Not recommended for realistic model
Main tag : pretzel pose,
Secondary tag : legs behind head,
LoRA str : 0.8-1.2
Prompt str : 1.0
Denoising str : 0.25 - 0.3
Training Pony LoRA Powershell info :
resolution = 768
Batch size = 3
network_dim = 16
network alpha = 16
mix/save_precision = BF16
optimizer_type = Lion
optimizer_args_other = weight_decay=0.1
GPU RTX 4060Ti 16GB
Torch 2.3+cu121 + Xformers + gradient_checkpointing
Vram used = 10.6GB
Speed = 1.98sec/it
Weight
UNet weight average strength: 0.01048306642834202
Text Encoder (1) weight average strength: 0.008954132907827908
Text Encoder (2) weight average strength: 0.007132517110300672
Description
FAQ
Comments (1)
Good stuff. Most models struggle to even begin to comprehend the concept of anything OTHER than someone's arms being behind their head. This has worked very, very well. At higher weights, it has some style-changing effects to it that can be undesired, but I've found this LoRA to provide enough push for the model to understand the concept even at strengths as low as 0.2. That said, haven't really needed to go below 0.5 for it to completely stop interfering.