About Image To Text

Give it one image or 20, pick the language, and it returns the words. Results land in editable boxes with a confidence score attached, so a mis-read character is a quick fix rather than a reason to start over. Take a single result with Copy text or Download .txt, or collect a whole batch into one file with Download all.

JPG/JPEG, PNG, WebP, GIF, BMP and AVIF all work, up to 25 MB each — the browser does the decoding, so if it renders in a tab it can be read here. Recognition covers 30 languages across Latin, Cyrillic, Greek, Arabic, Hebrew, Devanagari, Thai and CJK scripts.

A toggle controls what happens to line breaks. Keeping them as-is suits receipts, tables and code; merging them into paragraphs suits prose you are about to paste somewhere else, and it stitches words that the original layout split across two lines.

The recogniser is Tesseract 5 — the OCR engine that started at HP in the 1980s, was open-sourced in 2005 and now runs on an LSTM neural network — compiled to WebAssembly and executed in a web worker so the page stays responsive. Four steps stand between your file and the text:

1. Load. On the first scan the browser fetches the WebAssembly engine (about 3 MB) and the trained model for your language (0.5–4 MB). Both are served from this domain, not a third-party CDN, and the model is kept in IndexedDB afterwards so the next scan skips the download entirely.
2. Normalise. Your image is decoded with createImageBitmap and redrawn onto a canvas with its longest edge pushed into the 1200–2600 pixel band the model was trained around. A small screenshot gets enlarged, a 12-megapixel photo gets reduced, and transparent areas are filled with white so a cropped PNG is not read as light text on black.
3. Recognise. Tesseract binarises the canvas, works out the page layout — blocks, lines, word boxes — and runs each line through the LSTM model, which emits characters with per-symbol probabilities that get assembled into words.
4. Tidy. Trailing whitespace goes, runs of blank lines collapse, and if you asked for paragraphs the lines inside each block are joined and hyphenated word-splits are repaired.

Everything above is standard browser machinery: the File API to read your image, canvas to normalise it, WebAssembly in a worker to recognise it, and the Clipboard and Blob APIs to hand the text back.

On your own machine, from the first byte to the last. Nothing about your image — the pixels, the filename, the text that comes out — is transmitted to this site or anywhere else, because there is no upload step in the pipeline described above to transmit it during. The traffic runs the other way: the engine comes to your browser.

That also means the tool keeps working with the network off, once the engine and your language pack are cached. Only those two files are stored between visits; your images and their text live in the tab and are released the moment you close or reload it.

This describes what Image To Text itself does. It says nothing about what a browser extension, a corporate proxy or the device itself might see — only that this tool adds nothing to that list.