Type | |
Stats | 1,999 128 3 |
Reviews | (244) |
Published | Mar 11, 2023 |
Base Model | |
Training | Steps: 7,300 Epochs: 20 |
Hash | AutoV2 012DAAAB2F |
My pixiv link:
https://www.pixiv.net/users/76094070
This is the style of a deceased famous artist whose name you may not have heard, but you certainly know a galgame called "Shuusaku".
I used his two books of paintings as Lora training material.
And this time I used two different parameters for training. Two models were created.
Here I call them 01 and 02
Here are their training parameters, by the images created, I prefer the style of 02, 01 feels more retro to me.
01:
batch_size = 2
network_dim = 128
network_alpha = 128
use_lion = 1
no noise offset
no loo
02:
batch_size = 1
network_dim = 16
network_alpha = 8
noise_offset = 0.1
enable_locon_train = 1
use_lion = 1
Here are some training diagrams, my Lora name is HH02EP20, "HH" is the abbreviation of this master's name, "02" is the second model, "EP20 " is Training Epochs: 20,If you would like additional training epochs, please leave a comment below.
01
02
Please don't use it for commercial purposes, I just miss the work of the master, and the acg culture of that era.