lora-training / hina /README.md
khanon's picture
initial commit
7529c6f
# Sorasaki Hina (Blue Archive)
## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on 106 images curated by anon on /hdg/, repeated 10 times (1060 total images / 3 batchsize = ~354 iterations)
- Dataset included a mixture of SFW and NSFW
- Dataset included a lot of her swimsuit alt outfit, so it has a bit of a harder time producing her standard outfit (the military uniform)
- It does better with her more casual outfit ("sleeveless shirt, miniskirt") and obviously very well with her school swimsuit
- Dataset was tagged with WD1.4 interrogator. Shuffling was enabled.
- `hina, blue archive` were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens.
- Two resolution variants included (512px, 768px)
- The 768px version seems just generally better to my eyes but not overwhelmingly. They're definitely quite different so I've kept both.
## Usage
Hina needs a few tags to be summoned reliably. Some common tags in her dataset:
`1girl, halo, parted bangs, demon horns, long hair, white hair, purple eyes, low wings, small breasts`
You can add or ignore `hina, blue archive`; while they were in her captions, they don't seem to be particularly strong for some reason.
You can use the 512px or 768px variants. I want to say the 768px one is better, but it's hard to say definitively. Give both a shot and post your findings.
Weight 0.85-1.05 should work well depending on model.
## Dataset
Dataset curated by anon on /hdg/: https://mega.nz/folder/ayhWjJDA#xDGSB3T-yRnK6aRajsJ05g
/h/thread/7101537