lora-training / kazusa /README.md
khanon's picture
adds Haruka and Kazusa to index README
1c4a7b4
# Kyoyama Kazusa (Blue Archive)
杏山カズサ (ブルーアーカイブ) / 쿄야마 카즈사 (블루 아카이브) / 杏山和纱 (碧蓝档案)
[**Download here.**](https://huggingface.co/khanon/lora-training/blob/main/kazusa/chara-kazusa-v1c.safetensors)
## Table of Contents
- [Preview](#preview)
- [Usage](#usage)
- [Training](#training)
- [Revisions](#revisions)
## Preview
![Kazusa preview 1](example-001-v1c-16dim.png)
![Kazusa preview 2](example-002-v1c-16dim.png)
![Kazusa preview 3](example-003-v1c-16dim.png)
![Kazusa preview 4](example-004-v1c-16dim.png)
![Kazusa preview 6](example-006-v1c-16dim.png)
## Usage
- Use any or all of the following tags to summon Kazusa: `kazusa, 1girl, animal ears, colored inner hair, halo, short hair`
- For her normal outfit: `black choker, black jacket, black pantyhose, green sailor collar, hooded jacket, white skirt, miniskirt, hairclip, pink neckerchief, school uniform, sneakers`
- For a closed jacket, you might need to add `open jacket, white shirt` to the negative prompt.
- For raised hood: `hood up`
- For her "trying to look cool" expression: `expressionless, blush, sweatdrop, v-shaped eyebrows, embarrassed`
- For selfies: `selfie, (reaching towards viewer:1.2)`
- For eating: `eating, :t, holding food, food on face`
- Accessories: `macaron, cake, fork`
## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 280 images.
- 7 repeats
- 3 batch size, 4 epochs
- `(280 * 7) / 3 * 4` = 2614 steps
- Kazusa has a lot of good art so her dataset is much larger than usual. This seems to have let me train for longer without overfitting.
- 0.08 loss
- Initially tagged with WD1.4 swin-v2 model. Tags pruned/edited for consistency.
- [Detailed tagging methodology here.](tagging%20methodology.md)
- `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 16 after training
- Training resolution 832x832.
- Trained without VAE.
## Revisions
- v1c (2023-03-02)
- Initial release.