Bunnii O'Hare - all 6 styles (Anima) - character LoRA

SETTINGS

Base Anima-Base v1.0. Trained against Base, so it carries over to

Anima-Turbo, JANIMA and the all-in-one; the reverse is not true.

Strength 1.0

Sampler er_sde Scheduler beta

CFG 4.5 Steps 30

Size 832x1216 portrait, or 1216x832

There is no clip skip. Anima encodes with a Qwen3 text encoder and an

adapter, not CLIP. Any clip-skip value is ignored.

TAG LISTS — one per style

Each style has a costume block that was trained on every single one of its

images. The trigger already carries it: writing it out changes nothing, and

it is reproduced here so you know what the token contains.

bunniibase purple hair, blue eyes, purple lips, white hood, gold trim,

navy blue bodysuit, white chest plate, white gloves,

white armor, gold armor, gold emblem, white boots,

rabbit tail, bodysuit, high heels

bunniihoppers blonde hair, blue eyes, pink lips, white hood, pink trim,

dark teal bodysuit, white chest plate, white gloves,

pink armor, white emblem, pink boots, rabbit tail,

bodysuit, high heels

bunniiauric blonde hair, brown eyes, gold lips, gold hood, orange lining,

black bodysuit, gold chest plate, black gloves, gold armor,

gold emblem, gold boots, rabbit tail, bodysuit, high heels

bunniiphase pink hair, pale lips, iridescent hood, iridescent bodysuit,

holographic, pearlescent, opalescent, iridescent armor,

iridescent gloves, white boots, rabbit tail, bodysuit,

high heels

bunniirime purple hair, cyan lips, white hood, silver trim,

cyan bodysuit, white chest plate, cyan gloves, silver armor,

white emblem, white boots, rabbit tail, bodysuit, high heels

bunniicosmic purple hair, dark purple lips, starfield, galaxy print,

nebula bodysuit, cyan gradient, magenta gradient, dark hood,

dark gloves, dark armor, dark boots, rabbit tail, bodysuit,

high heels

bunnii purple hair, blue eyes, purple lips

(out of costume — no style token, no bodysuit)

WHAT IT DOES — tested, with the pass and fail conditions written down first

Ten conditions, two seeds each, criteria fixed BEFORE generating. Eight

passed, one failed, one works poorly.

Works action and motion; complex environments (night streets, rain,

crowds); civilian clothes with no costume piece bleeding through;

forced anime rendering; stacking with a style LoRA; two characters

in one frame; nudity with correct anatomy; hands holding an object;

alternate styles under scene lighting rather than flat backdrops.

Fails hood down. 1 success in 4 attempts. See below.

Poorly two characters in DIFFERENT styles — you get two figures, but the

colourways bleed into each other and the composition drifts toward

a character sheet.

WHAT IT DOES NOT DO — known defects, named rather than hidden

HOOD DOWN IS UNRELIABLE — 1 success in 4.

hood down appeared on 43 of the 125 in-costume captions against 82 for hood up.

It is not absorbed, it is outvoted. Two things help:

- name NO hood in your prompt. white hood together with hood down gave

0 successes in 2; removing it gave 1 in 2.

- expect to reroll the seed.

THE EAR LINING COMES OUT CYAN ON AURIC AND COSMIC.

Auric's lining should be orange and Cosmic's should be a starfield. Five of

the six styles share a cyan lining in the reference captures, so the minority

cases had no contrast to separate against. No tag fixes this.

bunniiauric AND bunniihoppers CARRY 3D RENDERING AND CANNOT SHED IT.

3d appears on 23/23 auric captions and 18/18 hoppers captions, against

26/29 for base. At 100% a tag is absorbed by the token, so these two styles

will look more like game capture than the other four. This is asymmetric and

it is a limitation, not a preference.

DEFAULT RENDERING IS ANIME, NOT GAME CAPTURE.

Despite 114 of the 143 training images being 3D game captures, the LoRA

renders in anime by default. 28 counter-examples were generated on purpose to

make the render style detachable, and they over-corrected. Whether it can

produce the game look on demand has NOT been tested.

OUT OF COSTUME, IDENTITY IS THIN.

Without the suit the character is carried by hair, eyes and lips only. There

is no bodysuit-free variant of the five Legacy styles: bodysuit sits at

100% in all six style concepts, so it is absorbed by the token. For higher

ratings, OPEN the suit unzipped bodysuit, bodysuit pulled down) rather

than removing it.

HOW IT WAS MADE

Trained with kohya-ss/sd-scripts (anima_train_network) against Anima-Base

v1.0, on a single 8 GB card. 1 h 29, peak VRAM 7997 MB of 8192.

unet-only, dim/alpha 32, CAME (weight_decay 0.01), lr 7e-5 / unet 2e-5,

cosine with 5 warmup steps, sigmoid timestep sampling, l2 loss, 1024 with

aspect-ratio bucketing and no upscaling, batch 1, bf16, 1600 steps, seed 42.

Checkpointed every 200 steps so the epoch could be chosen by looking rather

than by taking the last one.

Dataset: 114 in-game reference captures across the six styles, plus 28

generated images whose only job was to introduce VARIATION the captures did

not have. Without them 3d sat at 114/114 and bodysuit at 114/114 —

constant, therefore absorbed, therefore imposed on every output. With them:

80% and 87%, therefore detachable. Verified afterwards, not assumed.

Seven concepts: the six styles plus one out-of-costume concept that carries

no style token, because there is no style to designate on a woman in a

t-shirt and attaching one would teach a false association.

Captions were built by merging a hand-written fact sheet per style with what

a WD14 tagger actually saw in each image, so the vocabulary describes the

pixels rather than a marketing blurb.