first run downloads about 4 MB — the recognition code and the English model — from shrink.run; after that your browser has them cached.
- Read the result in the text box. It is editable, so you can fix a stray character or delete a header before you take it.
- Press Copy text to put it on the clipboard, or Download .txt to save it as a file next to the original.
What OCR actually does
A photograph of a page is just coloured dots. Nothing in the file says "this is the letter e" — that meaning only exists in your head when you look at it. Optical character recognition is the software that puts it back: it finds the lines of text, cuts them into characters, and matches each shape against a model trained on millions of examples of printed type.
The engine here is Tesseract, the open-source recogniser Google maintained for years, compiled to WebAssembly so it runs inside your browser at close to native speed. It is genuinely good at what it was built for — clean printed text — and genuinely bad at everything else, which is worth knowing before you trust a result.
| Source | Typical outcome |
|---|---|
| Flatbed scan of a printed page | Near-perfect |
| Screenshot of a website or app | Near-perfect |
| Straight phone photo of a document | Good, with the odd wrong character |
| Photo taken at an angle, or in shadow | Patchy — lines get lost |
| Table, form or two-column layout | All the words, none of the structure |
| Handwriting | Not supported |
Getting a better result
OCR quality is decided almost entirely by the picture, not the software. Four things make most of the difference:
- Shoot it square. Hold the camera parallel to the page rather than leaning over it. Curved or slanted lines are what breaks recognition most often.
- Fill the frame. Letters need to be roughly 20 pixels tall before the engine can tell them apart. A page photographed from across a desk will not read; the same page filling the viewfinder will.
- Light it evenly. A hard shadow or a bright reflection across the text costs you that whole line. Diffuse daylight beats a phone flash.
- Hold still. Motion blur softens the letter edges the engine matches against, and a blurry photo cannot be rescued afterwards.
Two of our other tools genuinely help before you run OCR. Cropping down to just the block of text removes the desk, the margins and anything else the engine would try to interpret, and rotating a sideways or upside-down scan upright matters because Tesseract reads left to right and does not turn the page for you. Both run on your device too, so nothing about the document goes anywhere.
Why this one does not upload
Think about what people actually put through an OCR site: a payslip, a tenancy agreement, a doctor's letter, a passport page, an invoice with a bank account on it. The standard free OCR service takes that file, uploads it to a server, reads it there and hands the text back. Whatever the privacy policy says, the document has left your machine and you have no way to check what happened to it.
shrink.run does the recognition in the page instead. The engine, the WebAssembly core and the English model are all served from this domain — we deliberately do not load them from a third-party CDN, because even that request would tell someone else that this device is running OCR right now. Once those files are cached, the page works with the network switched off, which is the simplest proof there is that nothing is being sent.
You can check it yourself: open your browser's network panel, drop an image in, and watch. You will see the engine files come down from shrink.run once, and no request go out with your picture in it. How this works, in more detail.
Frequently asked questions
Is my image uploaded to a server?
No. The recognition engine is a WebAssembly build of Tesseract that downloads to your browser from shrink.run and runs on your own device. The picture is decoded in the page and never sent anywhere. This matters more here than on most tools: the images people run through OCR are receipts, contracts, ID cards, prescriptions and bank letters, and every other free OCR site we are aware of uploads that file to its own servers to read it.
Which languages can it read?
English only, for now. Each language needs its own trained model — roughly 3 MB per language — and we ship the English one rather than making every visitor download a menu of them. Latin-alphabet text in other languages will partly come through, because the letter shapes are shared, but accented characters and the dictionary-based corrections will be wrong. More languages may follow if people ask for them.
Can it read handwriting?
No. Tesseract is trained on printed and typeset text. Handwriting, signatures and handwritten form fields come out as nonsense, and no setting fixes that. Neat block capitals sometimes partly work, but it is not something to rely on.
Why is the first run slow?
The engine and the English model are about 4 MB together, and they download the first time you use the page. Your browser then caches them, so the second image — and every image on a later visit — starts reading almost immediately. Recognition itself usually takes a few seconds per page on a laptop and rather longer on an older phone.
What kind of image gives a good result?
A flat, straight, sharp image where the text is dark on a light background and each letter is at least about 20 pixels tall. Flatbed scans and screenshots are close to perfect. A phone photo works well if you hold the camera square to the page, fill the frame, and have even light with no shadow across the text.
Why did it get my document wrong?
Almost always one of five things: the photo is blurry, it was taken at an angle so the lines curve, the light is uneven, the text is too small in the frame, or it is a low-contrast design — light grey on white, or white on a photo. Cropping down to just the text and rotating the image upright before you run it fixes a surprising number of bad results.
Does it keep the layout of the document?
Only loosely. You get the text in reading order with line breaks, which is what you want for a letter, a receipt or a page of prose. Tables, multi-column layouts and forms lose their structure — the words are all there, but the columns run together. The text box is editable, so you can tidy it up before copying.
Can I read a PDF this way?
Not directly — this tool takes images. Turn the pages into images with our PDF to JPG tool first, then drop those here. Bear in mind that many PDFs already contain real text you can select and copy in any PDF reader; OCR is only needed when the PDF is a scan.
Is there a limit on file size or how many images I can read?
No artificial limit. Drop a whole batch and they are read one after another, each with its own text box. The real limit is your device: a very large photo takes longer and uses more memory, and a phone will give up sooner than a laptop.