# Yuragi Momoka (Blue Archive)
由良木モモカ (ブルーアーカイブ) / 유라기 모모카 (블루 아카이브) / 由良木桃香 (碧蓝档案)

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

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

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

## Usage
Use any or all of the following tags to summon momoka: `momoka, halo, short twintails, horns, bright pupils, pointy ears, hair ornament, ahoge`
- Add `(dragon tail:1.3)` for her tail (even though I'm not quite sure Momoka is truly a dragon?)

For her normal outfit: `sleeveless dress, collared dress, blue necktie, white open jacket, off shoulder, loose socks, white shoes`
- Add `frilled dress` if the frills at the bottom of her dress are not correctly displayed.

For her accessories: `potato chips, bag of chips, holding food`

For her smug expression: `smug, open mouth, sharp teeth, :3, :d`
- Alternatively, `smug, grin, sharp teeth, smile` for a toothy grin

[Here is a list of all tags including in the training dataset, sorted by frequency.](all_tags.txt)

## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 94 images.
  - 13 repeats
  - 3 batch size, 4 epochs
  - `(94 * 13) / 3 * 4` = 1654 steps
- 0.0737 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 24 after training
  - This LoRA seemed very slightly overtrained, perhaps due to smaller dataset, so resizing to 24 appeared a bit better than 32.
- Training resolution 832x832.
  - This one also came out better at 832 vs 768.
  - It's not clear to me why some LoRAs perform substantially better at 768 and others at 832.
- Trained without VAE.
- [Training dataset available here.](https://mega.nz/folder/fi5zxDpb#J6ABI5i8ZFnTONVYiRlKHg)

## Revisions
- v1c (2023-02-19)
  - Initial release.