| # 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. | |