# Character LoRA for ANIMA — v1.1

## Version changes

- v1.1 — Added the second outfit variant xiabubu \(shirt\)): fitted white shirt with heart buttons, purple bowtie with heart gem, off-shoulder purple jacket, short green pleated skirt with black waist ribbon, black argyle torn thighhighs, black loafers. Dataset expanded from DATASET_SIZE_V1 to DATASET_SIZE images (all hand-picked, deduplicated, danbooru-style captions).

- Face / hair consistency re-anchored on the new images; white hair, purple eyes, pointy ears, star hair ornament and hair intakes are now stable across both outfits.

- Multi-outfit tagging split into two trigger groups so the two looks do not bleed into each other (see below).

## Trigger words

Identity / face (always use):

xiabubu, white hair, long hair, purple eyes, pointy ears, hair ornament, star (symbol), hair intakes, blush, large breasts

Outfit A — dress ver. (v1.0 reference):

purple bow, bow, sailor collar, white shirt, purple skirt, heart, skirt, white sailor collar, bowtie, school uniform, purple bowtie, upper body, clothing cutout, long sleeves, purple dress, flower, high-waist skirt, black pantyhose, loafers

Outfit B — shirt ver. (this version):

xiabubu \(shirt\), white shirt, purple bowtie, purple bow, heart, jacket, purple jacket, off shoulder, open jacket, long sleeves, pleated skirt, green skirt, ribbon, black ribbon, star hair ornament, hair ribbon, purple ribbon, jewelry, black thighhighs, torn thighhighs, argyle legwear, argyle clothes, zettai ryouiki, loafers, standing, looking at viewer, closed mouth, simple background, white background

Combined prompt for a mixed look: put the identity block first, then pick one outfit block. Do not merge both outfit blocks in a single prompt — the skirt color and legwear will mix.

## Recommended settings

| Item | Value |

|---|---|

| Base model | ANIMA |

| LoRA weight | 0.7 – 1.0 (0.8 sweet spot) |

| Resolution | 832x1216 (portrait) / 1024x1024 |

| CLIP skip | 2 |

| Sampler | Euler a / DPM++ 2M Karras |

| Steps | 28 – 35 |

| CFG | 5 – 7 |

| Hires. fix | 1.5x, denoise 0.40 – 0.50 |

## Negative prompt (suggested)

lowres, bad anatomy, bad hands, extra digits, fewer digits, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry, extra limbs, deformed

## What this LoRA does well

- Front-facing standing full-body and half-body shots; identity holds at 0.7 – 1.0

- Outfit A: sailor-collar white top, purple heart bowtie, high-waist purple skirt with white flower accents, black pantyhose, loafers

- Outfit B: white shirt with heart buttons, off-shoulder purple jacket, green pleated mini skirt, torn black argyle thighhighs, zettai ryouiki, loafers

- Outfit swapping works: keep the identity block, drop the clothing tags, add your own outfit

## Known limitations

- Outfit B is easiest to break: off shoulder + open jacket together can drop one sleeve entirely at weight > 1.0

- torn thighhighs sometimes spreads the tear to both legs; delete that tag if you want clean legwear

- The black waist ribbon in Outfit B can merge into the skirt hem at low weight — raise weight to 0.9 or add ribbon, black ribbon

- Background is not part of the dataset — describe scenes explicitly; simple background alone gives the flat white look from the samples

- No back view in Outfit B; rear angles are guesswork

## Training details

- Images: DATASET_SIZE (Outfit A: N_A, Outfit B: N_B), all captioned with danbooru-style tags

- Repeats: REPEATS, Epochs: EPOCHS, Steps: STEPS

- Network dim / alpha: DIM / ALPHA

- Learning rate: LR (UNet), TE_LR (text encoder)

- Optimizer: OPTIMIZER, Scheduler: SCHEDULER

- Base: ANIMA

## Changelog

- v1.1 — Second outfit (shirt ver.) added, dataset expanded, tag split into identity / Outfit A / Outfit B blocks.

- v1.0 — Initial release with the dress ver. outfit only.