hey there,
so if you read my training guide and visit some Youtube videos, you still don't really know what is
"Over- and Under- Training"
to get an idea, it is important to generate samples during training.
the samples should start with a prompt that you should choose.
may "a young woman look at me"
so in the begining the training generates samples of a woman that look at you.
from about 30% of the total training, there should slowly be some similarity to the training images.
of about 70%, there should be some samples that are really very similar.
when at the end of training, the object of interest looks cut out, then overtraining begins.
the goal of a good model is to determine the learning state before overtraining begins, so you can play with it in a large range
CU and have fun
start -> 30% -> 70% -> 95%
you also see it wile the learning, say the loss is arround 0.1 it should be stay around 0.1 is the loss dropping down 0.9 -> 0.8 -> 0.7 its over fitting!
in addition for SDXL training, the image get more and more dark points ore lines so its over trained ...
more helpful videos
https://github.com/XavierXiao/Dreambooth-Stable-Diffusion/issues/124
very good actual video !