Steps/Loss: 1428*2

Datasets: 34

Repeat: 14

Epoch: 6

Clip Skip:

learning rate: Both

Text Encoder: 0.00001

UNET: 0.0002

LR Scheduler: cosine

Optimizer: AdamW8bit

DIM/ALPHA: 32/32

Gradient Accumulation Steps: 2

Noise offset: 0.05

Multires noise discount: 0.3

Multires noise iterations: 10

Others: