Image to Text (OCR)

Recognize and review printed text

1. Choose one local PNG, JPG or JPEG up to 5 MiB and 4 million pixels, with neither side above 8192 pixels. Non-animated, non-interlaced PNG and header-readable JPEG are supported. Use English example selects an English image and recognition model without starting OCR.
2. Select the language printed in the image, then choose Recognize text. English, Russian, Simplified Chinese, Portuguese, Italian, Spanish, French and German are available in every page language. This selection is not automatic language detection or translation.
3. Compare the recognized text with the prepared image and correct mistakes in the editor. Copy text and Download text use the full current valid draft. Restore recognition removes local edits without running OCR again; changing the image or recognition language clears the result.

Understand what OCR can and cannot read

OCR estimates characters from image pixels. Clear horizontal printed text is the intended input. Blurring, small letters, skew, complex layouts, unusual fonts and a mismatched language can cause missing, substituted or wrongly ordered text. No result is guaranteed accurate, and the tool does not display a confidence score as an accuracy percentage.

The image is decoded at its oriented resolution and placed on an opaque white background without resizing, cropping or automatic upright detection. The preview uses those same prepared pixels. White lettering on transparent backgrounds may disappear against white; prepare a contrasting source image first.

Page language controls the interface, not the document's contents. Recognition starts with the page's matching model but remains an independent choice afterward. The Portuguese model is not claimed to be Brazil-specific, and Simplified Chinese recognition does not promise Traditional Chinese or vertical-text accuracy. Only one model is used at a time.

This version produces editable plain text, not Word documents, tables, searchable PDFs, translated text or reconstructed page layouts. URLs and markup in the result are inert text. No text detected is a recognition outcome, not proof that an image contains no writing.

Review output and local corrections

The English example is designed as a clear 1200 by 400 pixel image containing Hello 2026. For a returned text value of Hello 2026 followed by LF, the editor contains 11 Unicode scalars and the UTF-8 download has 11 bytes. This explains the counting and export policy, not an accuracy guarantee for every image.

Adding an exclamation mark before that LF changes the current draft to 12 scalars and 12 bytes and marks it Edited locally. Restore recognition returns the stored engine text. Editing does not validate the transcription or start another recognition attempt.

Characters and bytes differ across writing systems. The text café, a space, 你好, a space, an emoji 😀 and LF contains 10 Unicode scalars but 18 UTF-8 bytes. The emoji uses two UTF-16 units, and combined writing marks can span several scalars. These counts are not words or user-perceived characters.

Input, runtime and text bounds

Original files are limited to 5242880 bytes, oriented images to 4000000 pixels and 8192 pixels per side, and the prepared PNG to 16777216 bytes. Strict header budgets may reject an otherwise valid image. Animated PNG, known multi-picture JPEG, WebP, GIF, SVG, PDF, camera, URL and batch input are not supported.

Preparation has a 15000 millisecond watchdog within a 120000 millisecond overall attempt watchdog, including static engine and language loading. Browser scheduling can delay cancellation observation. A worker error, unsupported capability, timeout or cancellation produces no partial exported result and never falls back to an upload or main-thread recognizer.

Received text and editable drafts fit 200000 UTF-16 units, 100000 Unicode scalars and 400000 UTF-8 bytes. Unpaired surrogates, NUL or blank drafts cannot export; text is never silently repaired or truncated. The UTF-8 download keeps current spaces and line endings, with no added BOM or final newline. Clipboard applications may handle line endings differently.

Handwriting, equations, vertical text, mixed-language accuracy and document layout preservation are not promised. OCR and the chosen fast models can make mistakes. Compare important characters and numbers with the image rather than interpreting success, output size or a model name as proof of correctness.

Local recognition and static asset loading

This tool adds no image upload, account, saved history, source-bearing share URL or image/text logging. It downloads fixed engine assets and the selected language from this site's origin when recognition starts; a cold attempt can need about 8 MiB. Engine IndexedDB language caching is disabled, though normal static HTTP caching can occur. Existing site services remain subject to site policies; local processing is not certification of those services, deployment or offline availability.