Hello there, my dear friends and readers. I hope you're doing at the very least fine ;)
I've started doing Loras - more or less successful ones at least - barely a few months ago. It's natural that I am still learning these things. Honestly, I don't like bothering other people that much, but I also hate walking around blindly. Trial and error is a natural way of things, but everything has its limits. Even though I mostly figured the amount of steps it takes to train a Lora with a given number of images, I still can't really figure out what I should tag in the image and what not. I should also say that I rely on autotagging(30 tags with 0.4 setting) and then add/delete tags manually as I see fit. Can you folks give me a few pointers? Maybe you got yourself in simialr situations, how have you dealt with them?
Just a few examples.
In my Lisker-type zoanoid Lora I had just a few dataset images that were 3d and they all came from the same source. I tagged them accordingly, and now... whenever I put a "3d" tag, it tries to recreate that same 3d model no matter what I do and what 3d-styled Loras I use. Yep, it could be said that overtrained model is part of the problem, and still I think that I should have avoided using tags such as "realistic, 3d, photo, artwork", etc during training.
Then goes my Crysknife Lora which is rather more successful and yet it struggles with certain things. For example, close up portraits are messed up most of the time just because I tagged several images as "close up" during trainig. To be specific, those were close up shots of a hand holding said knife. Now, most of the time, when I try to make a close up portrait of someone holding a crysknife, it only works 1/10 times - it mostly tries to make that hand to stick out from behind the frame, bend it at a strange angle, just puts a knife(rather oversized knife) somewhere in the image, or even tries to burry that person in the sand up to their waist!
Then there's my Lora of a Rust dragon from HoM&M III. There was just one image with trees, maybe two images of disembdoied head portrait, and one image with it breathing fire. I tagged them all accordingly(except for disembodied head) and now... well. Every time I mention a tree or a forest it tries to recreate that image style, with pencil art and withered trees. Most of the time the dragon breath comes with the same style that one image had. Most of the close up portraits come out with a disembodied head, even if I put "disembodied head" in the negative tags.
Then there's my Shao Kahn Lora. Many images of Mortal Kombat II Shao Kahn come with "spiked knuckles", tagged these as such. It resulted in 9/10 generations of Knuckles character from Sonic. Not much of a problem really, just don't use that tag and everything is fine, but I'm kind of a perfectionist and Captain Control Freak, so it irritates me to no end.
There were also a lot of things I decided not to tag and now their appear anyway, as a Lora probably recognises them as a natural part of the concept. Such as ropes or rope bridge from my William Corvinus Lora - they tend to appear everywhere, in a cave, in a forest, etc. And nothing can be done about it, even putting "rope" in the negative tags doesn't help, because the Lora probably doesn't even understands that those things are ropes.
So, well. You're effed if you don't tag, effed if you do. Maybe you can give me a few suggestions about what to tag in the image? Or perhaps share your own experiences with this stuff and how have you dealt with it?