(Pictured: Frontiers in Glory Days version of Asticassia Uniform in the style of Gundam X.)
Like my earlier guide for styles, I'm not a clothing specialist, but I think I've done enough (and this set of LoRAs does provide half my yellow buzz earnings) to write down some basics to supplement my general training guide with points special to clothing LoRAs, and maybe prompt someone better at this than me to chime in with their own advice. Note my experience here is mainly with Illustrious, so not everything will line up with Flux or more modern models.
Have multiple wearers with different physiques
LoRAs are ultimately about teaching patterns, and your LoRA will (like it or not) learn all the patterns its given: If every picture you feed the trainer has the exact same waistline (etc.), that waistline (etc.) will become part of the clothing and be very hard to correct in generation. The main way to control what patterns are learned is selecting data so the only real patterns across all the data are things you’re trying to train, and as much else as possible either varies, is so universal as to not matter (e.g., humans are human shaped, “white background” is white), and/or be isolated into a known tag (if all the data has dark skin, but you applied the “dark-skinned male” and “dark-skinned female” tag it will just push gens with “dark skin” towards the exact tones you showed it and those without that tag won’t be effected). Luckily, since you need double digit images to really teach a concept, this is harder to mess up than I make it sound if you’re paying attention to it (it’s also basic LoRA fundamentals).
The different physique part is really only going to not be the case across multiple wearers when you’re dealing with video games, some (cheap) animation, and (poor) artists where the different characters are pure head/skin-tone swaps (maybe with some horizontal+vertical stretch).
A quick note is that mannequins, and unworn clothes do work for this.
Have multiple angles
While character LoRAs can make do with no having back pics and come out OK (though any back shots will be totally made up and thus suboptimal) since the base model knows so much about humans, its not at all advised for clothing. If you’re trying to teach something that’s only ever going to be seen when viewed from the front (like a particular necktie) this isn’t as serious.
Don’t worry about overlap in angle tags, especially if they are very specific for that tag by definition (“straight-on” is exactly 12 o clock, “profile” is exactly 3 o clock). Pushing a model towards “from side” being exactly 2:30 is not ideal, but entirely workable.
Have a tag for every piece that can be removed separately, and what changes within the dataset
You do not need to tag every detail of an outfit, only a tag for each piece (e.g., hat, shirt, pants), and things that aren’t consistent within the dataset (e.g., male characters wear pants while most female characters wear a skirt but the uniform is otherwise unchanged, or one character wears a novelty tie while everyone else’s is plain black). Picking which tags is generally straightfoward (as straightforward as tagging normally is): Just pick a relatively narrow tag where possible (e.g., “ankle boots” over “boots”) for max compatibility and, if you’re really stuck, pick a tag based on interaction with other tags, as it will inherit some compatibility with other tags (verbs, nouns, and some adjectives) from what you base it on, even if it's not strictly the right name.
When possible, take off heads like Henry VIII or a French revolutionary.
One great way to not have heads and hair contaminate a clothing LoRA is to just get rid of them with a little cropping. Clothing store pictures often come framed like this. Most of the time you just need to tag these images with the applicable of “lower body” or “head out of frame” (or, occasionally, eyes out of frame), though sometimes you’ll have an extra focus tag available as well (e.g., footwear focus).
(This is another article I had mostly finished in a text document and just needed to put the finishing touches on and publish then remembered I had and finished.)