Guide
Extract text from an image
A photograph of a page is not a page. The words are in it, but nothing can search them, copy them or add them up — and retyping is how an afternoon disappears. This reads the picture itself, layout and handwriting included.
Open an image PDF, Word, PowerPoint, PNG, JPG, WebP, Text · XLSX, CSV, Parquet, JSONOpen the picture
PNG, JPG, WebP, AVIF or HEIC — whatever the phone or the screenshot key produced. It is shown to you first and resized in your own browser, which is also where a photograph's location data is left behind: what travels is a re-encoded JPEG with no EXIF on it.
Say what you want from it
The whole text, a summary of it, the figures in it, or an answer to a question about it. Tables and columns are read as tables rather than flattened into a line, which is where most character recognition falls apart.
Read the answer
It comes back in the language of the page, whatever language the picture is in. Layout is described where it carries meaning — a total at the bottom of a column is a total, not a number that happened to be last.
The picture is sent, and then it is gone
An image has to leave your device to be read; there is no honest way around that, and the page says so before you press the button. What does not happen is storage: it passes through our server to the model and the answer comes straight back. No image store, no database, no log — and nothing is used to train a model.
Does it only extract text, or analyse the picture too?
Both, and they are different jobs. Extracting gives you the text off a photo, a screenshot or a scan. Analysing answers a question about what the picture shows — what a chart says, what went wrong in a screenshot, the objects in a room and how they sit together. Ask for whichever you need. It is not a face-matching service and will not tell you who someone is.
How exact is it?
On print, close to exact, columns and tables included. Where it pulls ahead of character recognition is the awkward page — photographed at an angle, stamped, two languages at once — because it reads the page rather than matching shapes to letters.
Does it read handwriting?
Often, if it is legible. Print and typed text are reliable; handwriting depends on how clear it is, exactly as it would for a person.
Is there a separate OCR step?
No. The model reads the image directly, which is why a photograph taken at an angle, or a page with columns and stamps, usually still comes out right.
Does it work on a scan or a photographed page?
Those are the ordinary cases rather than the hard ones. A scan is a picture of a page, and a picture of a page is what this reads — there is nothing to switch on for it.
How do I get the text out of a screenshot?
Open it as it is — no cropping needed. Menus and sidebars are told apart from the part you want, and asking for just the error message works.
Can I get a table out of a photo?
The table is read as a table and comes back row by row, ready to paste into a cell. It does not produce a finished .xlsx file.
Can it pull the total off a receipt?
That is one of the commonest uses. Ask for the total, the date and the seller, and it reads them from where they sit on the page rather than guessing.
Can it translate text inside an image?
Ask for the translation directly and you skip a step. The text is read off the picture and comes back in the language you asked for, with the layout intact, without a copy-paste in between.
Does it read WebP, AVIF and HEIC?
WebP and AVIF are read like any other picture: your browser decodes them and what leaves your machine is a JPEG either way, so the format stops mattering the moment you drop the file. HEIC, the one an iPhone saves by default, only decodes on Apple devices — anywhere else, share the photo rather than the file and you get a JPG.
Is the photo too big?
Almost certainly not. Large pictures are scaled down in your browser before anything is sent, because a twelve-megapixel photograph carries no more readable text than a two-thousand-pixel one. It makes the answer arrive sooner, not worse.
What about a document with several pages?
One picture at a time here. For something with more than a page or two a PDF gives a better result, and there is a page for that.
Do I need an account, and is it free?
No account, and yes — three files and twenty questions a week without signing in. A paid plan buys more files and more questions; the reading itself is the same on every plan.