Trying to get ArtiWaifu to waifu better
>>75222218>Is that a procedure to make custom stylesyep
Mind you, I'm using very different settings compared what most use here with kohya
LR Schedule: REX
Optimizer: CAME
TE1 LR: 1e-6
TE2 LR: 1e-5
UNet LR: 1e-4
Batch size: as high as you can go, doesn't really matter
Gradient checkpointing: True
Epochs: 1 (increase for smaller datasets)
Accumulation Steps: 1
Rescale Noise Scheduler: True (Should be set to False for SD1 models unless it's specifically a Zero Terminal SNR model) (You can set it to False too, maybe it is ignored for SDXL?)
Loss Weight Function: DEBIASED_ESTIMATION
Loss Scaler: GRADIENT_ACCUMULATION
