experiment:
unet_lr = 0.0003
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 0.08
network_args = [ "train_double_block_indices=1,4,6,17,18", "train_single_block_indices=0,3,7,9,12",]
network_dim = 2
network_alpha = 16
resolution = "512,512" => I can in 1024,1024 but for a character, don't necessarily have better results...
50 images x 16 = 800 repeats / batch 4 x 10 epochs = 2000 steps
experiment:
unet_lr = 0.0003
lr_scheduler = "constant_with_warmup"
lr_warmup_steps = 0.08
network_args = [ "train_double_block_indices=1,4,6,17,18", "train_single_block_indices=0,3,7,9,12",]
network_dim = 2
network_alpha = 16
resolution = "512,512" => I can in 1024,1024 but for a character, don't necessarily have better results...
50 images x 16 = 800 repeats / batch 4 x 10 epochs = 2000 steps