About the Image to Text (OCR)
Copy the text out of a photo, scan or screenshot. Choose an image (or paste a screenshot with Ctrl/Cmd+V), pick the language, and the text appears in a box you can copy or download as a .txt file.
Recognition runs on your device with the open-source Tesseract engine. The image is never uploaded; the only download is the language model (a few MB), which comes from this website the first time you use a language.
How it works
Optical character recognition (OCR) finds lines of text in the picture and recognises each character with a neural network trained for the chosen language. VelixaKit uses tesseract.js 7, a WebAssembly build of Tesseract 5.1, with Tesseract's "best" language models in their integer (faster) form.
The text is then cleaned up: line endings are standardised, repeated spaces are collapsed and accented letters are stored in their standard Unicode form, so é is one character rather than e plus an accent.
OCR reads printed text well and handwriting poorly. It can't guess words that are cut off, covered or out of focus, so always check the result before you rely on it.
Assumptions and limits
- Languages: English, Spanish, French, German, Italian, Arabic and Urdu. Choose the language of the text; the wrong language model gives many more mistakes, especially with accents.
- Works best on sharp, evenly lit, straight-on images where letters are at least about 12 pixels tall.
- Handwriting, decorative fonts and textured or patterned backgrounds are not reliably recognised.
- JPG, PNG and WebP images. iPhone HEIC photos need converting to JPG first.
Worked examples
Copying a quote from a screenshot
Take a screenshot, open this page and press Ctrl+V (Cmd+V on a Mac). Choose English and select Extract text. Clean screenshot text is usually recognised without errors; in our tests, nine clean sample images in four languages were read with no character errors.
Reading a photographed receipt
Photograph the receipt flat, in good light, filling the frame. Small print near the edge of readability is still read well (one character wrong in 52 in our tiny-text sample), but crumpled paper or shadows can introduce mistakes in prices, so compare the numbers with the receipt.
When OCR struggles
On a sample written in a handwriting-style font, about 1 character in 5 was wrong, and on text printed over a speckled paper texture most characters were wrong. For these, retake the photo with a plain background and more light, or type the text instead.
Measured accuracy on our test images
Character error rate = wrong, missing or extra characters ÷ characters in the original. Measured on 19 generated test images on 13 October 2026; real photos vary.
| Test image | Character error rate | What it means |
|---|---|---|
| Clean printed text (12 images, all 7 languages; Urdu in Naskh style) | 0% | Every character correct |
| Urdu in Nastaliq style | 8.3% | 2 wrong characters in 24 |
| Slight blur, 4° rotation, heavy JPEG compression | 0% | Every character correct |
| Very small (11 px) text | 1.9% | About 1 wrong character in 50 |
| Handwriting-style font | 22% | About 1 wrong character in 5 |
| Text on a speckled paper texture | 77% | Mostly unreadable |
How to use the Image to Text (OCR)
- Add an image. Choose a JPG, PNG or WebP file, drag it onto the page, or paste a screenshot.
- Pick the text's language and extract. Select the language the text is written in, then choose Extract text. The first run downloads the language model.
- Check, then copy or download. Read through the result for mistakes, then copy it or save it as a .txt file.
Frequently asked questions
How accurate is image to text?
On clean printed text it is very accurate: our clean test images were read with no errors. Accuracy falls with blur, tiny letters, handwriting and busy backgrounds; a handwriting-style font had about 22% of characters wrong. No OCR is perfect, so always proofread the result.
Is my image uploaded anywhere?
No. The recognition runs in your browser on your own device, and the image never leaves it. The first time you use a language, its model file (a few megabytes) is downloaded from this website, then your browser can reuse it.
Can it read handwriting?
Only neat block capitals, and not reliably. The language models are trained mainly on printed text, so joined-up or casual handwriting produces many mistakes. For handwritten notes, typing them out is usually faster than correcting the OCR result.
Which languages are supported?
English, Spanish, French, German, Italian, Arabic and Urdu. Pick the language the text is written in before extracting: using the English model on French text, for example, often loses accents and misreads words.
Why is the text in the wrong order?
OCR reads blocks of text roughly top to bottom, left to right. Multi-column layouts, tables and text wrapped around pictures can come out in an unexpected order. Cropping the image to one column at a time usually fixes it.
Sources
Data last checked: 2026-10-13.
This tool runs entirely in your browser. What you enter or upload stays on your device and is never sent to our servers.