>>53818
The thresholds for images needed for a style is a subject I'm not entirely sure on myself, with dataset quantity thresholds being poorly documented (even numbers needed for a cartoon character were all over the place when I started. Now that I've done enough character models with low data I can safely say for those). As a sometimes style maker, frequent style user, and frequent guide reader I can say the following with confidence
>base models vary in dataset quantity needed for training (Flux apparently needs a lot less than Illustrious)
>image size and quality matters (and don't upscale, you'll teach upscale artifacts)
>at least for certain styles (pixel art likes low noise) your training settings plays a big part
>training settings are easily the least understood part of the process even for experienced trainers (most people leave everything on default or only changes a few settings to their default preference)
>a model that only wants to replicate one specific type/framing of image (e.g., video game portraits or promo character art on a blank background) needs a lot less than a style able to do "anything", but it won't be any good outside of that specific thing (only ~30+ of that style with flip augmentation is good enough?)
>~50 works fine for a general purpose style when it works, but some things are just frustratingly completely undoable by the resulting model
<for example my LucasArts cutscene style model at 58 really sucks at adult women because there's only a small number of those in the dataset, and the default anime bias in the base model will overshine it on them
>the 398 images for the Jimmy Neutron style and nearly 1k images I used for my Gundam X style was almost certainly overkill
>do not try to handtag 1k images in a single go, you will burn yourself out
>making a bunch of smaller LoRAs for characters with screenshots, then combining them and some generalist data, is a much healthier way to get a big handtagged dataset
>if a model is going to make characters (instead of just objects), it needs subject variety to avoid turning every character into the most seen characters (there was a guy who made series styles with mass automated screenshots and autolabel, which worked but they a habit of making everyone's shape and clothes match the main character due to screentime bias)
I'm thinking ~100 is
probably enough for a general style. I can't tell you how big an impact hand-tagging (with or without suggestions from auto-tagger) vs. auto-tag only vs. cleaned up auto-tags (e.g., correcting 1boy vs. 1girl, replacing hand on hips with the actual booru tag) is in practice for styles, though consistent (a "blouse" is a "blouse" every time it appears, never a "shirt") is essential to getting the most out of low datasets for characters.
For a
character or something like a pose instead of a style, the answer is a lot easier and more definite: Results get way stiffer and less accurate with under 20 images and improve as you get to 30. After 30 images there seems be enough diminishing returns that excluding the lower quality images is as big an improvement as including more images. Extra costumes need extra images, but less than a whole new character. ~10 properly tagged seems to do it, but it seems to vary by outfit complexity and if you don't have enough the worst that will happen is they won't work or gens won't be a perfect match (e.g., if you only have 3 pics of a character in a particular bikini and add them to 30 of the character in normal clothes, at worst you'll just get bikinis with a vague resemblance to that official one when prompting for bikini).