The ZImage version alters the entire image far too much to make it usable for me. I would be grateful if you could change that.

- Using Z-Image-Turbo v1.1. -

Unfortunately I can confirm what someone else said, that the personal characteristics bleed through, though I didn't find it to be detrimental to my current creations but it does limit the characteristics I can acquire so, for instance, any character LoRA I used would be about 50 percent obliterated by the use of this LoRA with a strength of 0.8 and, like someone else said, lower values are pretty much not useful.

There's a thing called "masking" for training, which I don't use because it's a lot more work than what I do use, which I've chosen to also call "masking" which prevents the model from learning things that are unrelated to the LoRA and, in this case, would be whatever facial features were present in the dataset during training.

There's another way to do this, and that's to use different subjects for each image, though sometimes the trainer will lock onto a single image and take those characteristics with it, so that's not a consistent solution.

However, when using this non-traditional method of masking, one would simply paint over areas that shouldn't be learned, leaving enough detail for the model to know where the concept fits in. While this can also have its drawbacks, with the white mask permeating the generations, it does lend itself well to the intent and, using an alpha of 1, disparate from the rank, seems to almost completely get rid of the mask bleed-through, which I discovered recently.

One might get the idea of using an alpha or black instead of white, which I've already done and won't attempt to discourage testing but it didn't work for me. I didn't try other colors.

Another thing I discovered, and I'm telling you it's starting to sound like black magic but almost all of the information I've read about LoRA training was simply bogus, the captions can contain, effectively, a system prompt, which I use in order to discourage learning some bits and pieces. I add instructions to my masked images and I tested both sides of the fence, I get almost no mask bleed through with the instructions vs without. Go kick that can down the street a little bit.

I'm sure some guri will want to weigh in on this, to refute my findings, but I won't be listening because I do all of my own testing, and I try to encourage others to do the same, but a lot of people just want to "word" to sound intelligent or important, or maybe just want to dissuade people from succeeding, whatever it is it just blows my fkn mind and I just want to help this person with their stuff.

Good luck over there \o

Even when you feed it an image with tons of cum on the face already, this LoRA manages to get rid of the cum even at 0.55 denoising strength with img2img. The least I can say is that it doesn't really work :/ At best you'll have a tiny drool on the lower lip.

There seems to be a huge margin for improvement in your dataset's captions and images. Cum's known, like any layer on top of a very well understood concept (skin), to be hard to train properly. So good luck finding ways to improve the learning.