I'm really liking '(by asanagi, by shinomu \(cinomoon\), by asura \(asurauser\):0.8)'
Examples with various characters, many using NAI loras. From what I can gather, it combines asura and asanagi's shading, largely replaces asura's face composition with a mix of asanagi's and cinomoon's, and adds cinomoon's clean hard lines
https://files.catbox.moe/ft6a2u.jpg>>66011958I use pretty much the same settings as I do for NAI loras, just with all the vpred and zsnr options on, sometimes I bake them a bit longer but that has a limited effect for making them better. I still haven't tried boosting my learning rates; sometimes using a NAI lora for a character works better with certain details (hair) while other details (clothes) are be dropped entirely when using a NAI lora on EF. In the case of concepts, I switch to locon and by default go for 12 epochs (I usually do 6 for characters) and keep the best one. Style loras are the most hit-or-miss I'd say, with some having a better effect depending on which artist in HLL is used; sometimes the style lora just adds noise (not literal noises by pointless detail or might even mess stuff up like the hands)