Greetings once again from purgatory
[Prism/Prima Project] Mako Sameshima PonyXL LoRA v3(actually more like v6)
After making many loras with it, it seemed to me like my juiced up config I made for making loras, what, 30-60 minutes faster? was extremely flawed. They MOSTLY work just fine, but it's especially shitty I guess when used with adequate or plentiful datasets, and sometimes had weird behavior with warped faces when they take up a large amount of the image. For a while I was even just setting and forgetting them for 2000 steps and choosing whichever step worked best. With this lora I instead used the old 0.0001/0.00005 lion config for 2000 steps and step 2000 seems to be really good. The most accurate it's ever been while also being at its most flexible.
The dataset was also bolstered a little, the end result being increased design accuracy as well as the capability to prompt her stomach tattoo (though it's sometimes a bit choppy) and other hairstyles. The 'streaked hair', 'multicolored hair', 'blunt bangs', 'long hair', 'medium hair', and 'skull hair ornament' tags can be used as desired to prompt her alternate hairstyles. There is little support for removing the head fins and replacing them with dog or cat ears however, and her stupid ahoge toggles aren't there either; I'll need more data for that (or prune data for balancing). For nudity it seems necessary to include 'petite' if you want the body to remain somewhat accurate, otherwise autismmix (or pony in general) is heavily skewed towards generic pinup fuckdoll bodytype when 'large breasts' is prompted. All examples in folders.
I'll probably redo the Kai lora with this config.
Haha wow that's a lot of text for a single lora
https://mega.nz/folder/dfBQDDBJ#3RLMrU3gZmO6uj167o-YZg/file/xCRUUBIC 