Settings for 400 instance images?

[kaylazy]

I'm training a model to a produce a specific type/style of photography collages. The photos do contain people as subjects but I don't care about training on the subjects, just the style of images. It's kinda like a before and after photo thing.

I want to use 400 images to add more variation to the model, I don't want to settle for any less.

I was wondering what unet and text encoder learning rate to use and how many steps for each to use when training on 400 instance images?

[TheLastBen]

Set the text encoder to 350 and the unet to 1200, then test the model, use a good prompt.

[TheLastBen]

the learning rate of 1e-5 requires very few steps, you will get even better results with the V2-512

[kaylazy]

my results arent as great as they were with smaller datasets. It's not even hitting close to the style, whereas before they were always close enough. I suspect I might need 3000 steps or so since thats what I used on the smaller sets that had better results. Should I tick Style_Training if I am going to train the model further?

[TheLastBen]

My recommendation is to not go above 20 images for a style, but since you said "I don't want to settle for any less", then you're gonna have to try various things to get it to work

[gilroff]

selecting the best 20 and adding the rest to concept images and using a high number of concept image training steps. will this work?

[TheLastBen]

yes, is should work