mednasserallah/manga-panel-detector-yolo26n-onnx

Manga panel detector (YOLO26-nano) — ONNX export

This is not an original model. It is an ONNX export of leoxs22/manga-panel-detector-yolo26n by Leandro Narosky, published here because the original repository ships PyTorch and TFLite weights only — and exporting the ONNX requires ultralytics, which requires PyTorch, which is more to install than most applications that want to run this model.

All credit for the model belongs upstream. If you use it, cite the original (citations below), not this repository.

What changed

Nothing but the format. No retraining, no fine-tuning, no quantisation, no change to the weights.

pip install ultralytics onnx
yolo export model=manga_panel_detector_fp32.pt format=onnx imgsz=1024 opset=17

Source file	manga_panel_detector_fp32.pt (14.8 MB)
This file	manga_panel_detector_fp32_1024.onnx (10.1 MB)
Exported with	ultralytics 8.4.117, onnx 1.22.0, onnxslim 0.1.95
Opset	17
Input	1024×1024, letterboxed
Classes	0: panel, 1: text

Note the input size. The original was trained at 640×640; this is exported at 1024, which suits full manga pages where the thin panels down the side of a busy page are lost at 640. Export it yourself at another size with the command above if you want a different trade-off.

Usage

import onnxruntime

session = onnxruntime.InferenceSession(
    "manga_panel_detector_fp32_1024.onnx", providers=["CPUExecutionProvider"]
)

The page is letterboxed into a 1024×1024 square, run through the session, and the boxes are scaled back to the page's own pixels. Class 0 is a panel frame, class 1 is text.

Licence

Apache-2.0, the same as the original model — see the LICENSE file in this repository. Attribution and the notice of modification above are required by §4 of that licence.

The training data (Manga109-s) has its own terms. This model was trained on Manga109-s. Per condition 5 of the Manga109-s licence, results obtained from machine learning experiments — pre-trained models included — may be used for commercial purposes provided the use of the dataset is clearly indicated, which is what this section does.

Citations

The model:

@misc{leoxs22_manga_panel_detector_2026,
    author={Leandro Narosky},
    title={{Manga Panel and Text Detector (YOLO26-nano)}},
    year={2026},
    publisher={Hugging Face},
    url={https://huggingface.co/leoxs22/manga-panel-detector-yolo26n}
}

The dataset:

@article{multimedia_aizawa_2020,
    author={Kiyoharu Aizawa and Azuma Fujimoto and Atsushi Otsubo and Toru Ogawa and Yusuke Matsui and Koki Tsubota and Hikaru Ikuta},
    title={Building a Manga Dataset ``Manga109'' with Annotations for Multimedia Applications},
    journal={IEEE MultiMedia},
    volume={27},
    number={2},
    pages={8--18},
    doi={10.1109/mmul.2020.2987895},
    year={2020}
}

@article{mtap_matsui_2017,
    author={Yusuke Matsui and Kota Ito and Yuji Aramaki and Azuma Fujimoto and Toru Ogawa and Toshihiko Yamasaki and Kiyoharu Aizawa},
    title={Sketch-based Manga Retrieval using Manga109 Dataset},
    journal={Multimedia Tools and Applications},
    volume={76},
    number={20},
    pages={21811--21838},
    doi={10.1007/s11042-016-4020-z},
    year={2017}
}

Integrity

sha256  e66667bc6d5f00013ff27efc15d21e521825369d44dfd5d7f6e43cda2ca512b7
bytes   10068534