PNG to Text

Almost every PNG with words in it started life as a screen rather than a scene. Operating systems save screen captures as PNG by default, design tools export mockups as PNG, and anyone posting a bug report, a chat thread, or a receipt into a ticket is almost certainly attaching one. That origin matters, because it makes PNG the single best-behaved input a character recogniser can be handed. There is no lens, no lighting, no motion blur and no perspective. The glyphs sitting in the file are the exact glyphs a font renderer drew, at whatever resolution your display runs. Feed one in above and the engine — Tesseract 7, built to WebAssembly and served from this site rather than a third party — reads it inside your browser tab. The first visit pulls down the engine along with an English model of roughly 1.9 MB; from then on the browser cache serves both and the tool keeps working with the network switched off.

Lossless storage is why PNG reads cleanly

PNG compresses without discarding anything. Decode a PNG and you get back the identical pixel values that were encoded, byte for byte, which is a fundamentally different arrangement from the lossy formats that dominate photography. For text, that guarantee is worth a great deal. Screen type is rendered with anti-aliasing: the edge of every stroke is a short ramp of intermediate grey values that fools your eye into seeing a smooth curve. A recogniser leans on those ramps too. It thresholds the picture into ink and background, and a clean, consistent ramp gives it an unambiguous place to draw the line. Lossy formats mangle exactly that region, smearing faint halos around each stem where the ramp used to be. PNG leaves it untouched. The same screenshot saved once as PNG and once at middling lossy quality will often produce visibly different transcriptions, and the PNG is the one that comes out right.

There is a second, quieter advantage. PNG stores colour at full resolution rather than at half, so light-grey secondary text on a white panel, or a coloured link inside a paragraph, keeps its edge definition instead of bleeding. Interface design is full of low-contrast text like that, and it is precisely the text that degrades first everywhere else.

Error dialogs and terminal output

The most useful thing this does all week is turn a screenshotted error into a search query. Someone drops a PNG of a crash dialog into a ticket, or a colleague photographs their terminal because copying out of a remote session was awkward, and the exact string you need to look up is trapped inside an image. Stack traces and console output are close to ideal for recognition: monospaced faces have generous, evenly spaced letterforms, terminals run high contrast by default, and the text is rendered rather than photographed. What deserves care is punctuation. Paths, package names, and exception messages are dense with brackets, slashes, colons, and underscores, and those are the characters where a recogniser is least confident. Read the extracted string once before you paste it into a search box or, worse, into a shell.

Chat logs, receipts, and order confirmations

The other big category is evidence. A conversation captured before someone could delete it. An order confirmation screen that never arrived by email. A booking reference, a transfer receipt, a delivery slip shown only inside an app that offers no export. People screenshot these because it is the only way to keep them, then discover the record is unsearchable and impossible to quote. Recognition converts the capture into something you can paste into a note, a claim form, or a spreadsheet. It is also the category where running everything locally stops being an abstract preference: a payment receipt or a private thread is not a file you want sitting in an unknown company's processing queue, and here there is no queue, because the analysis happens on the machine in front of you.

When a PNG still reads badly

The format removes most of the ways an image can be ruined, but not all of them. The commonest failure is scale. If the capture was taken on a low-density display, or the interface was zoomed out, body text may be only eight or nine pixels tall, and below roughly that size there is genuinely not enough shape left to identify a character. Enlarging the PNG afterwards does not help, because upscaling adds pixels without adding information. The fix is upstream: zoom the page or increase the application font size, then capture again. The second failure is a PNG that is not really a PNG any more. Screenshots forwarded through chat apps or pasted into some web tools get converted to a lossy format en route and then saved back out with a .png extension, arriving with all the artefacts the format was supposed to prevent. Very heavy visual styling causes trouble too: condensed display faces, letter-spaced headings, text laid over a photographic hero image, or a gradient running underneath a caption. And a full-page capture that is fifteen thousand pixels tall will work, but it is a lot of picture to grind through on your own processor, so expect it to take a while.

Layout is flattened, and the model speaks English

What comes back is plain text plus a confidence figure, and plain text has consequences for interface captures specifically. A screenshot of a settings panel arrives as labels and values run together in reading order. A captured table loses its columns. A two-pane layout interleaves. Nothing is colour-coded, nothing is bold, and nothing preserves the spatial relationship that made the original readable at a glance, so a dense dashboard capture will need reformatting by hand once the words are out. Two other limits are worth stating rather than burying. The engine reads printed and rendered characters, so it is excellent on interface type and poor on handwriting; and version one ships the English model only, with more languages coming. A German error message or a Japanese app screenshot will still return something, which is exactly the problem — you would have no signal that it was guessing beyond a low confidence figure. And PDFs are not an input the tool accepts today; it works on pictures.

Before you share the capture onward

Screenshots leak. A capture taken to show one dialog routinely includes a browser tab bar with a private URL, a sidebar of contact names, a notification banner, or an email address in the corner. Once the text you actually needed has been extracted, take thirty seconds over the picture itself: pixelate the regions that were never meant to travel, or blur them if a softer look suits the document better. If the file is going into a ticket or an email with an attachment limit, compressing the PNG costs nothing now that its legibility no longer has to survive.