Question about training #2

by ymzlygw - opened

Hi, thanks for your good model, the result is surprising!

My question is about :

  1. Training way
    It looks you train it under dreambooth. But as I think your dataset is very large, it has potion, shield, sword and so on.
    How you set the instance prompt and class prompt? And did you train with a prior-preserving loss?

sorry, somehow really missed this question. in case this still helps anyone:

I used Dreambooth (JoePenna repo). Dataset was 118 images. instance token is just "fantassified", class is "icon". I did use prior-preservation but I don't remember anymore how many reg images were used. probably 1000.

Amazing work! Does 118 images were all potion?

Sign up or log in to comment