Saw another lora training guide posted last thread (or was it the one before?) and checked it out to see what's different from the previous guide -
https://rentry.org/59xed3I used a leftover Kanata dataset I had from an old embed I tried to make that didn't really work out and the general idea is fine but it's very attached to the idea of giving her a specific pose in basically all prompts I try. Weird since usually when something is too strong it's the outfit, never seen this sort of problem before.
It's a bit hard to believe but picrel is actually a different seed for every single image and the prompt is just masterpiece, best quality, 1girl, *, standing, cowboy shot
Clearly my dataset/tags aren't good, which I knew already. And the guide emphasizes very strongly the importance of good tagging. So, if whoever wrote it is reading this, I recommend when you get around to uploading a lora made with the guide as a sample that you also upload the dataset + tags too so people who are dumb (like me) can learn how to tag things properly
also as a side note, the guide says to not use the same venv as the webui does, but I ignored this step and everything still seems to run fine. Do at your own risk, though
