
This is my first upload of a Lora I trained after getting into SD and subsequently Lora training. So I am pretty new at this. Feedbacks welcome.
Some details on the Lora settings:
Trained with around 137 photos of Amau Kisumi, a popular Japanese idol. Photos are all publicly available on her Twitter site, please support and follow her -> https://twitter.com/amau_kisumi
Kohya Trainer
-
Base Model: chilloutmix
-
Network dim/alpha: 32/16
-
learning_rate: 1e-4
-
text_encoder_lr: 5e-5
-
lr_scheduler: constant
-
train_batch: 6
-
num_epochs: 15 (but I took epochs #12 after several testing)
-
steps: 15
Most of the parameters I just follow by reading a combination of a few guides. If you have any wisdom to share, please do.
To use please:
<lora:amaukisumi-v4:0.8> or <lora:amaukisumi-v4:1> amaukisumi
I found best results between 0.8 and 1.0, works best with chilloutmix.
But due to most of her photos being gravure, you may supress NSFW by putting in Negative prompts like "lingerie, nude, naked", and specify clothes style in Positive Prompts.
Also her chest might be a bit on big side, so put like (small perfectly shaped breast)
描述:
训练词语: amaukisumi
名称: amaukisumi-v4.safetensors
大小 (KB): 36981
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success