Yuuki Sakuna (結城さくな) - Illustrious LoRA (2 Costume)

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.
NO REPOST IN ANY WEBSITE (PixAI, Tensor . Art, Seaart, ...) !!!!!!!!

Version 3 description

New Dataset and captions
2nd costume is coming
Completely change any costume
Throw LoCon welcome UNet + CLIP-L (maybe future I will fork lycoris to support UNet + Conv + CLIP-L only)
Completely use only her costume to cosplay by another character.

Trigger Word (V3)

default costume

yuuki sakuna, yuuki sakuna \(1st costume\), 1girl, long hair, blush, animal ears, pink hair, medium breasts, hair ornament, maid headdress, bow, maid, animal ear fluff, pink bow, puffy sleeves, apron, white dress, pink dress, blue ribbon, bowtie, collarbone, hairclip, hair bow, detached collar, white thighhighs, black footwear, shoes, black choker

2nd costume (maybe you can add some prompt for better output) some parts still missing due lack of dataset.

yuuki sakuna, yuuki sakuna \(2nd costume\), 1girl, cat ears, short hair, pink hair, hair ribbon, pink eyes, choker, dress, white dress, purple dress, hair ornament, hairclip, bandaid hair ornament, long sleeves, bow, frills, white apron, handbag, wings

any costume

yuuki sakuna, 1girl, long hair, blush, cat ears, pink hair, medium breasts, hair ornament, blue ribbon, hairclip, hair bow

hairstyles (you can select ones)

long hair, two sides up
short hair
single braid

Limitations

Second Costume outputs still not good enough.

Training Details (V3)

Network Module

standard LoRA with clip-l only

Dataset

1st costume 313 images
alternative costume 63 images
2nd costume (short hair) 17 images
2nd costume (single braid) 12 images
special dataset for unlocking ability to change any costume 50 images

Parameters

resolution = 1024
batch size = 2
dim,alpha = 16/16
mix/save precision = bf16/bf16
optmizer = CAME + weight_decay=0.01
UNet LR = 3.2e-5
TE LR = 1.6e-5
scheduler = cosine_with_min_lr (down to 0.25 ratio)
alpha mask
block weights LR
l2 loss

Steps

epochs = 10
total steps = 5580
repeats = 1,5,10,14,3

Tools

custom kohya-ss GUI
torch 2.7.0 cu128 + xformers
RTX 3060 12 GB

Average Weights

UNet average weights = 0.00975080994381146
TE1 average weights = 0.008148421067744493

PREVIOUS VERSION

Version 2 description

Big Improvement when changing costume (almost 99% can change into any costume except maid)

New Dataset
using Lycoris LoCon instead of LoRA

Version 1 description

First Illustrious LoRA
using Lion Scheduler

Trigger Word

default costume

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

remove black footwear, shoes, black choker, white thighhighs for half body

any costume

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

Limitations

nothing (if found please comment below.)

Training Details (V2)

LoRA method

Lycoris LoCon

Dataset

256 images

parameters

resolution = 1024
batch size = 2
dim,alpha = 16,12
ConvDim, ConvAlpha = 16,12
mix/save precision = bf16/bf16
optmizer = Lion + weight_decay=0.1
UNet LR = 4e-5
TE LR = 1e-5
scheduler = cosine_with_min_lr
l2 loss

steps

epochs = 10
total steps = 2560
repeats = 2

tools

kohya-ss GUI v24.3.0
torch 2.5.0 cu124
Runpod RTX 4090 24 GB + xformers + gradient_checkpointing

weight

UNet average weights : 0.009442168148234486
TE1 average weights : 0.007911427761428058
TE2 average weights : 0.0064397353562526405

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

*REPOST model is not allow if you found this model please report to me.