I thought I was done with this. There have been 5 main improvements/additions. Everything from total torso remains the same but I was unhappy with the eyes I was getting, freckles were showing up like painted on dots, pubic hair was unreliable and fake looking most times, and faces were suffering likely due to all the cropped-out chins from the other training images. Tattoos were simply something I have considered in the past.
Best overall lora weight is 0.8 in my testing and you will still want to start with a lower trigger weight on tags like "face freckles" and "hairy pussy". Hairy pussy is capable of going from stubble to Wookie.
Body tattoos will range from sleeve tattoos only, then chest, then stomach. (usually)
I included the eye color tags simply to show which were baked in but with an added bonus, "2tn eyes" If you know someone with hazel eyes or they say their eyes change color, these are those eyes. It's tricky to get a specific combination but even used by itself can result in a much more dynamic eye color.
The fifth improvement is a more subtle one. There are no triggers for it. twenty-six images of high resolution (1080 and higher) portraits were added with full BLIP captioning. Portraits of actresses and completely random people also.
The overall dataset image count without counting regularization images is now at 1325. I don't currently have any plans to add anything else to this project, but that's what I thought last time.
(For anyone wondering, the F in the name is simply my way of tracking how many times I tried different training methods, this was my 6th attempt at this version.
