lora-training / mari /README.md
khanon's picture
adds loss values to most LoRAs
6e1c7b7
|
raw
history blame
3.39 kB
# Iochi Mari (Blue Archive)
伊落マリー (ブルーアーカイブ) / 이오치 마리 (블루 아카이브) / 伊落玛丽 (碧蓝档案)
[**Download here.**](chara-mari-v5b.safetensors)
## Table of Contents
- [Preview](#preview)
- [Usage](#usage)
- [Training](#training)
- [Revisions](#revisions)
- [Alternate Version](#alternate-version---v4)
## Preview
![Mari portrait](chara-mari-v5b.png)
![Mari nun](example-001-Treebark-v5b.png)
![Mari gym](example-002-AOM2-v5b.png)
![Mari dress](example-003-DefmixRed-v5b.png)
## Usage
Use any or all of the following tags to summon Mari: `mari, halo, 1girl, blue eyes, orange hair, hair ornament`
- Hair and eye tags are optional.
- It is recommended to add `cat ears` or `animal ear headwear`, otherwise her ears may not appear consistently. This is likely because Mari's veil prevents the AI from strongly associating her ears with the `mari` tag.
- If Mari's ears are appearing through her veil, try adding `animal ear fluff` to the negative prompt.
For her normal outfit: `single braid, animal ear headwear, nun, habit, blue neckerchief, white sailor collar, puffy long sleeves`
For her sportswear alt: `low ponytail, gym uniform, gym shirt, ponytail, white headband, id card, bottle`
- Add `track jacket, sportswear` for her jacket. Use `open jacket, off shoulder` instead when the jacket should be open.
- Use `wet, wet clothes` for transparent/wet clothing
For her swimsuit: `black one-piece swimsuit, frilled swimsuit, frills, paw print`
- You will likely need to add `competition swimsuit` to the negative prompt.
## Training
*Exact parameters are provided in the accompanying JSON files.*
- Trained on a set of 253 images
- 3 repeats
- 3 batch size, 7 epochs
- `(253*3) / 3 * 7` = 1771 steps
- Increased steps for stronger fit due to larger, more varied dataset.
- 0.0823 loss
- 768x768 training resolution
- `constant_with_warmup` scheduler
- Initially tagged with WD1.4 Convnext model, then heavily edited.
- Pruned redundant tags and simplified outfit tags such that Mari'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)
- Trained without VAE.
- [Training dataset available here.](https://mega.nz/folder/2uYkWQiA#GlVNAYXubskusIMbrmVOGQ)
## Revisions
- v5b (2023-02-08)
- 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.
- v4 (2023-01-20)
- Initial release.
- **Issue:** caption files were completely ignored during training. As a result, prompting "mari" grants a very strong effect, but the LoRA is highly overfit.
## Alternate Version - v4
**v4** is also available. This version was improperly configured and ignored caption files, so it was trained on the word `mari` only. It looks quite good, however the model is very overfit and it is difficult to adjust Mari's outfit. In the example image below, the prompt includes `yellow dress` but the model ignores the color and instead uses a dress that resembles Mari's usual outfit.
**[Download v4.](v4/chara-mari-v4-epoch7.safetensors)**
![v4 Preview](v4/example-003-DefmixRed-v4.png)