# Yurizono Seia (Blue Archive)
百合園セイア (ブルーアーカイブ) / 유리조노 세이아 (블루 아카이브) / 百合園圣娅 (碧蓝档案)

[**Download here.**](https://huggingface.co/khanon/lora-training/blob/main/seia/chara-seia-v1b.safetensors)

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

## Preview
![Seia portrait](chara-seia-v1b.png)
![Seia preview 1](example-001-DefmixRed-v1b.png)
![Seia preview 2](example-003-DefmixRed-v1b.png)

## Usage
Use any or all of the following tags to summon Seia: `seia, 1girl, fox ears, parted bangs, halo, multicolored eyes, flower wreath, blonde hair`

For her normal outfit: `white dress, sleeveless dress, blue necktie, sleeves past wrists, sleeves past fingers, detached sleeves, white pantyhose, high heels`
- Her dress was also tagged with `collared dress` and `frilled dress` if those elements were visible.
- Choose `sleeves past wrists` or `sleeves past fingers` depending on whether her fingers should be visible.
- Negative prompt `detached sleeves` for alternative outfits where she should not be wearing the dress.

For her expressions: `expressionless` / `light smile` / `parted lips`
- Use `half-closed eyes`, `sleepy` or `tareme` if her eyes are too wide.

## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 96 images.
  - 13 repeats
  - 3 batch size, 4 epochs
  - `(96 * 13) / 3 * 4` = 1664 steps
- 0.0764 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 832x832.
- Trained without VAE.
- [Dataset can be found on the mega.co.nz repository.](https://mega.nz/folder/SupXFCiC#C992OYnVa8svZt3PdUObww)

## Revisions
- v1b (2023-02-12)
  - Resized v1 from network rank 128 >>> 32. This has no negative impact on image quality/coherency, and provides these advantages:
    - Reduced LoRA file size from 144mb to 36mb
    - Slightly reduced overfitting on training data
      - Less darkening/vignetting of backgrounds
      - Improved color saturation
      - Halo is a bit more consistent
      - Seia's hair gets confused with details on her clothing less frequently
- v1 (2023-02-12)
  - Initial release.