>>37013723>Personally im thinking maybe 2500, 3500, 4500,after that it gets into unsure territory.I should probably do this with the anya embeds I have, I also had small dataset so that might also mean it's overtrained and some of the earlier steps are better?
Pic is comparison of the 3 anyas I have after all this mess. Top row is xformers & 4 tokens, middle row is xformers & 8 tokens that was set to train to 100k (oops) but I stopped it at 10k, bottom row is no xformers & 8 tokens, with --precision full --no-half as command line arguments. All from the same dataset. The last one needs a bit of extra strengthening on the tag to really look like her. Personally I prefer the middle one.
