Koharu Layout HighRes

Koharu Layout HighRes

Detects text, sound effects, speech bubbles and panels on manga pages, returning boxes and masks. For people building manga translation and cleanup tools with the Koharu app.

Made by mayocream 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: Koharu Layout HighRes finds the parts of a manga page and outlines them. The model belongs to the RF-DETR Seg 2XL family and was made for the Koharu app. For four kinds of page element it outputs a box and a mask. It does no text reading and no reading order. It replaces an earlier 1152 layout model.

Who it suits: It suits people who build manga translation or cleanup tools and need to know where text, sound effects, bubbles and panels sit on a page.

Industry | Who uses it | How they use it | Example
Publishing | Manga translators | Locate text and bubbles before typesetting | Finding every speech balloon on a page
Software | Tool developers | Feed layout detection into a translation pipeline | Passing text masks to a reader or cleaner
Research | Vision researchers | Test comics layout detection | Scoring results on a manga test set

You can skip it if you work with Western comics or colour web comics, because the training data is Japanese manga only.

Where to use it: It is meant for the Koharu app, and it also loads through the rfdetr library in Python on Windows, macOS or Linux.

• The Koharu app, which it was made for
• Python scripts that use rfdetr
• Japanese manga pages and spreads

Use case | Who | What they get back
Text detection | Translators | Boxes and masks for dialogue and captions
Sound effect detection | Cleaners | Boxes and masks for onomatopoeia
Bubble and panel finding | Tool builders | Boxes and masks for balloons and panels

How to set it up: You need Python and the download tool. The repository holds the weights, a configuration file for Koharu, the license and a notice file. The weights are about 186 MB in F32.

• Python
• About 186 MB of disk for the weights
• Koharu or an rfdetr-based script to run it

1. Install the Hugging Face download tool.
2. Download the repository files.
3. Point Koharu or your rfdetr script at the downloaded folder.

pip install -U huggingface_hub
hf download mayocream/koharu-layout-highres

How to use it: Give it a manga page image and it returns a box and a mask for each element it finds, labelled with one of four classes. Treat the output as layout only, since it does no reading of the text.

Class | What it marks
text | Dialogue, captions, titles and other non-sound-effect text
onomatopoeia | Sound effects
bubble | Speech and text balloons
panel | Comic panels

Example: You feed it a scanned manga page. It returns boxes and masks for each balloon, each block of dialogue, each sound effect and each panel, which a translation tool can then use to cut out and replace the text.

For newcomers and experts: If you are new, think of it as a tool that draws outlines around the pieces of a manga page. Download the files and let Koharu use them.

If you know the field, it is fine-tuned from the earlier 1152 model in two stages at rising input resolution on manga109-segmentation v3.0.0, using four NVIDIA H200 GPUs. On test pages, box AP rose from 0.653 to 0.824 and mask AP from 0.631 to 0.701.

Manga page image
RF-DETR Seg model
Boxes and masks for four classes
Koharu uses the layout

Good to know: Sound effects inside bubbles can be mistaken for text, and the same lettering can be detected twice.

• License: Apache 2.0, and the NOTICE file carries RF-DETR's notice
• Cost to run: no API key, account or paid service is named
• Data it reaches: the page images you give it, plus files downloaded from Hugging Face
• Not for: reading text, reading order, colour web comics, or non-Japanese comics

Title: Box AP on test pages
1152 model: 0.653
This model: 0.824

On Hugging Face, 13 November 2026

• Task: image segmentation
• Library: rfdetr
• Parameters: 46.6M
• Weights: about 186 MB in F32; about 23 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 125
• Likes: 5
• License: apache-2.0
• Languages: ja
• Tags: rfdetr, safetensors, manga, comics, rf-detr, instance-segmentation, object-detection, layout-analysis, text-detection, image-segmentation

Links

• Model: https://huggingface.co/mayocream/koharu-layout-highres
• Files: https://huggingface.co/mayocream/koharu-layout-highres/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mayocream/koharu-layout-highres

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