ToolsTray

Star a tool to keep it here.

AI OCR — Image & PDF to Text

Extract text from an image or a scanned PDF, using a neural engine that reads 50 scripts and copes with a page shot at an angle.

Downloads the text recognition model the first time you run it, then works offline.

Files

Drop the pages here — images, PDFs or both — paste one, or choose files

JPEG, PNG, WebP, a PDF, or a mix · read in this order, back as one document

Recognized text

Each page lands here as it is read, opening with a seam that names its file and page.

Reads the words off a page and works out the script itself

Output

Before reading

Extracted

—

Nothing queued yet

Everything is read on this device — the only thing fetched is the engine. Extract text is held until a page exists.

How to use

  1. Drop your images and PDFs into the Files box. Mixing the two is fine, several at once is fine, and every page of a PDF gets read in turn.
  2. Look over the queue. It lists the files in reading order with each PDF's page count; Remove takes out a row you added by mistake, and Add files puts more at the end.
  3. Leave Engine as you found it unless you have a reason. The hint under it says why it was picked, and the line at the top states the download, whether it is already on this device, and, once a run has gone through on the processor, this device's own time per page.
  4. On Classic, set Language to match the document. The pack is fetched on demand and then cached. Sharp hides that field, because it identifies the script on its own.
  5. Untick Also create a searchable PDF if plain text is all you want, since it starts ticked. Tick Black and white first when the scan is faint or unevenly lit.
  6. Press Extract text. Cancel stops the run and keeps the pages already read. Afterwards, correct any page in its field, then copy the text, take the .txt, or download the searchable PDF.

About this tool

Text locked inside a screenshot or a scan is annoying in a very specific way. You can read the words. You just cannot select them. Point this at a JPEG, a PNG or a PDF and it extracts the text from the image and hands it back as something you can copy, edit, or save as a .txt. Leave the searchable PDF box ticked and your original pages come back too, with an invisible text layer laid over them, so find-in-page works on what used to be a flat picture. When a PDF already has selectable text, PDF to Text pulls it out without any recognition at all and is the faster route.

Two engines sit behind the Engine menu and the page picks one for you, saying why under the menu. Classic is Tesseract compiled to WebAssembly: six languages, a small download, dependable on a flat clean scan, and it reads a two-column page column by column. Sharp is PP-OCRv6, a neural detector and recogniser that covers about 50 scripts including Chinese, Japanese and Korean, works the script out by itself, and reads a page photographed at an angle that Classic turns into nonsense. It gets selected when your browser has WebGPU, for that reach rather than for speed: on an Apple M4 laptop a dense 1,500-word page took about 5 seconds on Sharp with WebGPU and about 6 on Classic. Without WebGPU, Sharp falls back to the processor and the same page took 28 seconds, so Classic stays selected there. Both engines execute locally; the only thing pulled over the network is the model, and the line at the top of the tool states its size before you press anything.

Sharp Pro is the third entry, a much heavier model that needs WebGPU and earns its place only when Sharp has already struggled with a particular document. On ordinary pages it reads no better. Everything else comes down to the scan itself, since these models are trained on upright, high-contrast, reasonably sharp text. Tick Black and white first for a faint or greyish original. Files are read in the order you added them, each page lands in its own field named for its file and page, and you can fix a misread word there before copying. A page with no text keeps its field and says so, and each row in the queue reports how many words came off that file.

Good to know

  • Handwriting is the honest weak spot: both engines are trained on printed and typed text, so neat block capitals sometimes survive and cursive comes back as gibberish. A two-column page is caught: Classic reads it column by column, and Sharp marks it and offers to read the columns one after the other, though an angled photo of a multi-panel layout such as a menu may still read straight across.
  • Classic only knows English, Spanish, French, German, Italian and Portuguese, so anything outside that list needs Sharp.
  • An image past 67.1 megapixels, 8192 × 8192 or the same area in any shape, is turned away when it is added. Pages are read at full size and that is the most Safari on an iPhone or iPad can draw, so a huge panorama or a stitched scan goes through the Image Resizer first.

Frequently asked questions

How do I make a scanned PDF searchable?

That is what the searchable PDF output does. Each page gets rendered, recognized, and rebuilt with the words placed invisibly over the part of the image they came from. The page looks exactly as it did before, but selection, copying and find-in-page all start working. It saves under your first file's name with a .pdf extension.

Which engine should I choose?

Start with whatever it picked; the hint under the menu says why, and the line at the top gives the download and, after a run on the processor, your device's time per page. Classic suits a flat, clean scan in one of its six languages and fetches under 7 MB with English. Move to Sharp for another script, or for a photo taken at an angle, and accept the roughly 30 MB model it fetches once. Sharp Pro is worth trying only after Sharp has already disappointed you on a specific document.

Does it read handwriting?

Not with any reliability. Careful block capitals sometimes come through, but joined-up writing usually returns something close to noise. Treat any handwritten result as a draft you will correct by hand. That is where optical character recognition stands generally, not a quirk of this page.

What does the black-and-white option actually do?

It runs Otsu thresholding across each page before recognition, turning a grey or unevenly lit scan into hard black text on white, which often lifts the result a long way. There is nothing extra to download for it. One catch on the Classic engine: with both boxes on, the page images inside the searchable PDF are the black-and-white versions rather than your originals, and a note under the boxes says so. Sharp keeps your original pages in the PDF either way.

Missing something in AI OCR — Image & PDF to Text? Suggest a feature →

Tell someone who needs this.