Convert GIF to TXT

Almost every GIF with readable words in it is a screen capture. Older screenshot tools defaulted to GIF, forum software converted uploads to it, wikis and bug trackers accepted it when they accepted little else, and a great many support tickets from the last twenty years have a GIF of an error dialog attached to them.

That history is why this conversion is worth having and also why it is harder than reading a scan. A screenshot saved as a GIF has been through a palette reduction that a screenshot saved as a PNG has not, and text is the content most sensitive to it.

The recogniser reads the file directly, and its own documented format list covers GIF for non-animated files. A screenshot, a captured dialog, a scanned page saved as a GIF: all of those are inside what it reads, and nothing on this site has to decode them first.

An animated GIF sits outside that. Rather than guess at which frame would be used or whether the file would be read at all, the honest route is one extra step: convert the GIF to a PNG, which takes the first frame, and run the PNG through this tool. If the text you want is not on the first frame, extract that frame in an image editor first.

Screen text is antialiased: the edges of every glyph are drawn in intermediate shades so the strokes look smooth. A GIF has at most 256 colours to work with, and when a screenshot is reduced to that palette the intermediate shades are the first casualty — replaced by the nearest available colour, or dithered into a checkerboard of two.

Recognition works by comparing the shape of each glyph against what it has learned, and a stroke edged with alternating pixels is a different shape from a stroke edged with a smooth ramp. This is the main reason a GIF screenshot reads slightly worse than the same screenshot as PNG, and it is also why enlarging the GIF before running it through does not help: enlarging a dithered edge produces a larger dithered edge.

Every other image conversion on this site decodes the picture to raw pixels first and works from those. This one deliberately does not — the file is handed over untouched, because the recogniser has its own loader and forcing the image through a canvas would flatten anything transparent to black.

On a scan of white paper that single detail is the difference between a readable page and a black rectangle, and it is worth knowing if your GIF has a transparent background. Because the file is passed straight through, what the recogniser sees is what the file contains, transparency and all.

Four language models are available: English, German, French and Spanish. The setting is a choice rather than something detected, because detection is itself a guess and a wrong one would be invisible.

A wrong choice does not fail. It produces text — plausible-looking words, in the wrong language, mapped from shapes onto whatever vocabulary the model holds. That is the most dangerous failure mode this tool has, because everything about the output looks normal. If the result reads like a badly translated version of what you can see in the picture, the language setting is the first thing to check.

What comes back is lines of text with the blank lines the recogniser saw between blocks, and no other processing at all. Nothing collapses those blanks, joins lines that look like one wrapped sentence, or rebuilds columns.

Every one of those tidying steps is a guess about layout. A poem, an address block, a code sample and a table all lose meaning when their line breaks are rearranged, and the tool has no way to tell which of those it is looking at. Leaving the breaks alone means a screenshot of a table comes back with its columns run together, which is inconvenient and recoverable — unlike a rewrapped address.

Starting the recogniser means downloading and compiling several megabytes of WebAssembly and then the language model. That is seconds of waiting before any recognition begins, and it is why this is the slowest tool on the site by a wide margin.

It happens once. The model is cached on your own machine after the first run — local storage, nothing shared — so a second GIF starts almost immediately, and a batch of twenty pays the startup cost once rather than twenty times. Switching language reloads a different model, so it is worth doing all the English files together.

Handwriting, essentially. Text at an angle, text over a busy background, text rendered so small that the strokes are one pixel wide, and text in a decorative typeface will all produce errors ranging from a few characters to complete nonsense.

The other case worth naming is a GIF that was scaled down after the screenshot was taken. Halving the size of a screenshot destroys the glyph shapes in a way no recogniser recovers, and it is extremely common because people resized images to fit forum limits. If you have any route back to the original capture, use it — this conversion cannot restore what the resize removed.

The recogniser, its WebAssembly core and every language model are served from this site rather than fetched from a public CDN. That took deliberate work, because the library’s default is to fetch all three from elsewhere, and the point is not the image — the image would never have left either way — but the fact of the request.

Somebody reading a payslip, a medical letter or a private conversation out of a screenshot should not have that announced to a third party by the URL of the page they are on. The recognition happens in your browser, there is no account, and the text is never stored.