Not recommended for realistic models
Main tag : Milking machine,
Secondary tag : nipples,
Pose tag : leaning forward,
DELTA NORM Anima
Unet - Avg = 0.020503
TE - Avg = 0.006345
Training LoRA Powershell info :
train_speed_mode = fast
resolution = 512 [Anima/Illus 768]
Batch size = 2
network_dim = 16
network alpha = 16
mix/save_precision = BF16
optimizer_type = AdamW
optimizer_args_other = foreach=True
GPU RTX 4060Ti 16GB
Torch 2.9.0+cu130 + SDPA+ gradient_checkpointing
compile_mode = max-autotune-no-cudagraphs
Description
LoRA is a beta with reduced DIM.
Increased the number of images trained, but the steps during training remain the same
Update in Beta
No need to tag leaning forward and upper body
Sometimes use the upper body if the resulting image is distorted.
I'll probably add a step in v.1.2 and reduce the LR later