A unique merge of SteinXL (my Pony model merge series) and SteinIllustrious, with enough differences to compel me to make it into its own thing. This is the first time I've been able to perform a successful merge across the two architectures.
Recommend ADetailer for face compositions, and DDPM sampler for best quality.
Description
Launch edition. I realize it was called V3, and that's not strictly accurate, as it is the initial version.
FAQ
Comments (5)
Wow!! I wonder how to merge pony and ill model? And is there any problem with pony's aesthetic tags like "score_9, score_8, score_ 7, ...". THX
It's able to work with or without score tags, like Illustrious, but it has a bit of the flair that comes from using those tags; though, it does have a tendency to try to print them somewhere in the image as text.
The trick to get the merge to work at all was to use an exceptionally low merge percentage (0.3%) for the model not to become corrupted. As it is, this merge is an academic exploration, and I'm not entirely sure where I can go with it.
@AI_nstein any idea what the best way is for learning about what score tags do? Like what they mean?
@ShlongMcDong I'm not exactly an expert on it, but my understanding is that the archetypal Pony model lacked some specific training data or tagging data related to aesthetic score tags, which basically equates to a score (as seen in the prompts used with that model) denoting the overall aesthetic quality of an image. Kind of like how a lot of prompts start with "masterpiece", for instance, the idea is to nudge the AI into generating a more pleasing result, based on training data rated more aesthetically pleasing.
That, and Pony seems to profit from having underscores connecting parts of the prompt, like_this.
As for what these tags actually DO in a prompt, honestly, I think it's simpler just to have it down to trial and error; try using them, then try without them, and see what result you like better.
@AI_nstein
theoretically.
it could be worth looking into extensions that remove a text encoder from a pony model and then DARE merging the unet blocks with illustrious low in01-05 medium to high in05-08, 50/50 mid blocks, high out08-05, medium high out05-01, I would think the text encoder from illustrious will be transferred in the process into the new merge.
or it could be worth trying to find out how to separate the te from illustrious, merge unet pony with complete or only unet illustrious and then merge te from illustrious after merging pony and il unets?
could be how we go forward because pony loras work well in illustrious and noob models merged with pony.
could be as simple as 0.5 weighted sum or add 😂