>>2852170There are two schools of thought and it depends on how close you want it to the original/how much control you want on it.
In my methodology, implicit tags (things that describe the character all the time, not just sometimes) should be removed after auto-tagging. Using the tags in your gens (even without them in the training data) will help guide outputs more as well if the AI didn't "learn" something explicitly and is barely grasping certain things.
Even with a furry subject (a pokemon) in a LoRA I've trained about 100x out of autism, I add tags that describe ears/tail to direct the output.
