>>61366955Nothing wrong with "Page 1/3". If the TOC table can be replicated among all 3 of them that would be even better. Just harder for you to manage, I guess.
>>61367673>>61368199My Suggestion, do one on Civitai if they give you enough points and then look at the Settings used to make it and replicate them on your Kohya Gui.
They already show you some of the settings when setting up your training in "Advanced Settings" and after it is done you can load up the lora trained there in your A1111 and check the little 'i' icon to display more settings.
Some settings I use that come up with good results most of the time:
Optimizer: AdamW
Optimizer extra arguments: "weight_decay=0.1"
LR scheduler: cosine_with_restarts
LR number of cycles: 3
LR warmup: 5%
Unet LR: 0.0002
Text LR: 0.00005
Network Dim: 64
Network Alpha: 16
Epochs: 12
Dataset balancing: Aim for 100 with repeats (so 1 repeat if you have 90-100 pics, 2 repeats if you have 45-50 pics)
It should end up with around 1200 steps +/- divided by batch size(which depends on the amount of VRAM your GPU has)
The rest depends on how good your dataset is.
>Prompt yoinked from CDB