---
license: cc-by-nc-4.0
language:
- en
base_model:
- stabilityai/stable-diffusion-3.5-medium
---

# Text-Image-to-Image Trained on MANGA109 Pose HA Manga Images

This repository contains an image generation model trained using manga images from [MANGA109 Pose tools](https://github.com/kuri-lab/MANGA109-Pose-tools).
Please create the conditional input images using the repository linked above.

## Training Parameters
|Argument | Value |
| ---- | ---- |
|resolution | 512 |
|train batch size | 4 |
|learning rate | 1e-05 |
|mixed precision | fp16 |
|max train steps | 200,000 |

## Training Dataset
The MANGA109 Pose HA dataset was split into training, validation, and test sets in an 8:1:1 ratio.

## Author's Environment
  - GPU：H100NVL (1 unit)
  - CUDA：12.4
  - PyTorch：2.6.0+cu124
  - diffusers: 0.33.0.dev0

## Computation Time
Training was conducted on a single H100 NVL GPU (94GB) and took 88 hours.
Each training step took approximately 1.58 seconds.

## License
This repository is licensed under the [Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) ](https://creativecommons.org/licenses/by-nc/4.0/deed.en).

## Citation
If you use this repository in your research, please consider citing it using the following BibTeX entry:
```
@article{okada2025manga109pose,
  title={MANGA109 に姿勢情報を追加したデータセットの構築による姿勢を制御した漫画キャラクター画像生成},
  author={岡田 湧路 and 北川 峻 and 渡邉 謙吾 and 稲葉 通将 and 橋本 敦史 and 栗原 聡},
  journal={人工知能学会全国大会論文集},
  volume={JSAI2025},
  pages={2O1GS1005-2O1GS1005}
  year={2025}
}
```

Update History
* 2025/04/25: [Public Release]