Diffusion Single File
comfyui

Tagging for lora training or finetuning?

#105
by CappyAdams - opened

I know how to train loras in general but I still wanna ask this:
Should I use natural language or danbooru tags in my datasets? or both?

I think the official way to do it is mixing, while keeping the format as close to the reference as possible which is [quality/meta/year/safety tags] [1girl/1boy/1other etc] [character] [series] [artist] [general tags/natlang]

author said they use dropout, so should have that enabled. Other than that, keep to the order, and then do a mix of tags, tags + natlang, natlang, natlang + tags, is what i've found from my own training runs, and been told to do from others.

Gotcha! Thank you so much for your reply!!

CappyAdams changed discussion status to closed

Sign up or log in to comment