Hi :-D thanks for the nice words. It's not that I haven't tried, the issue is the translation of the make-up and the skin when training. If you'd see the type the type of images that come out of the training process you'd see that it is not training the face, it makes an interpretation of it which is why even copies of celebs never really catch the character of that person right. When you translate these "ghost faces" back into the generating process it's not simply doing what a face swapper would do, it's building the face from the ground with some features stored in those weird face images. Make-Up is a variable and not a feature which it will keep interpreting differently based on your prompt situation. I tried and tried and tried...
Whenever I went into hand painting make up in a make-up program (PortraitPro Max) and added more vectors and used a higher step count, the training model went straight to copying real life faces that people used when building the base model and I ended up deleting those girls because they looked like a real person. (recently deleted my 8 best performing models) I think the only way to achieve this is with an add on embedding created for just that girl or even a lora but that will result in face distortions. I'll keep trying :-D I do like Goth girls, they are deadly nice :-D

one thing you could try is using controlnet or After Detailer, to apply the goth make-up. If you use ADetailer simply put the embedding in the front of the prompt inside the add on or the base model will pain a different face on top of it.