Images

Image to text: read the words out of a picture

Read the words out of a photograph.

On this device

Choose a picture

A screenshot or a flat, well-lit scan gives the best result. The reader is downloaded the first time you use this tool, which takes a moment.

    Opened on this device · never transmitted

    What this tool does, exactly

    Optical character recognition, running here rather than on somebody's server: a neural detector finds every line in the picture and a neural recogniser reads it, on your graphics card where the browser offers one. It works well on screenshots and flat scans of printed type — a photo to text, a picture to text, a scan to text, free and with no page limit — less well on photographs taken at an angle, and it shows you which lines it was unsure of instead of hiding the doubt.

    What it does
    Finds every line of text in the picture, then reads each line character by character, and gives you the lines back in reading order as plain text and as Markdown. The models are PaddleOCR's PP-OCRv6, run with ONNX Runtime in this tab; measured on an Apple M4 Mac, an A4 page at 200 dpi reads in about 0.6 seconds on the graphics card, about two seconds on one processor thread in Chrome and six to eleven seconds on one thread in Safari.
    It does not read — it guesses well
    There is no understanding here, only shape-matching against a model of what printed text looks like. That is why the lines it was least sure of are listed rather than only one reassuring average. A wrong digit in an account number reads perfectly and is the error that costs the most.
    Which languages
    One model reads English, French, German, Spanish, Portuguese, Italian, Dutch, Polish, Turkish, Vietnamese and thirty-seven more Latin-script languages, plus Chinese and Japanese. Russian, Ukrainian and Belarusian are a second choice in the options, because their recogniser is a separate eight-megabyte download and most visitors should not pay for it.
    What it needs from you
    A picture where the text is upright, in focus and reasonably large. A screenshot is close to ideal. A flat scan is nearly as good. A photograph taken at an angle, in poor light, or of a page that is curved will produce noticeably worse results, and the confidence figures will say so.
    What it cannot do
    Handwriting has not been measured, so it is not promised: the models are trained on print. An old-style zero — the kind that looks like a lower-case o — is read as the letter. Tables come back as lines of cells rather than as a grid, and formulas as their words.
    What it costs to load
    About sixteen megabytes compressed the first time — the runtime, the reader and its two models — downloaded only when you open this tool and cached by your browser afterwards. It is the heaviest thing on this site, which is exactly why it is not loaded until you ask for it — except the Deep reader behind the Pro door under the button, 660 MB from Hugging Face, once; the AI OCR page, linked from the door, explains both readers.

    About this tool

    How accurate is it?

    On a clean screenshot or a flat scan of printed text, high enough to paste and skim: measured on a synthetic 200 dpi scan of an English contract, 0.2% of characters wrong, on Portuguese 0.7%, on Russian 0.6%. On a photograph taken at an angle or in poor light, low enough that you should read it against the picture before using it. The tool lists the lines it was least sure of, so the checking has somewhere to start.

    Can it read handwriting?

    It has not been measured on real handwriting, so this page does not claim it. The models are trained on printed type; on a cursive-shaped font — which is not handwriting — they got about 2% of characters wrong, ten times the print rate, and on a real hand they will produce confident-looking text that you must check against the picture. A page that promises handwriting will exist when there is a number behind it.

    Which languages does it handle?

    English and forty-six other languages written in the Latin alphabet, plus Chinese and Japanese, with one model. Russian, Ukrainian and Belarusian with a second recogniser you choose in the options, because it is an extra eight megabytes that most visitors should not download. Other scripts — Korean, Arabic, Greek, Thai, Devanagari, and Cyrillic outside those three languages — exist as models of the same family and will arrive the same way, as a choice.

    What runs on the graphics card?

    Both models, when the browser offers WebGPU — Chrome, Edge and Safari 26 do. Where it does not, the same models run on one processor thread with the same result: about two seconds a page in Chrome or Edge, six to eleven seconds in Safari, measured on an Apple M4 Mac. The result panel names which one read your picture.

    Can it read the text in a scanned PDF?

    Yes, and it has its own tool: OCR a PDF draws every page and reads it in one step, and gives the scan back as a searchable PDF as well as text. Use that one for a document; this one is for a picture you already have.

    Is my picture uploaded?

    No. The reader and its models are downloaded to this browser and the picture is read here; nothing about it leaves the device. The page can be watched doing exactly that in the browser's network panel: after the models arrive, no request carries your picture anywhere.

    More images tools · All 52 · How they work