Ok. First thing I wanted to say is that you don't have to remove the stars when the data is still linear but I like to remove them early on so I can visualize the data without the stars and it helps me get a better stretch. For my RGB data I remove them after PCC and for the lum before deconvolution. You could alternately do a minimal stretch on L and RGB and remove the stars after the LRGB combine.
If you want to remove the stars while still linear, the following tutorial illustrates how to do that:
https://thecoldestnights.com/2020/06/pixinsight-deconvolution-on-starless-images/For the lum I like to use an initial masked stretch so as not to blowout the highlights. The lum should look dull and gray so as to work well with the RGB when combined. For the RGB I do a small ArchSinH stretch followed by a masked stretch keeping the colors good and saturated. Then do an LRGB combine and follow that with an Exponential Transform or HT. This second stretch is really where the details can be brought out.
I then save the starless LRGB to a TIFF file and load into Topaz Denoise AI to address the low signal noisy areas. This is done judiciously. I little noise is better than an oversmoothed look. Topaz also has sharpening but as I mentioned before if overdone it creates artifacts that may initially look good but are completely bogus so I recommend using Topaz sharpening with constraint. Something I had to learn through experience.
Then I load the resulting denoised TIFF image back into PI and have to do an ICCProfileTransform to get it back to my default ICC profile otherwise weird things can happen with the colors. Topaz seems to change the ICC profile even though it says it does not.
So for blending the stars back, I really can't get into too much detail because of the terms of service of my Adam Block Studio subscription. But I highly recommend his videos if you are not already a member. We spend a lot on our equipment but last year I made the commitment to invest in my processing skills and I feel it has really paid off. The basic premise is that with masking you replace the residue left over by StarNet for the bigger stars in the starless LRGB image with an image representing the background (created with MLT) and then blend an instance with the stars in it that is stretched to taste with the starless LRGB denoised image. That way you get increased detail of a more intense stretch without bloated stars with big halos. This process may be made simpler with StarXTerminator in place of StarNet which results in little or no star residue being left over when the stars are removed but StarXterminator requires a paid license however there is a 30 day free trial.
Final steps would be LHE or MLT bias adjustments to higher layers, dark structure enhance script if applicable and curves for contrast/saturation adjustments.
For some dark nebulae with Ha regions I also sometimes add Ha data but that gets more complicated.
Good luck on your image processing.
Jeff