!! thanks for sharing - how many pics in your training dataset and how many training steps?
DDD
U
D
JJD
D
�
��This LoRA was trained on two layers single_transformer_blocks.7.proj_out and single_transformer_blocks.20.proj_out
It is possible to train on blocks.7.proj_outlayer only if you don't want the dataset style transfer, the size depending on the dim can be below 4.5mb (bf16), the model is powerful enough to offer advantages such as single layer training, so let's do that.
A Flux LoRA with 99% face likeness with great flexibility can be as small as 580kb (single layer, dim 16)
NN�
МN
M
FFFFFFFFFF
VV
allow_pickle=False to raise an error instead."V
DDDDFFF
TFFFTFFFD
D
D
allow_pickle=False