lora-training / reisa /README.md
khanon's picture
adds loss values to most LoRAs
6e1c7b7
|
raw
history blame
No virus
2.19 kB
# Uzawa Reisa (Blue Archive)
宇沢レイサ (ブルーアーカイブ) / 우자와 레이사 (블루 아카이브)
[**Download here.**](https://huggingface.co/khanon/lora-training/blob/main/reisa/chara-reisa-v3.safetensors)
## Table of Contents
- [Preview](#preview)
- [Usage](#usage)
- [Training](#training)
- [Revisions](#revisions)
## Preview
![Reisa portrait](chara-reisa-v3.png)
![Reisa V](example-001-AOM2NMMElyv3.png)
![Reisa pose](example-002-DefmixRed-v3.png)
Unfortunately the model frequently struggles with Reisa's halo.
## Usage
Use any or all of these tags to summon Reisa: `reisa, halo, low twintails, pink hair, purple eyes, star hair ornament`
- Hair and eye tags are optional.
She likes to make her `:d` expression by default. I included images tagged with other expressions, so you should be able to change it, but if not try adding `:d` to the negative prompt.
For her normal outfit: `school uniform, grey shirt, blue jacket, pleated skirt, striped socks, backpack, sneakers`
## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 89 images
- 13 repeats
- 3 batch size, 4 epochs
- `(89*13) / 3 * 4` = 1543 steps
- 0.0743 loss
- 768x768 training resolution
- `constant_with_warmup` scheduler instead of `cosine`
- Initially tagged with WD1.4 Convnextv2 model. Tags were edited for accuracy and improved outfit consistency. Tags for facial expressions, camera angle, and image composition were added.
- Used network_dimension 128 (same as usual) / network alpha 128 (default)
- Trained without VAE.
- [Training dataset available here.](https://mega.nz/folder/jyJAhJ5R#ebcamjoSq1quEZWlUh1qAQ)
## Revisions
- v3 (2023-02-06)
- Completely re-trained to correct issue where caption files were not being loaded correctly. New version should have more flexibility but requires additional outfit prompting.
- v2b (2023-01-28)
- Initial release.
- **Issue:** caption files were completely ignored during training. As a result, prompting "reisa" grants a very strong effect, but the LoRA is highly overfit.
- [Download available on Mega.co.nz repository.](https://mega.nz/folder/yjwGiIBD#w79CxW2-Cj66QmObg0zOdg)