I have created a Locon model for Fujiwara Chika from Kaguya-sama: Love Is War. While there is already a Lora model available for her, it struggles with drawing the hair bow on her bangs.
My Locon model is designed to be easy to use, requiring fewer trigger words and making it easier to draw the hair bow. For generating original character designs, I suggest using a weight of 0.8-1.
If you want to change her clothing and use keywords to describe her features, I recommend using a weight of 0.5-0.6.
Here are my training configurations:
Training data: 59 images
Based model: anything-v3-fp32-pruned
network_dim: 40
network_alpha: 40
conv_dim: 40
conv_alpha:40
optimizer_type: DAdaptation
mixed_precision: fp16
clip_skip: 2
learning_rate: 1
unet_lr: 1
text_encoder_lr: 0.5
max_train_steps: 600
lr_warmup_steps: 50
batch size: 4