Gawr Gura Pony LoRA

description

• trained with more concept with experimental parameters.

• This LoRA is for studying LoRA training with new technique so do not use for damaging or any harmful the character and original creator.

Trigger Wordmain trigger word : gawrgura

Gura Debut Costume :gura1st, two side up, blue eyes, blunt bangs, multicolored hair, grey hair, blue hair, streaked hair, blue hoodie, shark hair ornament, fins, shark tail, long sleeves

with hood :

shark hood, hood, blue hoodies

with cat ears :

cat ears

• sometime it's random appear if you dont want put it in negative prompt

Casual :guracasual, cat ears, blue eyes, multicolored hair, blue hair, grey hair, streaked hair, cube hair ornament, side ponytail, blue dress, white shirt, shoulder bag, sleeveless, shark tail

Party Dress :guraparty, shark tail, blue eyes, multicolored hair, blue hair, grey hair, streaked hair, hair ornament, gradient dress, collarbone, jewelry, twintails, blue nails, jewel under eyes, gradient clothes

• some color might be inaccurate

New Year :guranewyear, two side up, multicolored hair, streaked hair, blue hair, grey hair, blue eyes, blue kimono, hair ornament, shark tail, detatched sleeves, fur collar, earmuffs, ahoge, tail ornament

• some color might be inaccurate

School Uniform :without jacket :

guraschool, one side up, multicolored hair, streaked hair, blue hair, grey hair, blue eyes, shark tail, yellow vest, blue skirt, blue necktile, bandaid on leg, hair ornament

with jacket :

guraschool, one side up, multicolored hair, streaked hair, blue hair, grey hair, blue eyes, shark tail, blue jacket, blue skirt, blue necktile, bandaid on leg, hair ornament

any costume :gawrgura, blue eyes, blue hair, grey hair, multicolored hair, alternate costume

• and add anything prompt

Gura Teeth :open mouth, sharp teeth

• If you want her open mouth and upper teeth.

Limitations

• guranewyear and guraparty full body might be not effective.

• some costume details is not accurate

Training DetailsLoRA size

• resize dimension to dynamic (max 8) with dynamic alpha

dataset

• debut outfit (no cat ears) : 50 images

• debut outfit (cat ears) : 15 images

• casual : 25 images

• new year : 15 images

• party dress : 25 images

• school uniform (without jacket) : 10 images

• school uniform (with jacket) : 6 images

parameters

• resolution = 102>

• batch size = 2

• dim,alpha = 16,16 (for training)

• mix/save precision = bf16/bf16 -> resize to fp16

• optmizer = AdEMAMix

• UNet LR = 2e->

• TE LR = 1e->

• scheduler = inverse_sqrt with warmup 100 steps

• l2 loss only

steps

• epochs = 10

• total steps = 6500

• repeat = 3,10,6,10,10,15,25

tools

• kohya-ss GUI v24.2.0

• torch 2.5.0 cu12>

• RTX 3060 12 GB + xformers + gradient_checkpointing

weight

• UNet weight average strength = 0.01693913135115065

• Text Encoder (1) weight average strength Clip_L = 0.011991257262133617

• Text Encoder (2) weight average strength Clip_G = 0.0118015129444758