better this way...read below
You are adding pimples and bad skin , instead and versus adding healthy skin pores that have no pimples or bad skin pores. if you optimized you training dataset of images, with images that have professionally retouched skin pores, the top tier ones, that look real and not over retouched, you will likely have the best skin lora around. Food for thought....
Thanks for the feedback @akos2
Which version are you using or which comparison image are you using to determine this?
I ask, as there was issues training the first version, which added artifacts and unwanted noise too, not only to skin but the overall image due to a bug in AI Toolkit, so I retrained it as version 1.1 which is more optimal, although has less of an overall effect due to the artifacts and unwanted noise being removed.
The dataset I used is a professionally curated high res set that is entirely focused on retouched skin, through professional cameras for portrait photography.
Having said that, I still have issues with how AI Toolkit performed the training, so I've been testing the same parameters with Misubi Tuner and my results are better, which will become version 1.2 and released over the next few days.