nitrosocke
commented
1 year ago Hi there! I haven't done a lot of training continues, only on aborted models with low steps. The updates I'm releasing are always completely new trained as I'm regularly changing the dataset completely.
From a friend's testing I'd say training with a refined dataset for face closeup and a lower LR might give you the best results. But make sure your model is not overtrained already
I switched back to polynomial as well since that gave me the best results. Can't recommend constant anymore unless the little speed increase is important for you.
Thank you so much for how open you've been with the community: It's led to a lot of others having great success, and I hope others will be as open as you have been. Would you be willing to provide advice on continuing training? For instance, let's say I train a style with 112 images for 11200 steps, and it's done a great job on bodies, but face details are still pretty messed up. I can imagine a few approaches from this point:
1) Start completely over, but include a higher percentage close ups on faces next time around? 2) Start with the model I've already created, since it's good on bodies, but now train it with more faces in the image set? Maybe with a different learning rate? 3) Trust that I've trained the model to do bodies, and just train it on faces only for a more limited amount of time?
This is a specific question I have, but more generally, I wonder if you would be willing to share what you've learned works when updating a model you've already made.
The other question I can't seem to find an answer to is why you use constant but so many others use polynomial, and if that is because you're training styles versus their training of faces, or if it's just: people are doing whatever has worked for them. :)
Thanks!