Ideal Training Settings

[ClaytonFarr]

In lieu of a project wiki, perhaps we can collect and share here what settings are working best for folks.

I'm still trying to get my hands around things but these are ones I've seen shared with good results thus far (I'm sure there are others). @cloneofsimo, @nitrosocke, if I've confused any of the settings for what you've done, let me know.

Training Goal	Person (Unknown)	Person (Known/Celeb)	Style
By	xsteenbrugge	nitrosocke	cloneofsimo
Link	link	link	link
Base Model	SD 2.1 base		SD2.1 base
Resolution		768	512
Images	20	5-10	10
Captioned	Yes	No	No
Steps	400 / img	100 / img	300 / img
Unet Learning Rate	1-e4	1-e4	1-e4
Text Encoder Learning Rate
Learning Rate Schedule		constant	constant
Training Batch Size		4
Gradient Accumulation Steps		1	1
Mixed Precision		fp16	fp16
Use 8bit Adam		Yes
Prior Preservation			false

What training settings are working best for you?

[ClaytonFarr]

[brian6091]

Great idea! I suggest adding the base model (SDv1-4, 1-5, 2.1) since there seem to be some differences in how they fine-tune.

[ClaytonFarr]

Great point. When I’m back at computer I’ll look up to see if current examples mentioned model and add those that do.

[cloneofsimo]

This is great! I think when there is enough info on this, we can add this table to the readme as well.