My first Lora model, so I expect I'll revisit occasionally to update maybe?
This Lora has been trained using the StoreBoughtGyozaMixv32.
It's just the model I tend to use for some random reason.
I didn't use any tagging or anything fancy, so I strongly suspect I'll probably upgrade this when I get time, or can be bothered to. Was just too lazy to do it the fancy way this time.
You will probably find sometimes it'll merge the rider legs with the ponygirls arms or legs. (this is common problem, not helped by me accidentally including a couple images with this problem in the training data...)
It tends to prefer kind of diagonal front on, but it can do back on randomly sometimes. It'll also sometimes have rider and ponygirl facing opposite directions which is I guess a thing.
I tried really hard to get it to make rider have clothes and ponygirl not have clothes. (but even still it'll sometimes make rider the skimpy one and ponygirl the clothed one for some reason...)
Well it's a start, you can play around, it's better than nothing... probably.
Tag it should've learnt to recognise, if I did it right?:
girlride
tags that probably go well with it? idk:
2girls, bdsm, riding, outdoors, [nsfw]
don't set lora strength to 1 it'll get frayed and stuff, won't look good.
Experiment yourself, but I recon between 0.6-0.8 probably best? Tbh I've just made it so idk try yourself.
Description
Increased training set size, fine tuned tagging and in general pruned training data so ponygirl legs were better quality, added gag to ponygirl in training set and made rider little more sadistic.
This lora version is focusing more on versatility and consistency over beauty.
Lora probs best around 0.75 wieghting
tags I recommend using:
always use: girlride
recommend additionally trying: ponygirl riding, pony play, horseback riding, boots (or any specific footwear),
makes worse: negative prompts generally, unfortunately especially with Lora's featuring more than 1 character, any negative prompts usually break things in my experience. I strongly recommend not using negative prompts.
known problems which will work on for next version:
ai tries to add whips to the scenes but doesn't quite understand how and what whips are entirely. cute attempt, will probably try to fix this in next version by separating the images with whips in and not in.
reins still needs some work in general, but it's a little better. Think this problem is just down to training data, will have to photoshop the training data with reins to improve this I think. Which is a bit of a pain, might do it though if I find myself with a nice quiet day.
Overall prettiness has taken a little hit I think. somewhere in the middle I had really nice aesthetic but since I added bitgags to all images using adetailer and increased number of training images, aesthetic changed subtely and not sure how I feel about it. I suspect splitting out the bitgags from training data will fix, the average training images resolution decreased a little due to a lot of 512 by 512 images being added to training data. Which might be related. Tempted to add a version 1.5 because I had an inbetween version which arguably some might prefer aesthetic compared to this. But idk won't bother probably. Since I plan to fix next version anyway.
I'd say in general it's improved in most aspects. I'm not on copium I hope, Have fun :)
also side note I trained everything using storeboughtgyoza model. So this checkpoint will probably provide best results.