PNG to Text

1. Drop one PNG, or up to ten at once, into the box above. Each file can be up to 25 MB.
2. Leave "Read as" on text for prose, chat threads and error messages. Switch it to tables when the PNG is a grid of figures, so rows and columns stay rows and columns.
3. Press "Read the text". Each image comes back as its own page, in the order you added them, with a confidence mark.
4. Copy what you need, or download it. Plain text and Markdown are free; a Word document or a spreadsheet costs 3 credits.

There is no account, no watermark and no email step for the free pages. Three PNGs a day are read free; past that, each image is 1 credit ($0.05).

PNG is lossless, so every pixel is exactly as it was rendered. Text in a PNG screenshot has perfectly hard edges and full contrast, which is why screenshots read more accurately than photographs of the same content.

The failure modes that remain have nothing to do with the format. Dark mode is one: light text on a dark background reads fine, but if the screenshot has a translucent panel over a busy background, the text sitting on top of it can be genuinely ambiguous. Retina exports are another kind of easy win — a 2x screenshot has twice the text height and reads slightly better than a 1x one.

One PNG-specific thing to watch: transparency. A PNG of black text on a transparent background is invisible against a black canvas. It is read correctly here because the image is composited onto white first, but if you have ever pasted such a file into a dark document and seen nothing, that is why.

Diagrams and charts read as the labels they contain, in reading order. The shapes and arrows are not described — for that you want a person, not a reader.

Almost every PNG reads cleanly. The ones that do not share a handful of causes, and each has a quick fix.

• Text that was tiny when it was saved. A 1x screenshot of a dense dashboard, or a PNG exported at thumbnail size, can leave small labels only six or seven pixels tall. Zoom in before you take the screenshot, or export at 2x, and the same labels read perfectly.
• Pale grey on white. Placeholder text, disabled buttons and light watermarks have so little contrast that they are sometimes skipped. If you need them, screenshot that region on its own.
• Text on a photo or a gradient. Slide titles over a picture and captions on memes read well when the text has an outline or shadow, and less well when it is the same brightness as what is behind it.
• Rotated labels. Chart axes printed at 90 degrees come back, but sometimes in a different place in the reading order. Check them against the image.
• A scan saved as PNG. A whole page scanned at high resolution can be a 20 MB PNG. It is scaled down before it is read, so it works, but a JPG of the same scan uploads faster with no loss in accuracy.

A run of screenshots from a long chat, a set of slide exports or the pages of a manual saved as images can go in as one job. The result keeps them numbered and separate, so after you paste it into a document you can still tell where each image ended.

If your PNGs are really pages of one document, combining them into a PDF first works too. The PDF is read page by page, and each page counts the same as one image.

Upload a picture of a page and the text appears below it, in the order a person would read it. That covers a photo of a book, a screenshot of an app that will not let you select text, a scanned contract, a whiteboard after a meeting, and a page of your own handwriting.

Each page comes back separately, with a confidence mark, so on a batch you can see which page needs a second look.

Nothing you upload becomes a public page. The file is deleted as soon as the job finishes, and the text stays attached to your job for you to copy or download.

Reliable: printed pages, receipts, forms, screenshots, slides, book pages photographed flat, and neat handwriting in the writer's usual hand.

Usually fine: cursive that a stranger could read, a whiteboard photographed square-on, a scan with a coffee ring on it, and mixed pages where a table sits in the middle of prose.

Genuinely hard, and it will say so rather than invent an answer: personal shorthand, faded thermal receipts, text at a steep angle, dense small-type columns, and anything photographed while moving.

Unreadable words come back as [illegible]. That marker is the point: a reader who cannot tell which words were guessed has to check every one.

Accuracy is decided before you upload anything, by how the picture was taken. Four things matter, in this order.

1. Fill the frame with the page. Text that is 40 pixels tall reads perfectly; text that is 8 pixels tall does not, whatever the megapixel count of the camera.
2. Shoot square-on. A page photographed from an angle has converging lines and uneven focus, and the far end of every line is the part that fails.
3. Find even light. One window beats one lamp; a shadow across the middle of a page costs you that band of text. Turn the flash off.
4. Hold still, or rest the phone on something. Motion blur cannot be recovered by any reader.

There is no step five. Resolution beyond about 2000 pixels on the long edge adds upload time and nothing else, which is why oversized images are scaled down here before they are read.

Switch "Read as" to tables and every grid on the page comes back as rows and columns instead of a run of words. That is the difference between retyping a supplier list and downloading it, and it is why the spreadsheet export exists: one sheet per table, with numbers stored as numbers so a column of prices adds up without being cleaned first. When the table is the whole point, Image to Excel is built for exactly that and adds a CSV download.

Screenshots are the easiest input there is, because the text was rendered rather than photographed. Error messages, chat threads, dashboards and anything inside an app that blocks copying all read cleanly.

PDFs are handled two ways. A PDF exported from Word, Pages or a website already contains its text, so it is pulled straight out of the file — no model involved, the result is exact rather than approximate, and those pages are free because nothing had to be recognised. A PDF that is really a scan is read page by page like any other image, and charged like one. You do not have to know which kind you have; the file is checked and the cheaper path is taken automatically.

Three pages a day, across up to three uploads, cost nothing. You see the text, the tables and the confidence marks, and you can copy all of it. There is no account and no email.

Upload more pages than your allowance covers and the pages inside it are read in full, the rest are listed as unread, and the result says exactly where it stopped. Nothing is charged. On a scanned PDF, only the pages you are entitled to are ever sent to be read, so a preview costs the same whether the file has four pages or four hundred.

Beyond that it is 1 credit for each page that has to be recognised. Credits are bought once, never expire, and work across every tool here: a 30-page scanned report costs 30 credits, roughly $1.30 from the middle pack. The Word and spreadsheet downloads cost 3 credits each, once per job — re-downloading the same file later is free.

A PDF that already carries its own text is the exception. Extracting it costs nothing on our side, so it costs nothing on yours: up to 40 pages a job, free, without touching your daily allowance.

Read the numbers first. Text errors are obvious on sight; a transposed digit in a total or a date is not, and those are exactly the values people copy without looking. Then check proper nouns, which have no context to be corrected from, and anything you intend to quote.

That is a minute of work per page — the same minute a data-entry service builds into its price. What you are buying here is the retyping, not the checking.