[ML] Manga Segment

📃 | Description

Manga Segment is a Python computer-vision project for comic and manga page analysis. It trains and runs segmentation models to identify areas of interest such as panels, speech balloons, thought balloons, caption boxes, and text regions, then exports masks, overlays, segmented images, reports, and serving utilities for downstream processing.

The objective is automatic comic translation: segment the areas of interest where text is located, process and generalize those regions to remove the old text, and add the translation on top of the redrawn image.

This project was structured and tested with U-Net and Yolo(v8, v11, v26).

Input	U-Net	YoloV8n (v0.1)	YoloV8s (v0.2)	YoloV11s (v0.2)	YoloV26s (v0.3)

Tune

Model	Tuning Time	Image Size	Epochs/Inter	Iterations	Fitness	Scatter Plots
YoloV8n (v0.1)	45.1h	1280x1280	100	100

Dataset versions

Images and annotations

Property	v0.1	v0.2	v0.3
Images	283	480	754
Train	249	420	660
Valid	22	40	94
Test	12	20	0
Annotations	3293	4832	60419
Comic annotations	1258	1938	3522
Speech-balloon annotations	2035	2894	4194
Caption-box annotations	❌	❌	283
Thought-balloon annotations	❌	❌	749
Text annotations	❌	❌	51671

Preprocessing and augmentations

Property	v0.1	v0.2	v0.3
Auto-Orient	✅	✅	✅
Grayscale	✅	✅	✅
Resize	Stretch to 963x1400	⬜ Fit (white edges) in 963x1400	Fit (white edges) in 992x1440
Auto-Adjust Contrast	Using Contrast Stretching	Using Contrast Stretching	Using Contrast Stretching
Crop	❌	❌	0% Minimum Zoom, 20% Maximum Zoom
Shear	❌	❌	±10° Horizontal, ±10° Vertical
Rotation	❌	Between -15° and +15°	Between -15° and +15°
Exposure	❌	Between -10% and +10%	❌
Flip	Horizontal, Vertical	Horizontal, Vertical	❌
Blur	0.5px	⬆️ 1px	❌
Noise	0.5%	0.5%	❌

Comparison (Unet vs YoloV8 vs YoloV11)

Property	Unet	YoloV8 (v0.1)	Yolov8s (v0.2)	YoloV11 (v0.2)
Precision	0.7444	0.98424	0.96524	0.968
Recall	❌	0.95234	0.97068	0.965
Val Seg Loss	❌	0.7037	0.65568	0.69089
Val Clas Loss	0.23221	0.26816	0.31534	0.27078
Pretrained Model	❌	Yolo Nano	Yolo Small	Yolo Small
EarlyStopping	20	100	25	25
Stop Epoch	26	411	169	251
Image Set	3.882	283	480	480
Image Channels	4	3	3	3
Training Size	512x768	1280²	1400²	1400²
Dropout	0.2	❌	❌	❌
Kernel Size	3	3	3	3
Filter	[32,64,128,256,512]	[64,128,256,512,768]	[64,128,256,512,768]	[64,128,256,512,1024]
Artifacts	high	low	low	low

Details about Yolov8 Details about Yolov11

⚠️ | Disclaimer

This project does not hold the copyright to any work used in connection with the training of its models. Any manga, comic, or other material attributed to the training process remains the property of its respective copyright holders, and no ownership over such works is claimed here. The author does not have the financial means to afford a lawsuit, and this material is shared in good faith.

This project was developed as part of a practical academic study and was built over the course of two years. It is intended for research and educational purposes only. If you are a rights holder and have concerns about any content, please open an issue or contact the author at Matheusn.biolowons@gmail.com, and it will be addressed promptly.

📝 | Cite This Project

If you use this dataset in a research paper, please cite it using the following BibTeX:

@misc{ manga-segment_v2_dataset,
  title = { manga-segment_v2 Dataset },
  type = { Open Source Dataset },
  author = { Ashu },
  howpublished = { \url{ https://universe.roboflow.com/ashu-biqfs/manga-segment_v2 } },
  url = { https://universe.roboflow.com/ashu-biqfs/manga-segment_v2 },
  journal = { Roboflow Universe },
  publisher = { Roboflow },
  year = { 2026 },
  month = { jun },
}

⚙️ | Requirements

Program	Version
Python	>= 3.12
uv	latest

📦 | Setup

Training, inference and the production proxy now live in a single uv-managed folder (src/) built on a scalable, plug-in framework. See src/README.md for full details.

# Windows WSL2: https://www.tensorflow.org/install/pip?hl=en#windows-wsl2_1
# Install CUDA: https://developer.nvidia.com/cuda-downloads
sudo apt install nvidia-cuda-toolkit
sudo apt install -y libjpeg-dev zlib1g-dev

cd src
uv sync

CLI documentation

docs/CLI.md

💹 Production proxy

GET /image?url=<image-url>&mode=<segmented|annotated> — serves segmented or annotated image responses for integration with external tools.

YoloV8:

@software{yolov8_ultralytics,
  author = {Glenn Jocher and Ayush Chaurasia and Jing Qiu},
  title = {Ultralytics YOLOv8},
  version = {8.0.0},
  year = {2023},
  url = {https://github.com/ultralytics/ultralytics},
  orcid = {0000-0001-5950-6979, 0000-0002-7603-6750, 0000-0003-3783-7069},
  license = {AGPL-3.0}
}

Yolo11:

@software{yolo11_ultralytics,
  author = {Glenn Jocher and Jing Qiu},
  title = {Ultralytics YOLO11},
  version = {11.0.0},
  year = {2024},
  url = {https://github.com/ultralytics/ultralytics},
  orcid = {0000-0001-5950-6979, 0000-0002-7603-6750, 0000-0003-3783-7069},
  license = {AGPL-3.0}
}

Yolo26

@misc{jocher2026ultralyticsyolo26unifiedrealtime,
  title = {Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models},
  author = {Glenn Jocher and Jing Qiu and Mengyu Liu and Shuai Lyu and Fatih Cagatay Akyon and Muhammet Esat Kalfaoglu},
  year = {2026},
  eprint = {2606.03748},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  doi = {10.48550/arXiv.2606.03748},
  url = {https://arxiv.org/abs/2606.03748},
}

U-Net:

Ronneberger, Olaf, Philipp Fischer, and Thomas Brox.
"U-net: Convolutional networks for biomedical image segmentation."
In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 234-241. Springer, Cham, 2015.