# Akashi Junko (Blue Archive)
赤司ジュンコ (ブルーアーカイブ) / 아카시 준코 (블루 아카이브) / 赤司淳子 (碧蓝档案)

[**Download here.**](https://huggingface.co/khanon/lora-training/blob/main/junko/chara-junko-v1c.safetensors)

## Table of Contents
- [Preview](#preview)
- [Usage](#usage)
- [Training](#training)
- [Revisions](#revisions)

## Preview
![Junko portrait](chara-junko-v1c.png)
![Junko preview 1](example-001-v1c.png)
![Junko preview 2](example-002-v1c.png)
![Junko preview 3](example-003-v1c.png)

## Usage
Use any or all of the following tags to summon Junko: `junko, slit pupils, demon horns, halo, twintails, hair ribbon, pointy ears, demon wings`
- You can also use `low wings` if the wings appear too high.

For her normal outfit: `military uniform, short sleeves, black shirt, plaid skirt, red necktie, thigh strap, black boots`

For her New Year alt: `japanese clothes, yellow kimono, black hakama skirt, black boots, kinchaku`

For her normal expression: `closed mouth, smile, :3`
- You may need to prefix the colon with a backslash character.

For her hangry expression: `open mouth, wavy mouth, skin fang, (tearing up, crying with eyes open:0.5)`
- The AI is very aggressive about drawing tears/crying.  You may need to reduce the emphasis.

## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 140 images.
  - 131 normal images (9 repeats)
  - 9 "multiple views" images (6 repeats)
    - These were reduced because the AI was generating too many "multiple views" images.
  - 3 batch size, 4 epochs
  - `(131 * 9 + 9 * 6) / 3 * 4` = 1644 steps
- 0.0749 loss
- Initially tagged with WD1.4 swin-v2 model. Tags pruned/edited for consistency.
- `constant_with_warmup` scheduler
- 1.5e-5 text encoder LR
- 1.5e-4 unet LR
- 1e-5 optimizer LR
- Used network_dimension 128 (same as usual) / network alpha 128 (default)
  - Resized to 32 after training
- Training resolution 768x768.
  - Reduced from 832x832. Junko doesn't really have many very fine details that benefit from the higher resolution, and I think training at 832x832 may negatively impact the quality of images generated at lower resolutions.
- Trained without VAE.
- [Dataset can be found on the mega.co.nz repository.](https://mega.nz/folder/SnQDnCRD#ruLvChZGf2vQtWJN87Qs0Q)

## Revisions
- v1c (2023-02-15)
  - Initial release.