LoRA model for Asahi from Kantai Collection.

Trained using 49 images. A great thanks to all the artists.

Recommended settings

• LoRA network weight 0.6-1.0. I recommend starting around 0.8 if doing her regular outfit, and 0.6 for other outfits

• Please generate using highres fix for best effect.

• Some methods, especially ancestral samplers (e.g. Euler A), tend to increase the number of moles near her mouth. If this happens, please adjust the prompt weight or change the sampling method

• I recommend using adetailer to inpaint the face for better quality

For LoRA Block Weights:

• Due to the small training set certain poses and background elements are frequent if using all blocks.

• For best flexibility, try OUTD blocks only (1,0,0,0,0,0,0,0,1,1,1,1,0,0,0,0,0), however her back hair bow tends to not appear as well

• For to get the bow to appear (and shape properly), MIDD blocks (1,0,0,0,1,1,1,1,1,1,1,1,0,0,0,0,0) works better

Suggested prompts

General appearance

1girl, grey eyes, french braid, half updo, hair ribbon, white ribbon, hair bow, red bow, mole under mouth,

Regular outfit

1girl, grey eyes, french braid, half updo, hair ribbon, white ribbon, hair bow, red bow, mole under mouth,
white collared shirt, black bowtie, red jacket, open jacket, wide sleeves, long sleeves,
black skirt, high-waist skirt, long skirt, floral print, frilled skirt, black pantyhose, 
cross-laced footwear, black footwear,

• Her base form has a pure white ribbon. For her remodeled (construction ship) form, please add "grey ribbon" to the prompt. It may take some adjusting to come out properly.

Other useful tags & notes

black eyes, blue eyes,
buttons, 
open shirt,
pleated skirt,
wrench, holding wrench,

• Due to her chuuha graphic & others in the training set, if you try to generate only her white shirt it will tend to turn see-through. Please add "see-through" to the negative with enough weight if this is not desired.

• Adding "red clothing" to the negative can help avoid parts of her haori unwantedly appearing on other outfits (assuming those outfits aren't red).

• This LoRA model is fit to a frilled bikini similar to Mamiya's. If you'd like to generate other bikinis please keep the model weight low (around 0.6) and adjust the prompt as needed.

• If viewed from her right, sometimes a second braid will generate. However it's already biased to viewing her from her left side, so this shouldn't happen that often.

• Her in-game Halloween and Christmas outfits are not supported (yet)

On rigging

Due to the limited dataset, I intentionally excluded rigging from the training set to avoid unintentional appearances. If you'd like to attempt generating it, try Fonsion's Asahi LoRA.