Extract text from any image

Capture at the size you will recognise

The single biggest determinant of accuracy is how many pixels tall the letters are, and a screenshot is the one input where you fully control that. Zoom the page or application before capturing rather than enlarging the image afterwards.

Enlarging a small screenshot cannot add the detail that distinguishes similar characters, so the recognised text degrades in exactly the places that matter — digits, punctuation and short words.

Crop to the text and nothing else

Before characters are identified, the page is analysed for structure: blocks, columns, reading order. Interface furniture around your text — toolbars, sidebars, tab strips — is analysed too, and can produce output interleaved in an order nobody wanted.

Cropping tightly to the passage removes that ambiguity and usually fixes 'the words are right but jumbled' without touching any other setting.

Interfaces that are hard to read

Modern interface design works against recognition in two specific ways. Low-contrast grey text on a slightly lighter grey thresholds unpredictably, and text placed over photographs or gradients has no consistent background to separate from.

Dark mode is worth mentioning: light text on a dark background is handled by most engines, but a screenshot mixing dark panels and light ones can threshold inconsistently across the image. Capturing in light mode is a surprisingly effective fix.

Expect to proofread the ambiguous characters

Recognition resolves uncertain shapes using a model of the language, which is why errors cluster where that model has no help to offer.

• Digits and letters that share shapes: 0/O, 1/l/I, 5/S, 8/B.
• Serial numbers, licence keys and codes — no language model can validate them.
• Line breaks, which are guessed from spacing rather than read.
• Accented characters when the language is set wrongly.