rouwei supplument

~2500 images with teen-adult rating to soft porn photoset.

including 尊みを感じて桜O, 涼森れO, けんけO, 関根ささO, 藤木由O, 小倉優O, 本庄O, pyO eunjO and more

trained with rouwei 0.8 base so the resulting lora can properly communicate with the model, the typical merging method of CyberRealistic/Juggernaut fix doesn't work well because they don't share common language

total of ~20000 steps, nesterov adamw optimizer (torch.optim.nadam), 1e-5 learning rate, linear learning rate scheduling for fine learning rate decay

the dataset is captioned with the following tag:

"

cosplay

cosplay photo

realistic

photorealistic

photorealism

3d

human

person

female

women

girl

"

the cosplay tag distills the hair, accessories, and clothes, into this big umbrella of exotic cosplaying costume concept

the realistic family captioning is easy, it distill away the photorealism real-life look

lastly, the series of human person tag tell the model to look for human body feature, and it follow the mother class token methodology, using the hierarchy of human person > female/women > girl, from base category to sub category

the lora brings real human body proportion to the image, but it cannot help with stable diffusion hands/feet issue, the model still draw four fingers, six fingers, all the time.