Get the text out of a PNG in your browser, with nothing uploaded. PNG is the format OCR handles best, and it is worth knowing why: it is lossless, so letter edges are exactly as they were drawn, with none of the halos and smearing that JPEG compression leaves around high-contrast text. PNGs are also almost always screenshots rather than photographs, so the text was rendered by a computer — flat, evenly lit, consistently sized, no perspective. Nearly every variable that normally hurts recognition is absent. The one that remains is resolution: body text in a standard screenshot gives only about a dozen pixels per character, which is near the engine's comfortable lower bound. Zoom the page before capturing and accuracy goes up noticeably.
| Factor | PNG screenshot | JPG photo |
|---|---|---|
| Compression artefacts | None | Halos around letters |
| Lighting | Perfectly even | Shadows and glare |
| Perspective | None — flat | Usually at an angle |
| Character shapes | Rendered consistently | Varies with focus |
| Main limitation | Resolution | Everything above |
No. Recognition runs in your browser with Tesseract compiled to WebAssembly.
Because it is lossless. JPEG saves space by discarding fine detail, and the sharp edge of a letter is the first thing it discards — producing the halos that make an 'e' read as a 'c'. PNG keeps every pixel exactly as drawn, so the recogniser sees clean strokes.
Increase the resolution. Zoom the page or application before you capture, or capture on a high-density display. Body text in a standard screenshot has only about twelve pixels of character height, which is near the lower limit; doubling that makes a visible difference.
Usually, yes — the engine handles inverted text reasonably well. Very low-contrast themes, such as mid-grey on dark grey, are harder, and increasing contrast before capturing helps.
Reading order is preserved, but columns, tables and code indentation are flattened into lines. For a screenshot of a table, expect the text in order rather than in cells.
Several, chosen before you run it. Select the right one — an English model reading German will mis-handle umlauts, because it is not expecting them.
PNG is the easy case for OCR, and understanding why explains when this will and will not work well. The format is lossless, so every pixel is exactly what was drawn — no compression halos, no smeared letter edges, no artefacts around high-contrast transitions. PNGs of text are also overwhelmingly screenshots rather than photographs, which means the text was rendered by a computer at a consistent size, on a flat plane, with perfect lighting and no perspective. That combination removes nearly every variable that normally degrades recognition, and accuracy on screen-captured interface text is correspondingly high. The remaining limitation is resolution rather than quality: a screenshot taken on a standard-density display gives roughly 12 to 16 pixels of height per character for ordinary body text, which is near the lower bound of what the engine handles comfortably. Capturing on a high-density screen, or zooming the page before capturing, doubles that and measurably improves the result — which is the opposite of the usual advice, since here the problem is too few pixels rather than too much compression. Recognition runs through Tesseract in WebAssembly, entirely on your device, with batch and folder support.