>>59521322I created 10 loras about 3 months ago so idk if training methods have changed since then.
For image training set make sure backgrounds aren’t repeated too many times, think like an OLED display with burn in. If the background is always one color I found it shows up in the final output even when tagged. So variation helps
Finding images from different angles and poses is important. I find most vtuber loras have issues with this, as most of the training set is live2d standing upright, so trying to prompt them to do other poses is difficult.
I used automatic danbooru tagging then pruned tags
For pruning tags, if your character always has a certain appearance for example blue hair, red eyes, long hair. Prune those tags so when you use the Lora you don’t have to retype all those tags for their basic appearance. For clothing I typically leave those tags alone, to make it easier to change their outfit in final output
If your character has red hair in one image when her normal hair color is blue, leave it tagged so it doesn’t show up unprompted, same goes for outfits, accessories etc.
make an XYZ plot grid with the different epochs to make sure there is no overfitting after training
Here’s a rentry with my Loras, shylily one was the last one I did, pretty sure you can check the metadata in the lora for training info number of epochs etc along with tags used
https://rentry.org/t9ekk 