manga-ocr-torchless 0.2.0

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, with support for Hardware Acceleration

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

Full CLI Options

Argument	Description
image_path	(Optional) Path to an image file to OCR.
-m, --model	HuggingFace repo ID or local path to ONNX model.
-b, --background	Run in background, reading from clipboard (requires pyperclip).
-d, --directory	Watch directory for new images (requires watchdog).
--force-cpu	Force CPU usage even if GPU accelerators are available.
-v, --verbose	Increase output verbosity (shows DEBUG logs).
-q, --quiet	Suppress all output except OCR results.
--delay	Delay (in seconds) before reading new files in directory mode.

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

GPU Acceleration

By default, this package uses your CPU for inference. However, it automatically detects and uses hardware acceleration if you have the appropriate ONNX Runtime execution provider installed:

For NVIDIA GPUs (CUDA):

pip install onnxruntime-gpu

For Windows (DirectML - works on AMD, Intel, NVIDIA):

pip install onnxruntime-directml

For macOS (Apple Silicon / CoreML):

pip install onnxruntime-silicon # or standard onnxruntime which includes CoreML

For Intel CPUs/GPUs (OpenVINO):

pip install onnxruntime-openvino

Note: You may need to uninstall the standard onnxruntime first to avoid conflicts.

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