Your example images contain a lot of redundent tagging (for example "freckles" is listed multiple times), and a few misspelled ones ("eyeware"). I'd suggest double checking them when generating as it makes generation do some funny things. May also have been an issue in training data.

For user ease of use I'd recommend including the tags in the description, separated into one block for physical characteristics (character's trigger word, plus eye color, skin, hair style etc.) and a second (/third and so on if character has multiple outfits trained) block with the outfit listed top to bottom (its best to not specify "pants" and "shoes" if you're trying to generate a picture that only shows the head or upper body). This makes it easier to copy/paste into the generator than just putting them all into a single blob on the trigger word field