I think adjustments to this may be in the works for the next major alpha still, based on last comments in the 1.9 thread?
A couple things you might try -
I think .541a had improvements to this, so depends what version you are using. I still often use .536a as I prefer how it handles deringing, but 541 can be better with sample outlines.
Binning down the dataset can seem to help with sampling also, as long as you retain enough resolution for SVD to get a good deconvolve.
You might try going stronger on your OptiDev final global stretch and/or with enhancing modules just make sure stuff hasn't dimmed out too much. If I recall from discussions of starfishy sampling while still in 1.8, while the colors of the samples are brought from the data's linear state, the white outlining may be coming from stretched states. That's just off the top of my head and I haven't experimented with it, and I may be confusing it with the binning changes allowing better sampling.
I am not aware of any way to trick ST into using manual mask generation in 1.9.
EDIT: Looks like I was wrong about the level of stretch - did a quick test and the apod mask generation was exactly the same regardless. Oops.
Binning, however, did change a full field of unusable little white outlines into actual usable stars.