basically remove all the mis-tags, as much as you can tell. you can keep the one-eyed or single eye tags, but I removed them to emphasize the result, and give a new tag cyclopscpt instead so the outcome will be more consistent. cyclops is one of the easier one to train, since it only involves one trait, one eye. I'm planning to retrain the spider one, since the outcome is not too consistent, will need to edit the caption files again...

think it like you're teaching the ai an idea, you'd want the imgs to be more focused on certain traits with different art styles, so the model will get the idea and won't be narrow minded.

for example, if you are training the concept "human", you'll try to gather a bunch of human imgs, with different races, age groups, body shapes, genders, etc., instead of only showing people from certain area and one age group. the more diverse the imgs are the better (they'll need to focus on certain idea).

of course we'll have limited time and resources, so do what you feel the best fit for you.