loras / styles /kaoming /training notes.md
breakcore2's picture
kaoming
cee2436
|
raw
history blame
438 Bytes

Training Parameters

Dataset consisted of:

272 images repeated 5 times per epoch

Training Parameters:

$learning_rate = 6e-5
$lr_warmup_ratio = 0.10
$train_batch_size = 2
$num_epochs = 2
$save_every_n_epochs=1
$scheduler="constant_with_warmup"
$network_dim=128

notes: This lora was trained with pgs style as a token that was kept with each training example but in practice this is not required to be used for inference.