manga-ocr-onnx-q8

manga-ocr, quantized for the browser

This reads Japanese text out of a manga panel. It is kha-white's manga-ocr, converted to ONNX and shrunk so it can be downloaded and run inside a web browser.

It is 117 MB instead of 204 MB. Nothing else about it is new.

Credit

Everything that makes this work was made by other people.

• kha-white/manga-ocr-base — the actual model, trained on the Manga109 dataset. All the ability here is theirs. Apache 2.0.
• DigitalLarynx/manga-ocr-onnx — the tokenizer and image settings are copied from this repository. The export tool did not produce them, and without them the model cannot turn its output back into Japanese text.
• optimum — did the conversion.
• transformers.js — what runs it in a browser.

What was actually done

Two steps, both mechanical:

1. Converted the original PyTorch model to ONNX, the format that runs in a browser.
2. Quantized both halves. A model stores millions of numbers. Normally each takes 32 bits; quantizing stores them in 8 bits instead. The file gets about four times smaller and the model gets slightly less precise.

Existing ONNX versions of manga-ocr only quantize the first half, so they are around 204 MB. This one quantizes both, which is where the saving comes from.

Honest limitations

The accuracy has not been measured. Quantizing the second half is the new part here, and nobody has compared this against the full-size version on real pages yet. If you need certainty, use the full-precision files — they are in this repository too.

There is no speed improvement. A faster version would need something called a key-value cache. The conversion tool refuses to produce one for this model's architecture, which is why no manga-ocr export anywhere has it. The file named decoder_model_merged.onnx is the plain decoder under the name the browser library expects — the same thing every other manga-ocr export does, and the only way to make one load.

Japanese only, and it reads a cropped bubble rather than a whole page. Find the text first, then hand it over.

Files

The full-precision files are kept so you can choose, and so this can be compared against them.

Using it

import { pipeline } from '@huggingface/transformers';

const read = await pipeline('image-to-text', 'kimchireader/manga-ocr-onnx-q8', {
  dtype: { model: 'q8', decoder_model_merged: 'q8' },
});

Made for a manga reader that runs entirely in the browser — your pages are never uploaded anywhere.