[Illustrious/NoobAI] dooby3D

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, ...) !!!!!!!!

Description V1

• Detective Dooby.

• can change costume

Description

• Test with LoRA GGPO (test version)

• First NoobAI v-pred version LoRA / with illustrious too

trigger word

with white shirt

dooby3d, 1girl, solo, mouse girl, blonde hair, black hair, white hair, multicolored hair, green eyes, tail, short hair, gloves, checkered gloves, blue overalls, goggles on headwear, shirt, short eyebrows, hair ornament, blue hat, green neckerchief, medium breast

without white shirt (accidentally remove)

dooby3d, 1girl, solo, mouse girl, blonde hair, black hair, white hair, multicolored hair, green eyes, tail, short hair, gloves, checkered gloves, blue overalls, goggles on headwear, bare shoulders, collarbones, bare arms, checkered tube top, short eyebrows, hair ornament, blue hat, medium breast

any costume

dooby3d, 1girl, solo, mouse girl, blonde hair, black hair, white hair, multicolored hair, green eyes, tail, short hair, short eyebrows, hair ornament, medium breast

Limitations

• no [v1]

• only 1 costume [v0.1]

Training Details [V1 Illustrious]

LoRA type

• LoRA GGPO

dataset

• 75 images

parameters

• resolution = 768

• batch size = 2

• network dim,alpha = 16,8

• mix/save precision = bf16/bf16

• optmizer = AdamWSchedulefree

• UNet LR = 1.75e-4

• TE LR = 8.75e-5

• scheduler = no

• huber loss

steps

• epochs = 10

• total steps = 3000

• repeat = 8

tools

• sd-scripts

• torch 2.7.0 cu128

• RTX 3060 12 GB

avg weight

• UNet average weights : 0.00988162469796159

• TE1 average weights : 0.008170449989847839

Training Details [v0.1 NoobAI]

LoRA type

• LoRA GGPO

dataset

• 70 images

parameters

• resolution = 1024

• batch size = 2

• network dim,alpha = 12,12

• mix/save precision = bf16/bf16

• optmizer = CAME + weight_decay=0.01

• UNet LR = 4.5e-5

• TE LR = 2e-5

• scheduler = cosine with restarts (3 times)

• l2 loss

• extra args

--v_parameterization --zero_terminal_snr

steps

• epochs = 10

• total steps = 1750

• repeat = 5

tools

• sd-scripts custom version by me

• torch 2.5.0 cu124

• RTX 3060 12 GB

avg weight

• UNet average weights : 0.00935241486877203

• TE1 average weights : 0.008072614941435555

Training Details [v0.1 Illustrious]

LoRA type

• LoRA GGPO

dataset

• 70 images

parameters

• resolution = 1024

• batch size = 2

• network dim,alpha = 12,12

• mix/save precision = bf16/bf16

• optmizer = CAME + weight_decay=0.01

• UNet LR = 4.5e-5

• TE LR = 2e-5

• scheduler = cosine with restarts (3 times)

• l2 loss

steps

• epochs = 10

• total steps = 1750

• repeat = 5

tools

• sd-scripts custom version by me

• torch 2.5.0 cu124

• RTX 3060 12 GB

avg weight

• UNet average weights : 0.009963956746188078

• TE1 average weights : 0.008257410062166551

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

*If you see my model reposted on other image generation service please report to me.