khanon's picture
initial commit
7529c6f
|
raw
history blame
2.51 kB
# Kagura Nana, alt outfit (Virtual Youtuber)
Request for anon on /hdg/. Sorta rough, not much art for her and especially NSFW art.
## Usage
Tags: `1girl, cat ears, tail, two side up, two-tone hair, blue eyes, small breasts`
Outfit: `sleeveless dress, off shoulder, jacket, choker, neck bell, striped bow, skirt`
The original version of this was extremely overfit due to bad training data, and was highly resistant towards nudity because there was literally none of it in the dataset. I did some pruning, generated/found a couple of nude images, and cleaned up tags -- now her outfit is manually tagged and exclusively described using the tags above so it is generally more composable (ie. you can prompt for the `jacket` but negative prompt the `sleeveless dress` and it will work fairly reliably). I included `skirt` in images where the flared, ruffled bottom of her dress was visible even though it's not really a skirt but it may help with outfit composition.
It's still slightly overtrained so turning down the weights a bit (0.85 - 1) is recommended.
Three variations:
- v1: the original, in the `old` folder. Not really recommended unless you want it to hyperfocus on her details and outfit at the cost of NSFW capability. Might be okay for mixing?
- v3: new cleaned up dataset, somewhat overcooked
- v4: same as v3 with reduced steps, except for the small number of NSFW images which were boosted
I did something different with this one and tagged the four lewd images in her dataset with `nsfw`. I can't really tell if it helped or hurt, but you may as well prompt for that if you want porn.
## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on a set of 67 images initially provided by an /hdg/ anon, heavily curated and re-tagged.
- Dataset included a mixture of SFW and NSFW.
- Some NSFW images were outputs from v1 of the model. Does not appear to have caused down syndrome Nanas from model inbreeding.
- Dataset was initially tagged with WD1.4 interrogator and then heavily pruned and manually cleaned up.
- `kagura nana, virtual youtuber` were added to the start of each caption. keep_tokens=2 was set to prevent shuffling those tokens.
- All ambiguities or alternate tags for her outfit were removed and replaced with the standard set above. Prompt for `sleeveless dress` not `sleeveless shirt`, and `jacket` not `black jacket` for the highest correlation.
- Three variations with different training params, check the JSON files if you care.