Convert PNG to TXT

It is a remarkably common situation. An error message screenshotted into a ticket. A document received as a picture instead of a file. A chart whose labels you need. A page of a PDF that turned out to be a scan. A message forwarded as an image because that was easier than copying it.

In every case the words exist and the only route to them is retyping — until you read them back off the picture. That is what this does, and on a screenshot it does it well.

Everything that makes photographed text hard is absent. The focus is perfect because nothing was focused. There is no angle, because the page is the screen. There is no shadow, no uneven light and no camera shake. And the type was rendered by software rather than printed and then photographed, so the letterforms are exactly what the font specified.

The practical consequence is that expectations can be higher here than on the photograph pages. Clean interface text at normal resolution usually comes back essentially exact, and where it does not, the cause is nearly always one of the two things below.

Recognition needs enough pixels per character. A screenshot taken at full size on a modern display has plenty; one that has been scaled down to fit a document, compressed by a chat application, or captured from a low-resolution screen may not.

Enlarging the image afterwards does not help — it adds pixels without adding information, and often makes the result worse by softening the letterforms. If a result is poor, retake the screenshot at full size, or zoom the application in before capturing so the text is physically larger on screen.

The recogniser matches shapes against the letterforms and vocabulary of one language, and given the wrong one it produces confident nonsense rather than an error. This catches people out on screenshots specifically, because interface text is short and the errors look like plausible words.

It matters most for anything with accented characters or a non-Latin script, and for mixed content — a German interface with English technical terms in it will read better with the language set correctly than with either one guessed.

Screenshots of code and command lines are among the commonest inputs, and they carry a specific hazard. The characters that matter most in code are precisely the ones that look alike: the digit one against a lowercase L against a pipe, zero against capital O, and in some fonts a hyphen against an underscore or an en dash.

A misreading between any of those produces text that looks correct and does not work, and the error is in the character least likely to be noticed. Read anything you intend to run, and prefer a monospaced source — the fonts terminals use are designed to distinguish exactly these characters, which helps the recogniser as much as it helps you.

The output is plain text in reading order. A table becomes a run of values with no columns, a two-column layout can interleave, and headings arrive as ordinary lines.

For most screenshot work that is exactly enough — you wanted the error message, the reference number or the paragraph. Where it is not enough is a table you intend to use rather than read, and there the honest position is that this gets you the numbers to verify rather than a structure to import.

The recognition runs inside the browser tab on your own processor. Nothing is uploaded, so there is no server copy, no account and no daily allowance, and a folder of images converts as fast as your machine can read them.

Screenshots are a particularly sensitive category and people rarely think of them that way: they routinely contain account numbers, internal system names, customer records, colleagues’ names in a message list, and whatever else happened to be on screen at the edges of the capture. Not sending them anywhere is the straightforward answer to all of it.