Yuuki Sakuna (結城さくな) - Pony LoRA

Disclaimer

• Do not use my LoRA to produce AI image and tagging her fanart hashtag in Twitter/X or if you have an enough contribution in that image then fine.

Version 3 description

• reduce dim/alpha to 16/12

• using scale weight norm to 1.5 (prevent overfitting)

• change scheduler

• new dataset + captions

• her smile is less become :3 instead of :)

Trigger Word (V3 only)

• default costume

yuuki sakuna, long hair, blush, animal ears, pink hair, medium breasts, hair ornament, maid headdress, bow, maid, animal ear fluff, puffy sleeves, apron, white dress, bowtie, collarbone, hairclip, hair bow, detached collar, black footwear, shoes, black choker, white thighhighs

• any costume

yuuki sakuna, long hair, blush, cat ears, pink hair, medium breasts, hair ornament, bowtie, hairclip, hair bow

Limitations

• some costume cannot change properly (due to lack of datasets)

• full body still inaccurate (but much improvement)

Version 2 description

• using Experimental Optimizer (it seems improvement on full body but camera angle still not variable if not prompt it)

• bow still remains even change outfit

Version 1 description

• I like cat ears girl so, I trained as LoRA (don't know about her past at all :<)

Trigger Word

• trigger word

yuuki sakuna

• any costume (still not flexible enough)

yuuki sakuna, long hair, animal ears, pink hair, blush, cat ears, pink eyes, two side up, ahoge, colored inner hair, two-tone hair

• debut costume (good but some component still missing)

yuuki sakuna, long hair, hair ornament, bow, animal ears, pink hair, blush, cat ears, maid headdress, hair bow, frills, hairclip, pink eyes, pink bow, blue bow, maid, puffy sleeves, two side up, cat hair ornament, ahoge, heart hair ornament, puffy short sleeves, clothing cutout, pink dress, blue bow, colored inner hair, two-tone hair, cleavage, breasts

• for full body add this following text will help (shoes still not correct)

shoes, black footwear, white thighhighs

Limitations

• Cannot change costume properly (still have some debut costume component leftover)

• full body may be not effective

• LoRA still little bit underfitting (like medium rare pork) (in version 1)

• version 2 is improve some small detail but dataset still not variable enough (due to imbalance image)

Training Details (Version 3)

LoRA size

• reduced dimension to 8 with sv_fro=0.95

dataset

• 42 images

parameters

• resolution = 1024

• batch size = 2

• dim,alpha = 16,12 (for training)

• mix/save precision = bf16/bf16

• optmizer = AdEMAMix + weight_decay=0.025 betas=0.9,0.999,0.9999

• UNet LR = 2e-4

• TE LR = 1e-4

• scheduler = cosine_with_min_lr (min_lr_ratio 0.67)

• huber snr 0.85

steps

• epochs = 5

• total steps = 1575

• repeat = 15 (one concept only)

tools

• kohya-ss GUI v24.3.0 (Forked by me)

• torch 2.5.0 cu124

• RTX 3060 12 GB + xformers + gradient_checkpointing

weight

• UNet average weights : 0.0149531283161857

• TE1 average weights : 0.011002991641968642

• TE2 average weights : 0.009832777519477531

Training Details (Version 2)

LoRA size

• reduced dimension to 8 with dynamic alpha

dataset

• 38 images (most is half body)

parameters

• resolution = 1024

• batch size = 2

• dim,alpha = 16,16 (for training)

• mix/save precision = bf16/bf16

• optmizer = AdEMAMix (32 bit consume VRAM)

• UNet LR = 2e-4

• TE LR = 1e-4

• scheduler = inverse_sqrt with warmup 100 steps

• l2 loss only

steps

• epochs = 10

• total steps = 2850

• repeat = 15 (one concept only)

tools

• kohya-ss GUI v24.2.0

• torch 2.5.0 cu124

• RTX 3060 12 GB + xformers + gradient_checkpointing

weight

• UNet weight average strength = 0.015634962041489377

• Text Encoder (1) weight average strength Clip_L = 0.011193290141749815

• Text Encoder (2) weight average strength Clip_G = 0.010691167576002698

Training Details (Version 1)

dataset

• 38 images (most is half body)

parameters

• resolution = 1024

• batch size = 2

• dim,alpha = 16,16 (no resize for preserving quality if LoRA is good enough will do it :P)

• mix/save precision = bf16/fp16 (accidentally change)

• optmizer = AdEMAMix8bit

• UNet LR = 1e-4

• TE LR = 5e-05

• scheduler = inverse_sqrt with warmup 100 steps

• huber snr with c = 0.85

steps

• epochs = 10

• total steps = 2850

• repeat = 15 (one concept only)

• full_bf16 training

tools

• kohya-ss GUI v24.2.0

• torch 2.5.0 cu124

• RTX 3060 12 GB + xformers + gradient_checkpointing

weight

• UNet weight average strength = 0.008335085569112463

• Text Encoder (1) weight average strength Clip_L = 0.0073367764333498705

• Text Encoder (2) weight average strength Clip_G = 0.005826970830639767

description ref from Gtonero

*This LoRA is for studying LoRA training with new technique so do not use for damaging the vtuber (also support her too).