The removed star areas had a hatch pattern and the sizing didn't match perfectly.
This even showed up in the star mask I created, but a small blur fixed it for my uses. It could be a number of factors, grungy stars (I know they're not perfect), many of the stars have reflection nebula around them, oversampled stars, stars bleeding into each other, my processing, my stacking, the list goes on.
Don't take this the wrong way. I think this is a great piece of work, and it gets me 90% of the way to making an effective star mask, something I've struggled with in the past and need to get working as I transition from OSC to RGB to narrowband.
I'd love to be able to assist with training, but I have no decent examples of destarred photos to help with.
And an informal poll at work (the guys in my cube and another who's a wedding photographer) all preferred the original to the star reduced version. I prefer the star reduced version.
Don't worry, I know it does not work perfectly on all data. Just collecting some info on when it does not work so I could try to add relevant training data and make it better.
BTW. I actually also prefer original version!) lol)