hibiki/README.md

# Nekozuka Hibiki (Blue Archive)

## Usage
*Important:* This is a fairly temperamental LoRA due to the dataset, and it needs some wrangling to get good results.  It won't look good with vanilla NAI and the standard NAI negative prompt, despite being trained on nai-animefull-final.
- Use a strong negative prompt.  Consider using the bad_prompt_v2 embed at a reduced strength to dramatically improve things, though it does affect the style somewhat.
- Use a strong model.  AbyssOrangeMix2 and nutmegmix work well with this LoRA.
- Use the negative prompt liberally to suppress cheerleader Hibiki if you don't want her, otherwise her traits tend to take over.

To summon Hibiki, use the following tags.  Adjust strength as needed.
  - `1girl, halo, black hair, blue eyes, bright pupils, animal ears, dog girl, tail, hair bobbles, goggles, eyewear on head, medium breasts`

For regular Hibiki, add the following tags.  Adjust strength as needed.
  - Prompt: `(black camisole, jacket, black shorts:1.2), (fishnet legwear:1.1)`
  - Negative prompt: `(midriff, navel, skin tight, tight:1.4), (tattoo, arm tattoo, star sticker:1.30), ski goggles, wavy mouth, embarrassed`
    - Cheerleader Hibiki tends to have a permanent embarrassed/wavy mouth expression unless you use negative tags to get rid of it.

For cheerleader Hibiki, add the following tags.
  - Prompt: `cheerleader, midriff, embarrassed, wavy mouth, crop top, white pleated skirt`
  - Negative prompt: `fishnets`

You can add or ignore `hibiki, blue archive`; while they were in her captions, they don't have an especially strong effect.

Adjust the weight as needed.  Weights from 0.95 up to 1.25 work well (higher weights may summon Cheerleader Hibiki unintentionally).

## Training
*All parameters are provided in the accompanying JSON files.*
- Trained on 119 images, separated into two sets.
  - Hibiki (regular) -- 54 images, 10 repeats
  - Hibiki (cheerleader) -- 65 images, 6 repeats
  - Dataset included a mixture of SFW and NSFW.  Mostly SFW.
  - As you can guess, her cheerleader alt makes up the vast majority of her art, and artists are more consistent when drawing her. Training them all together did not work well, so I had to split up the datasets.
- Dataset was tagged with WD1.4 interrogator.  Shuffling was enabled.
  - `hibiki, blue archive` were added to the start of each caption.  keep_tokens=2 was set to prevent shuffling those tokens.
  - Her cheerleader outfit is much more easily recognized by the tagger, leading to stronger tags.  Even human artists can't seem to agree on what she's wearing in her normal outfit.
- Trained at 768px resolution.  I stopped training a 512 variant because it was almost always worse and added to training time.
- Two variants included.
  - v1: first attempt at splitting the dataset.  Works well, but not as coherent in some ways (halo particularly) and still tends to blend details from her two outfits.
  - v2: using the split dataset. Increased batch size slightly (3 >>> 4), reduced learning rate (3e-5 >>> 1e-5), increased epochs (1 >>> 3).  Generally an improvement, though sometimes v1 is easier to wrangle.