lora-training / miyako /README.md
khanon's picture
initial commit
7529c6f
# Tsukiyuki Miyako (Blue Archive)
## Usage
Use any or all of these tags to summon Miyako:
`1girl, halo, long hair, one side up, bangs, silver hair, purple eyes, fake animal ears`
She really doesn't need many, just `1girl, halo` is enough to summon an instantly recognizable version of her. The model may be a little overtrained but it looks good.
For her SRT outfit:
`school uniform, neckerchief, blue skirt, bulletproof vest`
The model is very slightly overtrained so weights in the range of 0.8 - 1 are ideal.
## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on 85 images on a modified dataset curated by an anon on /hdg/, repeated 21 times (1785 total images / 3 batchsize = ~595 iterations per epoch, 1 epoch)
- Dataset included a mixture of SFW and NSFW.
- 1.5e-5 learning rate.
- Used "cosine_with_restarts" scheduler instead of "constant_with_warmup", which seems to have helped address some coherency issues
- Dataset was tagged with WD1.4 interrogator. Shuffling was enabled.
- `miyako, blue archive` were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens.
- Two variants, identical except for VAE
- v3-NAI: uses NAI VAE
- v4-WD: uses WD1.4 (kl-f8-anime2) VAE