Bluefield is a god-tier artist who draws a lot of Hololive girls in fashionable outfits. I love their use of colors, their line-art and just the general look of their art. Illustrious somewhat knows the style but it's a bit ass so I made a lora.

Will remove if requested by the artist.

Generation settings

1.0 strength works great in my experience.

Tags

Main tag for Illustrious: Bluefield,

Main tag for Anima: @bluefield,

☝ That's all you need, but the dataset also has some specific outfits (and characters)

Characters

Mumei: nanashi mumei, yellow eyes, ahoge, brown hair, streaked hair, very long hair, feather hair ornament, hairclip, high ponytail, medium breasts,

Suisei: hoshimachi suisei, blue eyes, symbol in eye, star in eye, high side ponytail, light blue hair, medium hair, sidelocks, blue hair ribbon, striped ribbon, small breasts, skinny,

Cali: mori calliope, pink eyes, long hair, pink hair, blunt bangs, large breasts,

Kiara: takanashi kiara, pink eyes, colored inner hair, feather earrings, large breasts, long hair, orange hair, two-tone hair,

Alternatively, check any image of any character on danbooru and use those tags.

Outfits (in no particular order, combine as you like)

crop top, crop top overhang,

white sweater, puffy long sleeves,

black undershirt,

[color] choker,

pendant choker,

white sundress, pleated dress, short dress, spaghetti strap, cleavage, sleeveless,

[color] dolphin shorts, [color] trim,

[color] yoga pants, [color] trim,

high heels, black stiletto heels, strappy heels, pumps, ankle bow,

colored shoe soles, [color] soles,

Training details

Illustrious

Base model: Illustrious XL 0.1

41 images, 2 repeats + flip augmentation, 20 epochs

16 batch size

8 dim, 8 alpha

0.75 unet and text encoder learning rate

Prodigy optimizer, args: "decouple=True weight_decay=0.01 betas=[0.9,0.999] d_coef=2 use_bias_correction=True safeguard_warmup=False"

Anima

Base model: Anima Base v1.0

41 images, 3 repeats, no flip aug, 22 epochs

16 batch size

16 dim, 16 alpha

0.0007 unet learning rate, 0 TE/LLM learning rate

AdamW8bit optimizer, cosine scheduler

0.05 IP Noise Gamma

If you like the model, upload some images!