lora-training / iroha /README.md
khanon's picture
initial commit
7529c6f
# Natsume Iroha (Blue Archive)
## Usage
Use any or all of these tags to summon Iroha:
`1girl, halo, red hair, very long hair, wavy hair, grey eyes, small breasts`
For her normal outfit:
`peacked cap, open jacket, collared shirt, partially unbuttoned, red necktie, armband, sleeves past wrists`
You can add or ignore `iroha, blue archive`; while they were in her captions, their effects aren't especially strong.
The network seems simultaneously undertrained and overfitted in some ways so weights are kind of all over the place, just play around with them. 0.8-1.2 work alright.
## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on 95 images using a modified dataset curated by anon on /hdg/, repeated 7 times (665 total images / 3 batchsize = ~222 iterations per epoch, 3 epochs)
- Dataset included a mixture of SFW and NSFW
- Sightly increased learning rate from the last few (1e-5 >>> 1.5e-5)
- Seems a bit underfitted at times, overfitted others.
- Frequently forgets her jacket/shirt colors despite them being a very consistent part of her outfit, and her halo is pretty incoherent.
- She really likes doing a particular kind of toothy smirk that was too common in her dataset. When you don't want it, discourage it with `closed mouth` and negative `open mouth`.
- Dataset was tagged with WD1.4 interrogator. Shuffling was enabled.
- `iroha, blue archive` were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens.
- No variants. I trained 8 iterations of this LoRA and they all had issues of some kind, this latest one is the best of them.
## Dataset
Dataset curated by anon on /hdg/: https://mega.nz/folder/Hqg10AAI#9xsrFmuA8oqVBlhsj95y6w
/h/thread/7104198