Styles are hardly my specialty, but I think I’ve done enough I can make a supplemental article to my older training guide with some basic principles. If someone has more suggestions feel free to add them in the comments. Note my limited experience here is mainly with Illustrious, so not everything (particularly quantity) will line up with Flux or more modern models.
Start by deciding your scope
Start by deciding what, exactly, you want your style LoRA to include. Even within the same artist or work, a style can vary considerably. For example, an artist may have drawn very differently in the early 1990s and late 2000s, while different seasons (or episodes) of animation may have some clear subtypes (I've seen One Piece LoRAs done per animation director since the art style varies so much between episodes), and you may have to decide between making the style of a particular author, or a style of a work they contributed to (and maybe even led) but weren’t the sole hand of. Often times you’ll be forced to make a more composite style purely by virtue of how many pictures you have for training data (e.g., you can’t go narrower because you wouldn’t have enough data), but that’s OK as long as you’re cognizant of it.
How many images for a style and the two “types” of style?
Styles absolutely require many more images than most other types of LoRA. I think styles, and the images needed to make one, will fall into one of two categories: Narrow styles, only meant to make a limited "framing" of images, and general purpose styles, that should hopefully be able to do virtually anything.
Narrow styles seem to be doable with relatively few images, perhaps 30-35+ with flip augmentation on, but more is certainly better. The training data for these will generally have a consistent camera and/or subject type such as character art on a plain background, or portraits with relatively consistent facing. Most of these I’ve seen are for video games such as the promotional character art or as characters appear on the UI, though it’s not necessarily exclusive to them. Including unique on style but off subject pictures doesn’t seem to be as big a boost as on style and on subject pictures, but I think for these you might as well just throw in every image you have of a style. Narrow styles, especially ones made with fewer images, seem to be picky about what checkpoints and what character LoRAs they work with, sometimes having minimal effect when used on a character with high inherent style without lowering the weight.
General purpose styles are styles that should be usable for any arbitrary image within that style. While they want mass of images, seemingly somewhere over 100, they also need variety. I’ve found a style trained with few unmasked and non-monstrous humans in the dataset tended to struggle with human faces despite data quantity, and a narrow style with over 100 images still make a poor general purpose style if that dataset is narrow.
I don't really know where the diminishing return point for styles is, especially when lowered repeats with total number of images are accounted for, but I suspect it's 200 something. Since most narrow styles will be narrow because of a lack of quantity, you might as well throw everything you've got at training for those.
Avoid over-representing (the main) characters
A problem I often see styles LoRAs run into is that they overuse the main characters of a work in their dataset. It’s immediately obvious when you start seeing a style turn every male character into the hero and every female character into the heroine. Beyond just volume of data, this seems to occur when overly common tags are used for features that primarily show up on these characters (e.g., the main hero is most of the “short hair” data), so you minimize the issue by using more specific tags when possible and including non-main character examples of these tags. I think giving over-represented characters a proper tag to call them (even if it’s not enough for a full LoRA) and consistent attribute tags can reduce this, or at least allow the character name as a negative. Don’t be afraid to include (non-low detail) crowd shots or bystanders who exist for a single scene ("people" is a great tag for when you've got normal detail subjects but also low detail background people).
You don’t need or want a person in every picture
Tying into the above, it’s actually perfectly fine, and even preferred to have pictures in your dataset that don’t have people in them. Standalone backgrounds, or focus on objects, vehicles, animals (etc.) all make for good data. Just be sure to give some appropriate tags to these (no humans, [subject] focus, indoors, outdoors, object focus, vehicle focus, etc.) since auto-label tends to be terrible for no humans.
At the very least, make sure tags aren’t egregiously wrong
Styles generally don’t need the same degree of consistency of tagging that other LoRA types do (unless you’re trying to make certain things callable beyond the style), but auto-label is still horribly unreliable. At a minimum, I’d at least make sure the number of people and subjects (e.g., 1girl, 1boy, solo, no humans.) and make sure there’s not something totally crazy (auto-label loves to throw contradictory tags on together, especially for larger groups). Also take a look at your most common non-trigger tags and see if anything is oddly common.
