Used a few of my personal favorite generations to whip this up.
As a result, it is essentially a style lora that mixes some of the more popular semi-realistic models. I'm vanilla, so it is very good at whipping up basic "ultra hot" girls, rule 34, bodies, sexual acts, and expressions. Very good with multiple skin tones on both male and female subject as well.
Works best at around 0.8-0.9, 30-50 steps, Highly sensitive to CFG like most pony LoRAs,
3-3.5 CFG highly recommend, go no higher than 5.
Regular pony problems with weird faces/extra digits do occur, but easynegative/lazyneg type embeddings help a lot, this model is also very good at image to image variations for refining small mistakes it makes. Don't forget to make sure the trigger word is in the prompt somewhere.
Helps to have fairly specific prompts, I recommend using both my negative and positive prompts from the sample images in V3 as examples and editing them from there. Use booru tags and the innate pony character tags as required.
I made this to streamline my local generations, but I would love to hear feedback, V3 will likely be the final version. I may make a custom checkpoint at some point if there is interest, but for now....
Enjoy :)
Description
Added many more training images that I personally in-filled and edited to enhance the results.
Added many distinct tags, notably perspective and view tags to make it easier to frame the image. Too many to list here, but full training data is attached if you're curious.
Skin colors and regular pony errors have been minimized.
Use same prompts and weights. Sweet spot is between .75 and .9.
Clip Skip between 0.5-1 for best results on most local generators.