Quoted By:
So I tried to make an iA3 version of this Shiori lora I've been working on. I changed nothing about my dataset between the two, and just used the default settings from the prodiagy rentry. I'm assuming I did something wrong because it doesn't really look great. I did train it up to 2400 steps, but at that point it started to look funky.
I'm not sure if I need to do something to my dataset. I saw that you can go captionless and just put your trigger word in the folder name, but I think I'm not understanding what that really means. How does that work if I have multiple concept folders with different repeat values?
v3 in picrel was a regular lora trained with prodigy, and the v4s are the iA3.
