PP-OCRv6_manga

PP-OCRv6_manga

Finds and reads Japanese and Chinese text on manga, webtoon and anime pages, on a CPU. A small ONNX detector and recognizer pair from Kellenok, about 11.6 MB in FP16, for developers building comic reading tools.

Made by Kellenok on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: PP-OCRv6_manga is a small text-reading pipeline for manga and anime pages in Japanese and Chinese. A detector finds the text lines on a page, and a recognizer reads each one. It is light enough for a CPU, at about 11.6 MB in FP16 for both parts together.

Who it suits: Developers and researchers who build reading, search or translation tools for comics and need to pull text out of page images without a GPU. The files carry the Apache 2.0 license.

Industry | Who uses it | How they use it | Example
Software | App developers | Add text extraction to a comic reader | Selecting speech-bubble text on a page
Localization | Translation tool builders | Get source text from pages before translating it | Feeding detected lines into a translator
Research | OCR researchers | Test recognition on manga, webtoons and illustrations | Comparing error rates across page types

Anyone who needs to read printed documents or languages other than Japanese and Chinese can skip it, since it is tuned for comic pages.

Where to use it:

• A CPU, through ONNX and OpenVINO, with no PaddlePaddle needed
• PaddlePaddle, through the card's Paddle example
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Page text detection | Developers | Boxes around the text lines on a manga page
Text recognition | Tool builders | The characters read from each detected line
Benchmarking | Researchers | Error rates for manga, webtoons and illustrations

How to set it up:

• Python and the huggingface_hub package
• The openvino, opencv-python, numpy and pyclipper packages
• A page image, such as page.jpg, to run the example on

1. Install Python if you do not have it yet.
2. Download the model files into a local folder with the first command below.
3. Enter the folder and install the packages with the next two commands.
4. Run the example script on your page image with the last command.

hf download Kellenok/PP-OCRv6_manga --local-dir PP-OCRv6_manga
cd PP-OCRv6_manga
pip install openvino opencv-python numpy pyclipper huggingface_hub
python examples/onnx_openvino.py page.jpg

How to use it: You run the example script on one page image. It passes the page through detection, post-processing and recognition. The card recommends post-processing values of 0.15, 0.25 and 1.5, and a detector input with a long side of 960.

Tool | What it does
manga_det | Finds the text lines on a page, at 0.92 MB in FP16
manga_rec | Reads the characters in each line, at 10.71 MB in FP16
manga_lm | An optional language model for hard characters, which the card calls experimental
onnx_openvino.py | Runs the whole pipeline on one page

Example: You run the script on a scanned page you have the right to use. The detector marks the speech bubbles and the recognizer turns each marked line into Japanese or Chinese text.

For newcomers and experts: OCR means reading text out of a picture. A good first step is to run the quick start on one page image and look at what comes out.

The detector is a PP-OCRv6 tiny model and the recognizer is a PP-OCRv6 small model. FP16 halves the size, and the card reports a 0.02 point difference in error rate against FP32. The card advises plain greedy decoding without the language model for general use.

Load a page image
Detect text lines
Post-process the boxes
Recognize each line
Collect the text

Good to know:

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the download fetches files from Hugging Face, and the source names nothing else
• Languages listed: Japanese and Chinese
• The optional language model is experimental and not recommended for production
• The dictionary covers 19,053 symbols, including all of JIS X 0208 Level 1 and Level 2
• Use it only on pages you have the right to read and copy

Title: End-to-end error rate in percent, version 0.3, lower is better
Manga: 5.98%
Webtoons: 2.74%
Illustrations: 14.92%
Total: 7.57%

Title: Total error rate by version in percent, lower is better
Version 0.2: 8.27%
Version 0.3: 7.57%
Version 0.3 with the language model: 7.47%

On Hugging Face, 13 November 2026

• Task: image to text
• Library: onnx
• Downloads in the last 30 days: 437
• Likes: 14
• License: apache-2.0
• Languages: ja, zh
• Based on: PaddlePaddle/PP-OCRv6_tiny_det
• Tags: onnx, paddlepaddle, pp-ocrv6, text-detection, text-recognition, manga, anime, paddleocr, openvino, image-to-text

Links

• Model: https://huggingface.co/Kellenok/PP-OCRv6_manga
• Files: https://huggingface.co/Kellenok/PP-OCRv6_manga/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Kellenok/PP-OCRv6_manga

Then follow the usage in its model card. Nothing in this kit runs by itself.