lora-training / chise /README.md
khanon's picture
adds loss values to most LoRAs
6e1c7b7
|
raw
history blame
2.71 kB
# Waraku Chise (Blue Archive)
和楽チセ (ブルーアーカイブ) / 와라쿠 치세 (블루 아카이브) / 和楽千世 (碧蓝档案)
[**Download here.**](chara-chise-v2.safetensors)
## Table of Contents
- [Preview](#preview)
- [Usage](#usage)
- [Training](#training)
- [Revisions](#revisions)
- [Alternate Version](#alternate-version---v1)
## Preview
![Chise portrait](chara-chise-v2.png)
![Chise moon](example-001-Defmix-v2.png)
![Chise calligraphy](example-002-Defmix-v2.png)
![Chise swimsuit](example-003-Defmix-v2.png)
## Usage
Use any or all of these tags to summon Chise: `chise, halo, oni horns, hair flower, red eyes, blue hair`
For her normal outfit: `braid, japanese clothes, detached sleeves, bridal gauntlets, obi, tabi, geta, sailor collar, blue bow`
For her swimsuit alt: `side ponytail, swimsuit, striped bikini, see-through, see-through shirt, white sailor collar, side-tie bikini bottom`
- You may also add `pointy ears`, which are usually only visible in her swimsuit outfit.
## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 119 images; 76 swimsuit, 43 normal.
- 15 repeats for normal outfit
- 9 repeats for swimsuit outfit
- 3 batch size, 4 epochs
- `(76*9 + 43*15) / 3 * 4` = 1772 steps
- 0.0837 loss
- 832x832 training resolution
- `constant_with_warmup` scheduler
- Initially tagged with WD1.4 Convnext-v2 model, then heavily edited.
- Pruned redundant tags and simplified outfit tags such that Chise's outfit was always described with the same words
- Ensured unique traits such as `halo` were consistent with actual visibility
- Added many tags for facial expressions, camera angles, and image composition
- Used network_dimension 128 (same as usual) / network alpha 128 (default)
- Resized to 32dim after training
- Trained without VAE.
- [Training dataset available here.](https://mega.nz/folder/XzJ2hZ4Y#R_IEY-uO664-t6Vk935B-A)
## Revisions
- v5b (2023-02-12)
- 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.
- v1 (2023-01-26)
- Initial release.
- **Issue:** caption files were completely ignored during training. As a result, prompting "chise" grants a very strong effect, but the LoRA is highly overfit.
## Alternate Version - v1
**v1** is also available. This version was improperly configured and ignored caption files, so it was trained on the word `chise` only. It looks quite good, however the model is very overfit and it is difficult to adjust Chise's outfit.
**[Download v1.](v1/chara-chise-v1.safetensors)**
![Chise v1 Preview](v1/chara-chise-v1.png)