"---
license: cc-by-4.0
task_categories:
- object-detection
- image-to-text
language:
- ja
- en
- ko
tags:
- manga
- manhwa
- synthetic
- ocr
- multilingual
pretty_name: Mangium
size_categories:
- 10K<n<100K
![manga_synth_000000008_135b6d0c](https://cdn-uploads.huggingface.co/production/uploads/6855122609fb1b2a7291f170/S3h6uLAYWBqR6sgeRFKMy.jpeg)
![manga_synth_000000044_d270de85](https://cdn-uploads.huggingface.co/production/uploads/6855122609fb1b2a7291f170/OMzI1Bgj09jeskDYsHbun.jpeg)
![manga_synth_000000094_a3e9a185](https://cdn-uploads.huggingface.co/production/uploads/6855122609fb1b2a7291f170/K7yOgwDzTTSPkjvhG9CwE.jpeg)

- I am Indomie, an independent developer who is just starting out in this field. Driven by my passion for manga and my desire to build tools around it, I hit a major roadblock: the severe shortage of free, open-source datasets available for training models—especially datasets specifically tailored for manga itself. Because of this, I decided to build and generate this dataset myself to bridge the gap and help out others who are interested in this space.

Note: Since this is my very first project of this kind, the data might not be perfect or entirely error-free. It remains a work in progress, and I will be continuously improving future versions based on your feedback and usage.

-About Me & Project Context
I am Indomie, an independent developer who is just starting out in this field. Driven by my passion for manga and my desire to build tools around it, I hit a major roadblock: the severe shortage of free, open-source datasets available for training models—especially datasets that support Asian languages with complex, vertical text directions. Because of this, I decided to build and generate this dataset myself to bridge the gap and help out others who are interested in this space.

Note: Since this is my very first project of this kind, the data might not be perfect or entirely error-free. It remains a work in progress, and I will be continuously improving future versions based on your feedback and usage.


-Language DistributionI balanced the languages in this dataset to cover different reading systems (both vertical and horizontal alignment) according to the following breakdown:LanguageDataset ProportionSupported Reading DirectionJapanese (JA)40%Vertical (majority) / HorizontalKorean (KO)30%Horizontal / VerticalEnglish (EN)30%Horizontal


- Generated Graphic Components (Dataset Features)
I programmatically overlaid custom text and dialogue elements onto reference images. These components include:

Text Balloons: Multiple shapes including elliptical bubbles (standard dialogue), cloud-like shapes (internal thoughts), jagged/spiky bubbles (screaming/loud dialogue), and rectangular narrative boxes.

Visual Sound Effects (SFX): Free-form, angled, and curved texts that mimic expressive, hand-drawn sound effects in the background.

- Dataset Output Layout
The outputs are structured in a standardized directory format, making it easy to plug them directly into your training pipelines:

images/: Contains the final composited images in .jpg format.

annotations/: Individual .json files corresponding to each image name, containing:

High-precision polygon coordinates (Polygons) defined at both the unit/paragraph level and individual word level.

The actual text content (Transcription) mapped to each polygon.

Structural classification labels (e.g., dialogue bubble, sound effect, narrative box).

manifest.jsonl: A consolidated manifest file that maps each image path to its annotations on a single line, allowing for rapid batch loading during training.


- Target Applications
You can use my dataset to train and evaluate models on several key tasks:

Improving Text Detection: Training models to locate text lines that overlap with background art or are written in highly stylized, angled, or non-standard fonts.

Developing Manga OCR: Training classifiers to recognize multilingual characters oriented in both vertical and horizontal directions within the same page.

Layout Analysis & Segmentation: Helping models distinguish between regular dialogue bubbles, narrative boxes, and background sound effects to automate cleaning or translation pipelines.


- License
I have released this dataset under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. This means you are completely free to use, modify, and build upon this data for academic or commercial purposes, as long as you give appropriate credit to me.