
7th attempt. Decent enough to release, but not very flexible. You'll have to play with the weights depending on the rest of your prompt.
I find training on old photos gives flatter results, to tried setting noise_offset
to 0.1. I think it gives more pop to the images, but sometimes they get too dark.
It overtrained on smiling pics so you may need to place smiling
in negative prompt.
描述:
PICS: 86
BLIP: 10 - 75
num_repeats
: 15
num_epochs
: 8
min_snr_gamma
: 5
noise_offset
: 0.1
network_dim
: 64
network_alpha
: 16
steps: 12% 217/1736 [04:53<34:15, 1.35s/it, loss=0.105]
steps: 25% 434/1736 [09:59<29:57, 1.38s/it, loss=0.101]
steps: 38% 651/1736 [15:03<25:06, 1.39s/it, loss=0.103]
steps: 50% 868/1736 [20:08<20:08, 1.39s/it, loss=0.1]
steps: 62% 1085/1736 [25:13<15:08, 1.39s/it, loss=0.105]
steps: 75% 1302/1736 [30:18<10:06, 1.40s/it, loss=0.0964]
steps: 88% 1519/1736 [35:23<05:03, 1.40s/it, loss=0.0991]
steps: 100% 1736/1736 [40:27<00:00, 1.40s/it, loss=0.0949]
训练词语:
名称: romySchneider.safetensors
大小 (KB): 73849
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success