lora-training / sora /README.md
khanon's picture
adds loss values to most LoRAs
6e1c7b7
# Sora (Blue Archive)
ソラ (ブルーアーカイブ) / 소라 (블루 아카이브) / 空 (碧蓝档案)
Be forewarned that this LoRA is prone to generating NSFW results since Sora has an unusually high proportion of extremely lewd art. Use `nsfw, naked apron` etc in the negative prompt to prevent this.
[**Download here.**](https://huggingface.co/khanon/lora-training/blob/main/sora/chara-sora-v3.safetensors)
## Table of Contents
- [Preview](#preview)
- [Usage](#usage)
- [Training](#training)
- [Revisions](#revisions)
## Preview
![Sora portrait](chara-sora-v3.png)
![Sora](example-001-v3-defmix1.png)
![Sora fallen](example-002-v3-aom2nmm.png)
## Usage
Use any or all of the following tags to summon Sora: `sora, halo, mini wings, white wings, two side up, blonde hair, forehead`
For her usual outfit: `blue apron, white shirt, (black pleated skirt:0.8)`
- If needed, you can also add `off-shoulder, strap slip, bowtie, short sleeves`
- Sora's skirt is not visible in most training art. Consider adding `black pleated skirt` and negative prompt `miniskirt, pencil skirt` for more accuracy.
Sometimes her wings appear too high; `low wings` can help with this.
Weight 1 is suggested.
## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 128 images
- 9 repeats
- 3 batch size, 4 epochs
- `(128*9) / 3 * 4` = 1536 steps
- 0.0791 loss
- 768x768 training resolution
- `constant_with_warmup` scheduler instead of `cosine`
- Initially tagged with WD1.4 Convnextv2 model. Tags minimally pruned/edited.
- Used network_dimension 128 (same as usual) / network alpha 128 (default)
- Trained without VAE.
- [Training dataset available here.](https://mega.nz/folder/yjwGiIBD#w79CxW2-Cj66QmObg0zOdg)
## 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 (2023-02-01)
- Initial release.
- **Issue:** caption files were completely ignored during training. As a result, prompting "sora" 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)