manga-ocr-browser

manga ocr browser

i prepared smaller downloads of mangaocr for japanese text recognition in the browser. these are the existing q8 exports from onnx community, compressed with gzip. the weights are unchanged.

the encoder and decoder download in about 95.5 mb combined, down from 116.6 mb. decompress them before loading them into ONNX Runtime Web. compression reduces the download size. inference time and memory use stay the same.

in my manga samples, the model corrected several weak paddle predictions but introduced errors elsewhere. readest therefore uses it to retry uncertain lines. small reading guides beside kanji can still cause errors.

scripts and test results

use

download both .onnx.gz files and vocab.txt. decompress the model files with gzip. the source model documents the preprocessing and decoder settings. this repository includes file sizes and checksums in models.json and packed.json.

the model and vocabulary retain the apache license, version 2.0. see LICENSE and NOTICE for attribution. i used existing weights throughout, and the test pages remain private.