Image to Text (OCR)
Pull the text out of screenshots, photos and scans. Works with English, Simplified Chinese, numbers and mixed text. Recognition runs in your browser, so your images are never uploaded.
Note: the first time you run it, the page downloads about 21 MB of model and engine files, with a progress bar above. They are stored in your browser's cache, so later visits start almost instantly. Speed depends on your device; a phone screenshot usually takes 1–3 seconds on a typical laptop.
How to use
- Click the box above to pick images, drag them in, or take a screenshot and press Ctrl+V (⌘+V on a Mac) anywhere on the page.
- On first use the recognition model downloads first; recognition starts automatically when the progress bar finishes.
- The text appears in the box below. Fix any mistakes right there, then click Copy text to paste it into an email, document or spreadsheet, or Download TXT to save it.
- For articles and documents, choose Join into paragraphs. For chats, tables and lists where each line matters, choose Keep original line breaks, then select the image again.
How accurate is it?
The tool runs the open-source PaddleOCR PP-OCRv5 mobile model (from Baidu's PaddlePaddle project) through onnxruntime in your browser. Before launch we compared it on web page screenshots, a phone settings screen, a chat screenshot, an order slip mixing Chinese and English, and a blurry, noisy scan. Clean screenshots came out essentially word-perfect, including numbers and punctuation. On the noisy scan with small type it confused a few look-alike characters, which is typical for any OCR; cropping or enlarging the image helps. We also tested the popular Tesseract engine, which skipped whole lines in the chat screenshot and inserted stray spaces between Chinese characters, so we chose PaddleOCR.
Why on-device OCR is more private
The images people need to read are often contracts, receipts, IDs, chat logs and work documents. Most online OCR services upload them to a server, and you cannot see what happens to them afterwards. Here the model is downloaded into your browser and runs on your own device. Images are read only in local memory, the text appears only on this page, and nothing is sent to us or anyone else. Once the model has loaded, recognition keeps working even if you go offline.
Tips for better results
- Use the original image or screenshot rather than a copy that has been compressed by a messaging app. When photographing, keep the page flat, evenly lit and in focus.
- If the text covers only part of the picture, crop to that area first. It is faster and more accurate.
- Very small text (a whole page shot from a distance) benefits from enlarging to 150–200% with the Image Resizer first.
- Printed, horizontal text works best. Handwriting, decorative fonts and vertical text are less reliable, so proofread the result.
- To read text from a scanned PDF, export its pages with PDF to Image and recognize them one by one.
FAQ
Q: Are my images uploaded?
A: No. The recognition model runs inside your browser on your device. Neither the images nor the extracted text are sent to us or any third party.
Q: Why does the first run take a while?
A: The first run downloads about 21 MB of model and engine files and shows the progress. They are kept in your browser cache, so later runs load them locally without downloading again.
Q: Which languages are supported?
A: English and Simplified Chinese, plus numbers, common punctuation and mixed text. Many Traditional Chinese and Japanese kanji characters are also recognized, but accuracy is not guaranteed.
Q: Can it read handwriting and tables?
A: Neat handwriting is partly recognized; messy handwriting produces many errors. Tables are output line by line without their structure, so choose Keep original line breaks and tidy the result in a spreadsheet.
Q: What if the result has mistakes?
A: Edit them directly in the text box before copying. Many errors usually mean the image is blurry or the text is too small; crop, enlarge, or use a sharper original.
Q: Does it work on phones?
A: Yes, in modern mobile browsers. The model is about 21 MB, so load it on Wi-Fi the first time. Phones are slower than computers; a screenshot typically takes a few seconds.