manga-ocr-torchless 0.1.0

Torchless optical character recognition for manga focused Japanese text, lightweight version of kha-white's manga-ocr

manga-ocr-torchless

A lightweight, torch-free version of the excellent manga-ocr by kha-white.

This package uses ONNX Runtime for inference, making it significantly faster to install and run on machines without a GPU, eliminating the multi-gigabyte dependency on PyTorch.

By default, this package uses mayocream's onnx export of the original manga-ocr package, but you can use any onnx export based on the manga-ocr package such as l0wgear's onnx export of jzhang533/manga-ocr-base.

Features

• Lightweight: No PyTorch dependency (~2GB saved).
• Parity: Achieves 100% character-level parity on the original test suite.
• Fast: Optimized for CPU inference using ONNX.

Installation

pip install manga-ocr-torchless

Note: The required ONNX models (~400MB) will be automatically downloaded from HuggingFace on the first run, not during installation.

Usage

CLI

Process a single image:

manga_ocr path/to/image.jpg

Monitor clipboard (auto-OCR every time you copy an image):

manga_ocr -b

Watch a directory for new screenshots:

manga_ocr -d ./screenshots

Python API

from manga_ocr import MangaOcr
from PIL import Image

mocr = MangaOcr()
img = Image.open('image.jpg')
text = mocr(img)
print(text)

Custom Models

You can use a different ONNX model by providing a HuggingFace repo ID or a local path to the constructor or via the --model flag in the CLI:

Python:

mocr = MangaOcr(pretrained_model_name_or_path="user/repo-id")
# OR
mocr = MangaOcr(pretrained_model_name_or_path="./local_model_directory")

CLI:

manga_ocr --model "user/repo-id" path/to/image.jpg

Acknowledgments

This project is a direct port of manga-ocr by kha-white. All credit for the model architecture and training belongs to them. This version simply swaps the backend to ONNX for a leaner distribution.

License

Apache-2.0