Thanks! I've always pointed people at this as it's all I ever really did, but its slightly outdated now https://rentry.org/simplified-embed-training though the only real change is I think you might need to preprocess your images to the right size first (or leave them uncropped but it'll need more vram). The UI is now a little different but all the options are the same (on the "Train" tab)

@smythmark87855 the key is consistency, but only on the thing you're training for. So for example this embed, almost every image was a red ball gag with the loops at the side, for the sex positions it's using a consistent composition or else you end up with a mess. Characters you can be more varied as long as the character part is the same

Sorry wrong reply and can't change it apparently!

Thanks, yeah using the two together will give good results I think. Good luck with it!

Good spot! Despite the one here it doesn't work almost all of the time, so either I'll refine it or maybe release all the failures somewhere for people to tinker with

Please refer to the description on this, but also yeah hands and especially feet are hard in SD because most artists also struggle with them and/or also hide them. There is a my_hands hypernetwork I use sometimes but I don't know if it really improves hands or just changes the overall look a little, and there's a bad_hands embed too though again I don't know where to get it from :D Feet would be difficult, I can only see it working if you want feet that are in the same position every time. An embed doesn't really improve the model, it just gives it a very specific part of the prompt

I actually had a go at this exact thing a while ago but SD is so bad at feet, especially upside down and inverted feet, that it didn't really work so I didn't release it. It's here if you want to try it out though https://drive.google.com/file/d/12a-nRNEp9N74oax1b_XpWAPsOlJoNZYY/view?usp=share_link

Heh, thanks!