Of course, having a good enough knowledge of booru tagging you can accurately tag things in a reasonable time will certainly still help.
Consistently tag things you’ll want the generator to avoid doing unsolicited (but can’t avoid high representation of)
This is, to some degree, a consideration for any LoRA, though styles have a few unique considerations beyond it. Most of these are fairly standard, like “simple background” or “signature”, or something you always need to do in the medium like “blank speech bubble” when dealing with comics (blank out the text). Style specific things that comes to mind are common clothing and “figure” tags. It’s plausible you might run into some kind of uniform being common on humans, which would should be a matter of tagging it as some kind of uniform instead of common tags. Figures are a bit more varied, but includes stuff like children, who are overrepresented in many works but can skew every human into looking like a child (and for Civitai will get a LoRA the engagement killer checkbox) if not accounted for (tag “child”), or “weird” body types like fat characters being spherical (tag “fat”).
Avoid any but the most basic cleaning or upscaling
Like all LoRAs, style LoRAs learn the style of your training data, and that’s going to include artifacts from (non-next nearest) upscaling or advanced cleaning (e.g., “clean” filters to remove watermarks). Stick to “basic” things like cropping and removing things from plain colored BGs, and let the rest fall where it does, accepting it as part of the style. With post-SD 1.5 models lot of things you’d clean can be minimized with tag selection.
Include non-SFW content if it exists in your style
As far as I understand it, the idea that a LoRA needs non-SFW training data to support it is largely a holdover from the SD1.5 days where that model had a hard time understanding clothing wasn’t part of a character that wasn’t trained with multiple outfits or none at all in the training data. This is not a problem in newer models (SDXL onward), but it’s still nice to include official examples of such content in your training data when it exists. This isn’t so much to make it possible with your LoRA, but to make it accurate to how it’s done by that artist/house. That doesn’t mean this content should be over represented however (e.g., if only one woman is seen without her clothes, don’t put that many images of her in or every woman made with the resulting model will have her exact body shape)
Note: If the subject of your LoRA falls into that category of artist that does sometimes draw this type of content but is terrible at it, feel free to skip this data and let the base model’s understanding fill in.
Be aware Civitai’s trainer changes default settings based on image quantity
Short one, but the on-site trainer will, by default, lower the repeat count if you have a lot of (large?) images. I don’t know the ideal settings for training in general, but I think it could be worth spending the extra blue buzz to up this count.
Tag in blocks
Tagging data is something I find works best in bursts, and the ability to save your work with the on-site trainer is effectively limited to downloading images+tags, plus general instability concerns. For this reason I recommend not trying to tag images (even if you're just correcting auto-label) all at once. Instead separate them into blocks, Your block size can vary, even within a single project, based on your tagging "stamina" and other concerns. I like ~30-50 myself, but you could be lower, or could be higher. Once you have blocks, upload one, tag it, and download it. Repeat till you've got all blocks tagged, then upload them all and train.
(This is actually applicable to any large LoRA, but styles are the most likely to run into it. For characters with lots of costumes your blocks will be by costume for auto-label concerns.)
Lower noise when training pixel art styles?
Something QuattroBajina2 suggested. Seems to work.
Make a good initial gallery
While any type of LoRA wants good samples for best first impressions, styles have a unique need for more variety in their samples (where characters could deal with just front pic+back pic+alternate costume). Try to show off a variety of things your style can do: Different types of character (male/female, tags only without a name/those from loras, unusual bodies like muscular, aliens), objects, backgrounds, non-humans, and more. If you’ve got character and pose/concept LoRAs already, they make great options for samples since that will also give some attention to your older works. Also, I know seeing women without their clothes in a style is half the fun of making a LoRA, but don't forget to include at least one PG image in the gallery or it won't be visible to lots of users.
Updates
Feb 18th: Went to make a style and remembered I forgot to include the tag in blocks section.
Feb 20th: Added note on gallery needing a PG image.
April 24th: Various minor spelling/grammar fixes, made a few things clearer.