Image to Text

Three output modes, and when each one wins

For a spreadsheet, scan the table alone rather than the whole page: fewer surrounding words means cleaner tab separation when you paste into Excel or Sheets.

How to extract text from an image in three steps

1. 1Upload the pictureJPG, PNG or WEBP up to 6 MB. Screenshots, scans and phone photos all work; a straight-on photo in even light gives the cleanest result.
2. 2Pick the output modePlain text is fastest. Keep layout costs nothing extra and is the right default for anything with columns. Handwriting is tuned differently, because cursive letterforms break the assumptions printed text relies on.
3. 3Edit, copy or downloadThe result box is editable: fix a misread word, trim what you do not need, then copy it or download the .txt. Word and character counts are shown above the box.

Getting a clean scan of text

• Shoot straight on. Perspective is the single biggest cause of misreads, because it distorts letter shapes unevenly.
• Light evenly and avoid a flash: glare removes the ink and leaves a white patch that cannot be read.
• Fill the frame with the page; a page at 40 percent of the image wastes the model's resolution.
• Keep the page flat. A curved book edge bends the text and letters start merging.
• For a long document, scan page by page; a page with more text than the model can return in one pass gets trimmed.
• If the photo is poor, enhance contrast first, then convert: text extraction improves more than you would expect.

Handwriting: what works and what does not

Print handwriting reads well: block capitals, school-style cursive, forms filled in with a ballpoint. What struggles is fast personal cursive, fountain pen with heavy ink spread, pencil that has faded, and anything written at an angle on a curved surface.

The mode marks unreadable words as [unclear] rather than guessing, which is deliberate: for a medical note, a legal document or a family letter, an invented word is worse than a gap. Names are the hardest category for every system, so check proper nouns against the original before you rely on the transcript.

What people actually use it for

Expenses and admin are the highest volume: receipts, invoices and delivery notes that need to become text in a spreadsheet. Students photograph lecture slides and textbook pages; the extracted text becomes searchable notes. Support teams pull error messages out of screenshots so they can paste them into a ticket. Researchers transcribe archival documents and old letters.

Developers use it on diagrams and whiteboard photos after a planning session, where the value is not perfect formatting but getting the ideas into a document before they evaporate. Translators use it as a first pass, extracting the source text and then editing, which is far quicker than typing it out.

Languages, accuracy and limits

The vision model handles all major scripts, including English, Hindi, Bengali, Devanagari, Tamil, Telugu, Arabic, Cyrillic, Chinese and Japanese, and mixed-language pages where the script itself is the clue. Accuracy on clean printed text is high enough to trust with light editing; accuracy on degraded scans depends almost entirely on how much of the ink survived.

Limits are practical: 6 MB per image, one image per scan, and a flat 5 credits per scan regardless of output length. If a page is too dense to return in full, split it into halves and scan each, then join the results.

Privacy: what happens to your image

The image is sent to the vision model for the duration of the request and the extracted text is returned to you. Nothing is written into a public gallery, nothing is added to your history, and no email address is collected. That matters for receipts and documents, which is why the request is stateless by design.

As with any online tool, do not upload documents you are not permitted to share, and redact anything sensitive before scanning a page you only need the layout of.

Accuracy: what affects it and how to improve it

Extraction accuracy is decided before the model runs. Resolution sets the ceiling: text smaller than about twenty pixels of height is guesswork. Perspective introduces distortion that varies across the page, so a photo taken at an angle misreads progressively worse towards one side. Lighting decides how much of the ink survives: glare on a glossy receipt erases exactly the digits you need. Contrast decides the rest, because a greyish scan of grey text has very little for the model to separate.

The fixes are photographic, not technical. Shoot straight on and fill the frame. Use diffuse light and no flash. If the page is glossy, tilt it slightly so the reflection moves off the text. If the scan is low contrast, enhance it before extracting. And if the document is long, split it: one page per scan returns more complete text than a double spread compressed into one file.

Turning extracted text into something useful

Raw extraction is a starting point, not a deliverable. For a receipt, the useful move is to copy the lines into a spreadsheet with columns for date, vendor and amount; keep layout mode and the tabs usually land in the right cells, but verify totals by hand. For a contract or a letter, plain text plus a read-through is faster than fighting the formatting, because the model preserves words and you preserve meaning.

For research, the highest-value habit is to keep the image next to the transcript while you edit. Word-level checks are quick when you can see the source, and mistakes cluster in exactly the places you would expect: proper nouns, numbers, and anything handwritten in the margin. Proofreading against the picture takes a fraction of the time that typing the page would have taken.

What to keep out of an OCR tool

Anything you would not paste into a public web form should be considered carefully before uploading. That includes identity documents, bank statements with full account numbers, medical records and signed contracts with personal data. Redaction before scanning, or cropping to the section you actually need, reduces exposure without losing the function.

For internal documents, the practical pattern is to extract only the pages you need, keep the output in a local file rather than a shared one, and delete the source image once the text is verified. None of that is unique to this tool, but it is the habit that separates a useful workflow from an avoidable risk.

Forms, receipts and tables: the practical patterns

Three document types come up again and again, and each has a pattern that works. A receipt is best scanned straight and in keep layout mode, because vendor, date and total sit in predictable positions and the tabs usually survive the paste into a spreadsheet. A form is best read in handwriting mode with the labels visible, so the field name and the filled value stay together in the output. A table is best scanned on its own, cropped tight to the grid, with the header row included in the image.

In all three, the verification habit is the same: check numbers against the source, not against your memory of the source. Digits are where extraction errors concentrate, and a single transposed figure in an expense claim costs more time to fix later than the two minutes it takes to check now.

Keep going with the rest of the toolkit