lora-training / izuna /README.md
khanon's picture
initial commit
7529c6f
# Kuda Izuna (Blue Archive)
## Usage
Use any or all of these tags to summon Izuna:
`1girl, halo, yellow eyes, brown hair, animal ears, fox girl, fox tail, fox hair ornament, bangs, medium breasts, red eyeshadow`
For her normal outfit:
`blue sailor collar, red neckerchief, partially fingerless gloves, ninja, floral print, fishnets, single thighhigh`
For her swimsuit outfit:
`striped bikini, denim shorts, visor cap, highleg bikini, red neckerchief`
You can add or ignore `izuna, blue archive`; while they were in her captions, their effects aren't especially strong.
Weight 0.90-1.2 should work well depending on model.
## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on 67 images curated by anon on /hdg/, repeated 10 times (670 total images / 3 batchsize = ~223 iterations per epoch)
- Dataset included a mixture of SFW and NSFW
- Dataset included a lot of her swimsuit alt outfit but it seems to be able to produce her normal outfit readily.
- Reduced learning rate (3e-5 >>> 1e-5) and increased epochs (1 >>> 3). I'll likely do this from now on since it seems to produce more consistent results.
- Dataset was tagged with WD1.4 interrogator. Shuffling was enabled.
- `izuna, blue archive` were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens.
- No variants, she was a good girl.
## Dataset
Dataset curated by anon on /hdg/: https://mega.nz/folder/Hqg10AAI#9xsrFmuA8oqVBlhsj95y6w
/h/thread/7104198