Listen to this post: Getting Text Out of Images: What OCR Does Well and Where It Fails
Text trapped in an image is a specific kind of annoying. You can read it perfectly well; you just cannot select it, search it, or paste it anywhere. A photographed page of notes, a screenshot of an error message, a scanned form, a slide from a lecture you were not allowed to download.
Optical character recognition solves this, and Vibeland’s OCR scanner does it without an account or an upload limit. What follows is when it works well, when it does not, and what to do about the cases in between.
What OCR is doing
OCR takes an image, finds the regions that look like text, and works out which characters those shapes represent. Modern systems do this with trained models rather than by matching against a library of letterforms, which is why they cope with fonts they have never seen and with text that is slightly rotated or unevenly lit.
What comes back is plain text. Not layout, not formatting — text, in roughly the reading order the system inferred.
Where it works well
Screenshots of digital text. The easiest case by a distance. A stack trace in a terminal, an error dialog, text in a slide, a chunk of code in a screenshot someone sent you instead of pasting. The characters are crisp, the contrast is high, the baseline is straight. Expect this to just work.
Clean printed documents. A flatbed scan or a straight-on photo of a printed page in decent light. Standard body text in a normal typeface comes out well.
Signage, labels, packaging. Short bursts of large, high-contrast text — a serial number, an address, a wifi password on the back of a router — are usually reliable.
Where it struggles
Being straight about this saves you more time than any tip.
Handwriting is the hard case, and results vary enormously with how neat the writing is. Tidy, well-separated printing sometimes comes through usefully. Genuine cursive, or the scrawl of someone taking notes at speed, often does not. Treat any handwriting result as a draft you will need to correct rather than a transcription you can trust.
Poor source images. Blur, glare, deep shadow across half the page, a photograph taken at a sharp angle. OCR cannot recover information the image does not contain.
Complex layouts. Multi-column pages, tables, sidebars, captions running alongside body text. The characters may be read correctly while the reading order is scrambled, because the system had to guess how the blocks relate. Tables in particular tend to lose their structure.
Unusual typography. Decorative fonts, very tight tracking, light text on a busy photographic background, condensed all-caps.
Getting better results
Most OCR failures are input failures, and the fixes are simple.
- Shoot straight on. Hold the camera parallel to the page rather than at an angle. This matters more than resolution.
- Get even light. Diffuse daylight beats a single overhead lamp, which throws a shadow from your own hand. Avoid glare on glossy paper.
- Crop to the text. Remove the desk, the edge of the book, your thumb. Less for the system to interpret.
- One column at a time. If the page has two columns, crop and run them separately. This solves reading-order problems outright.
- Prefer the digital original. A screenshot of a PDF beats a photograph of a printout of the same PDF, every time.
Always check the output
The characteristic OCR error is not gibberish — it is a plausible substitution. A “1” for an “l”, a “0” for an “O”, “rn” collapsing into “m”. These survive a quick glance, which is exactly what makes them dangerous in anything containing numbers.
Reference numbers, dates, quantities, prices, codes: read them back against the image before you use them. For prose it matters less; for a figure going into a document, verify.
Where it runs, and what that means
Most of Vibeland is pure JavaScript on your device — nothing you paste is transmitted. OCR is one of the exceptions, and the site says so rather than blurring it. The models are far too heavy to sit in a browser tab, so the work happens on a server.
That is worth knowing before you use it. A screenshot of a public webpage, a photo of a whiteboard, a slide from a lecture — no real concern. A signed contract, a medical letter, anything under NDA — apply the same judgement you would to uploading it anywhere, and if the answer matters, use something on hardware you control.
The same applies to PDF extraction, which is the tool to reach for when your source is a PDF rather than an image.
After the text is out
OCR output usually needs a tidy. The whitespace remover strips the ragged spacing that comes from column detection, the find and replace tool fixes a substitution that repeats throughout, and the word counter tells you whether the extracted passage fits wherever it is going.
All three run on your device, so once the text is out, the rest stays with you.
