Although I previously stated that I wouldn't be releasing any updates, I went ahead and did so anyway. I’ve added some new datasets and tweaked the training methodology; however, I still haven't been able to achieve a 100% response rate. Nevertheless, compared to the previous version, the theoretical success rate has—in principle—risen from 15% to 60%. I have moved my previous post over to the "Articles" section; if you are looking for a way to achieve consistent results, please refer to that article.
https://civarchive.com/articles/29350/how-to-create-a-face-behind-a-mask-or-inside-a-bubble-in-a1111


Basic Prompts: thought bubble,circular inset,face in bubble,expression inside bubble,close-up,face only,portrait,looking at viewer,1girl,black latex bodysuit,full face mask,hidden face,featureless mask,cat ears,fake animal ears,shiny clothes,skin tight,contrapposto,Chastity lock,bulge,chastity lock,[color] inner ear,[color] pattern,bodysuit,


Ball Prompts: ball gag over mask,gagged,ball gag in mouth,mouth open,tearing up,


Thank you again for your patience. Although stability has improved slightly, you can see that the model still occasionally struggles to decide which face to render; therefore, I still recommend using regional segmentation for your drawings.