[Illustrious] Amagai Ruka (雨海ルカ)

Disclaimer

• My goals in LoRA training is not making fast but I want to experiment some stuff in LoRA training even characater exists in CivitAI.

• Do not use my LoRA to produce AI generated image and using her fanart hashtag in Twitter/X post.

• Don't ???? her...

Description

• CAME with special dataset to enable ability to change any costume even cosplay

• last picture?????

• unet + clip-l only

trigger word

default

amagai ruka, 1girl, solo, looking at viewer, grey hair, blue hair, long hair, multicolored hair, long sleeves, streaked hair, crown braid, hat, shirt, collared shirt, gloves, white gloves, half gloves, blue ribbon, plaid skirt, blue hat, ribbon, boots

blue jacket with hood

amagai ruka, 1girl, solo, looking at viewer, grey hair, blue hair, long hair, multicolored hair, long sleeves, streaked hair, crown braid, blue jacket, hood up

any (add weight prompt to change effectively)

amagai ruka, 1girl, solo, looking at viewer, grey hair, blue hair, long hair, multicolored hair, long sleeves, streaked hair, crown braid

Limitations

• some camera angle is not working properly

Training Details

LoRA type

• standard unet + clip_l only

dataset

• default costume 23 images

• blue jacket with hood 9 images

• others costume 10 images

• 50 images (special dataset to make lora very flexible)

parameters

• resolution = 1024

• batch size = 2

• network dim,alpha = 16,16

• mix/save precision = bf16/bf16

• optmizer = CAME + weight_decay=0.01

• UNet LR = 3.25e-5

• TE LR = 1.6e-5

• scheduler = cosine with min lr (0.25)

• l2 loss

• block weights lr

• alpha mask

steps

• epochs = 10

• total steps = 3400

• repeat = 10,20,22,1 (special dataset to make lora very flexible)

tools

• sd-scripts custom version by me

• torch 2.5.0 cu124

• Runpod RTX 4090

avg weight

• UNet average weights : 0.010260802101005207

• TE1 average weights : 0.007930390653200448