Image to Text (OCR)
Extract text from an image — Korean or English screenshots, scans and photos — with OCR that runs entirely in your browser, no upload.
Drop an image here, or click to choose (PNG, JPG, WebP…)
Downloads about 10 MB the first time you use this language, then keeps it.
🔒 Text is extracted in your browser — your image is never uploaded. Korean is read by PaddleOCR (PP-OCRv5); the Korean model also covers the Latin letters and numbers mixed into a Korean page, so 한국어 and 한국어 + English load the same models. English on its own uses Tesseract, which is a much smaller download. Each engine is fetched once and then cached. Before reading, Clean up the picture takes the grain out of a scan or photo and pulls a faded page back to full contrast — a clean screenshot is measured and left exactly as it is. Straighten a tilted page measures the tilt of the lines of text and turns the picture level first; it is offered on the English path because the Korean engine already finds each line and straightens it as it reads.
Turn an image into selectable text — in Korean or English
Drop a screenshot, scan or photo, pick the language, and this tool reads the text in it and hands you plain, copyable text. Two recognition engines sit behind the Language menu. 한국어 and 한국어 + English both run PaddleOCR's PP-OCRv5 models, which read Hangul properly; the Korean model's character set already covers Latin letters and digits, so those two choices load exactly the same models rather than two different ones. English on its own stays on Tesseract, the long-standing open-source engine, because it is a far smaller download. Everything runs on your own device either way — the image is never uploaded to a server.
What the first run downloads, and how well it reads
The engine you choose is fetched once and then kept by your browser: about 32 MB for Korean, about 10 MB for English. After that first download the tool starts immediately and keeps working offline, and because the recognition happens on your device even a private document never leaves your computer. Accuracy is best on clear, high-contrast print — screenshots, exported PDFs, clean scans — and the picture no longer has to arrive that way. Clean up the picture measures the grain, filters it out when there is any, and pulls a faded page back to full black-on-white contrast before the model sees it; a screenshot that is already clean is measured and passed through untouched. Straighten a tilted page measures the tilt of the lines of text and turns the picture level first, and appears on the English path because the Korean engine already finds each line and straightens that line as it reads. Both are on by default. Blurry photos, low resolution, unusual display fonts and handwriting are still harder, and the tool shows a confidence figure with every result so you can judge it.
Frequently asked questions
Can it read Korean?
Yes. Choose 한국어 or 한국어 + English and recognition runs on PaddleOCR's PP-OCRv5 models, which are built for Hangul. Those two choices load the same models, not two different ones: the Korean model's character set already includes Latin letters and digits, so it reads the English words and numbers mixed into a Korean page without anything extra. The models are served from this site and run on your device.
Is my image uploaded anywhere?
No. The image is read and recognized entirely in your browser with WebAssembly — PP-OCRv5 for Korean, Tesseract for English. It never leaves your device and the tool works offline after the first load.
Why is the first run slow?
It downloads the recognition engine for the language you picked and caches it: about 32 MB for Korean, about 10 MB for English. That happens once, with a progress bar. Every run after that starts straight away.
What images work best?
Clear, high-contrast printed text — screenshots, exported PDFs, clean scans. A scan or a photo does not have to arrive that way, though: "Clean up the picture" filters the grain out and opens the contrast back up before reading, and "Straighten a tilted page" turns a crooked page level first on the English path (the Korean engine straightens every line itself as it reads). Both are on by default, and a picture that needs neither is passed through untouched. Blurry photos, low resolution, unusual fonts or handwriting still reduce accuracy.