I have been using the settings you posted here, but failing the training to make a style LoRA, and am hoping you can help me out.

I used 10 images of an artist that did fantasy oil painting for D&D and similar games. They were all portraits of females, 10 in total. I also tried with a more diverse training set.

The results were a failure. There is too much similarity in the faces and scenes. Yours can be applied diversely, whereas mine was stuck on the faces in the training set. In addition, the results were unsatisfying, and I could achieve similar with regular text prompting without any embeddings. Both the LoRA's I trained and regular prompting fail to reproduce the artist's distinct style.

I should mention that I tested all of the even numbered epochs from 2-20, and in :0.X increments of .1.

So, now after trying all that I have a few questions.

1. Do you use regularization images? If so, what class keyword would you recommend? I'm doing a painting style, so I am experimenting with keywords like 'art style' and 'painting' to generate the regulation image sets.

2. To which of the 3 available options do you apply the learning rate of 1e3? (Learning rate, text encoder learning rate or Unet learning rate?)

3. Do training set filenames/tags matter? If so, any tips for how to write them effectively?

Here is the config file to confirm if I'm using the correct settings or not. https://pastebin.com/BrS6NzeT

Thanks for all your good work.

FOLLOW UP:

I was mildly successful training a LoRA 1mb. However, the facial similarity is still a problem.

Useful as a negative prompt?