File size: 2,264 Bytes
ccb87c0 a1bcd94 ccb87c0 acdbfd2 ccb87c0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 |
# Arona (Blue Archive)
アロナ (ブルーアーカイブ) / 아로나 (블루 아카이브) / 阿罗娜 (碧蓝档案)
[**Download here.**](https://huggingface.co/khanon/lora-training/blob/main/arona/chara-arona-v1.safetensors)
## Table of Contents
- [Preview](#preview)
- [Usage](#usage)
- [Training](#training)
- [Revisions](#revisions)
## Preview
## Usage
Use any or all of the following tags to summon Arona: `arona, 1girl, white hairband, bow hairband, halo, short hair, single braid`
- You can also add `blue eyes, two-tone hair, multicolored hair, blue hair, pink hair` if necessary.
- `mechanical halo` can sometimes improve the quality of Arona's halo.
- I tried tagging her different halo types (`heart halo`, `red halo`, etc.) but the AI doesn't recognize them very well.
For her normal outfit: `school uniform, blue shirt, white sailor collar, long sleeves, white bowtie, white choker`
- Images where her stomach was visible were tagged `midriff peek` or `midriff`. The AI seems to like showing her this way by default, but you can prompt those tags to strengthen the effect (or negative prompt them to try to remove it).
For her swimsuit outfit: `swimsuit, official alternate costume, black bikini, frills, frilled bikini`
Arona frequently appears in the ruined classroom, which I have tried to tag consistently. Use `ruined classroom, school desk, shallow water, cloud` to produce this background.
Not all tags may be necessary.
## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 250 images.
- 5 repeats
- 3 batch size, 4 epochs
- `(250 * 5) / 3 * 4` = 1667 steps
- Initially tagged with WD1.4 convnextv2 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)
- New: increased training resolution to 832x832.
- Trained without VAE.
- [Dataset can be found on the mega.co.nz repository.](https://mega.nz/folder/j25mVDxL#fN3r1SPNw77QFs9oaeehLw)
## Revisions
- v1 (2023-02-10)
- Initial release.
|