Page images stay local — model weights download on opt-in (size disclosed above the gate)
TrOCR + detect
Line OCR: RapidOCR detects boxes, TrOCR recognizes each crop (Xenova/trocr-small-printed via Transformers.js).
1. Download models (~64 MB + ~14 MB det)
TrOCR alone is weak on full pages — we pair RapidOCR EN detect then crop→recognize. If detect finds nothing, falls back to whole-page TrOCR (honest but usually worse).
Models not loaded
2. Image or PDF
Drop a page image or PDF here or click to browse
PNG, JPG, WEBP, or PDF — one PDF page at a time
One page at a time.
3. Plain text
Faster plain OCR? Try RapidOCR English.