# Nemugaki Fubuki (Blue Archive)
合歓垣フブキ (ブルーアーカイブ) / 네무가키 후부키 (블루 아카이브) / 合歓垣吹雪 (碧蓝档案)

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

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

## Preview
![Fubuki portrait](chara-fubuki-v1b.png)
![Fubuki preview 1](example-001b-v1b.png)
![Fubuki preview 2](example-002b-v1b.png)
_[Uses this first-pass ControlNet input (scribble preproessor)](example-002-controlnet-input.png)._

![Fubuki preview 3](example-003b-v1b.png)

## Usage
Use any or all of the following tags to summon Fubuki: `fubuki, 1girl, twintails, streaked hair, antenna hair, heart hair ornament, red eyes`

For her normal outfit: `police uniform, white jacket, blue vest, blue necktie, collared shirt, pencil skirt, black pantyhose, red sneakers`

For her accessories: `doughnut, holding food, police badge, traffic baton, police hat`

For her smug expression: `smug, jitome, smile` + `closed mouth` / `open mouth`
- `jitome` refers to the flat-topped eye shape and can be used without `smug` if you prefer.
- Using `smug` without specifying `closed mouth` or `open mouth` tends to look strange, as though the AI tries to do both.

[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 137 images.
  - 72 without "holding food" (10 repeats)
  - 65 "holding food" (8 repeats)
    - Fubuki is so commonly seen holding a doughnut that these needed to be reduced to avoid overfitting.
  - 3 batch size, 4 epochs
  - `(72 * 10 + 65 * 8) / 3 * 4` = 1654 steps
- 0.083 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.
  - For some reason, 832 performed better than 768 on Fubuki's LoRA, unlike the Junko's.  The difference was not substantial, however.
- Trained without VAE.
- [Training dataset available here.](https://mega.nz/folder/b25mDTIA#lGdhQu6tBUGXco6-aPEAOg)

## Revisions
- v1b (2023-02-18)
  - Initial release.