@forcekin44 Thanks. Worth noting that, in my experience (which absolutely does not cover realistic stuff), as long as you clear the 256/256 minimum without substanial bits of other characters and aren't jaypegged, you can indeed use "lowres" data. The "depth" that the trainer sees images as is closer to a vector than a human sees it (there's some tools to convert an image to an approximation of one out there). At most you'll lose fidelity of hands (SD1.5 just stuffing its base model with such is why it was so bad there) but those can be fixed with regional stuff if really a problem and will be a problem anyways. My pre-spark Chandra LoRA actually uses a lot of low res stuff to get to its quantity, and a lot of my stuff doesn't use much better (I love making LoRAs for old obscure stuff, including several 480p series and I've even gotten good results off teeny tiny sprites). Main issue is the inconsistencies in the art and not having a good shot of the regulator so it gets read as a pouch.

@forcekin44 For characters, I've found ~20 to be the threshold that I wouldn't bother if I couldn't exceed for characters. The few times I did go lower the LoRA preformed noticeably worse. I think ~30 is what I'd prefer for a single outfit character, past which one just gets diminishing returns and strong inherent style (though this also depends on the nature of what images are in the data). I don't really have data for outfits/poses/etc. but I think the minimum for them is about the same from what I've seen.

Styles are a different story. Based on what I've seen (both with my work on styles, the "inherent style" that will show up on characters trained on a single work, and the work of other people on stuff with such finite data I can make a good guess on how many images they used even without being explicitly told) they can be divided into two types "narrow" and "general purpose". "Narrow" styles are things that are only expected to make one kind of image (e.g., my Fallout Talking Head Style LoRA is only expected to make character portraits at a relatively limited viewpoint, while my Dawn of Souls Style LoRA is only expected to make character art on a blank background). These seem to only need 30-40 images with flip augmentation to work well, but don't hold up too well to making anything but their focus. General purpose styles, ones that can do arbitrary compositions and subjects, seem to need at least 100 to work well (My Carmen Sandiego style uses 104 and does well most of the time, though for some gens I think it could use more. My Gundam X style used just short of 1k handtagged images and I never found it to not work. My Jimmy Neutron style at 398 feels like it's held back by how well the checkpoint handles non-realistic 3D than the LoRA itself). Beyond quantity there's also the question of what exactly is put in (my Christopher Rush style LoRA has a relatively small number of "clear" humanoid faces in its 102 pics and I find such heads its biggest shortcoming) which I'm not entirely sure on yet.

As for clip skip: I think almost everything I've done has been on the default "1" because it was a mix of styles and I only bothered to change it for a few things that were all anime screenshots. I haven't noticed or been told about any impact on it. Only ever made 2 SD1.5 LoRAs (once for a request for a western comic character, once as a joke to post Ultra Lord every week in a row for set period) and I think both were left at default

@forcekin44 It did indeed take a long time and I will never do that many manually tagged again (especially as it's not even remotely needed). I did try to work around the idea with my Jimmy Neutron style LoRA by doing lots of characters on their own then combining them, but it's still only something I'd do as a fan of the series trying to pump my LoRA count. I don't think everything needs to be super well tagged for a style (my Carmen Sandiego and Code Lyoko styles are mostly auto-tagged with basic corrections) unless you want some items or characters callable.

Only real advice I can give on tagging generally is (like with most things) to "check the docs". In this case, the docs are the Danbooru wiki. Helps find related and more speciffic tags as well as learn exactly what each tag actually means (not to mention knowing what's an actual tag).

@forcekin44 Glad to hear it. I don't really have any requests though (the great thing about knowing how to make LoRAs is that you don't really need other people to make them short of not having access to the source material in good quality)