No cloud OCR request
The site is delivered as static files. After that, each selected image moves only between the page and a dedicated browser worker.
OCR without upload
Your image stays on your device
Queue up to 10 images or 100 MB. A worker reads them sequentially in your browser and returns editable text without sending images or results across the network.
✓ No uploads✓ English + Chinese✓ Works offline after caching
A strict network boundary
A Content Security Policy limits connections to same-origin application, model, and anonymous task-count requests plus Cloudflare page analytics. There are no ad, account, upload, or OCR endpoints.
One decoded image at a time
The queue retains small previews and results instead of keeping every full-size image decoded in memory.
Offline after the first cache
Fast mode can work offline after its models finish caching. High accuracy works offline only after its optional model has also been stored.
Common questions
Frequently asked questions
Where is OCR processed?
Inside your browser with ONNX Runtime Web. There is no image or recognized-text API endpoint.
Can the site see my text?
No. BrowserOCR has no server OCR fallback. Anonymous measurement is limited to page visits and three anonymous task steps—never images or recognized text.
What are the batch limits?
Up to 10 images, 50 MB per image, and 100 MB for the queue. Images are decoded and recognized one at a time.
What is downloaded?
The Fast detector and recognizer download when OCR starts. The 21 MB High accuracy recognizer requires an explicit confirmation.
What actually crosses the network
Checked: · Chrome 152.0.7977.76
The browser downloads the page, JavaScript, dictionary, ONNX models and WASM runtime from this site. OCR runs in a worker. The local test observed GET requests for these assets and POST requests to /api/demand-event for anonymous task steps. The production check on this date found no Cloudflare page-analytics beacon; the CSP allows it if enabled. Zero upload does not mean zero network activity.
The task endpoint accepts only task_started, task_completed and result_exported with a page path and an enumerated coarse source. It does not accept image bytes, filenames or recognized text. Use your browser Network panel to inspect request payloads while trying the synthetic image.
/models/ppocrv6-tiny-det.onnx+/models/ppocrv6-tiny-rec.onnx(Fast)/models/ppocrv6-small-rec.onnx(High accuracy)/ort/ort-wasm-simd-threaded.wasm
When offline OCR works
This sample was processed again after the models and page assets were cached, the browser went offline and the same page reloaded. The recorded offline output appears below. First use needs a connection. High accuracy needs its optional model downloaded first. Other pages, formats and PDF assets must be opened or used online before relying on them offline. Clearing site data, private browsing or storage eviction can remove the cache.
Observed offline result: 烯OCR纬摧PRIVATE