I came across similar issues in DM-28294, where ci_hsc has regions of the image with a partially masked out PSF in the I-band:

I've tried to test the impact of the masking problem in the data by doing a separate run where I throw out any star that has masking. To increase the number density of stars I set the reserved fraction to zero. This results in roughly the same number of stars as the original settings. I processed 40 randomly selected i-band visits and then compared some different diagnostics.

Rho1 averaged over all visits:

Rho1 @ 1 arcmin for each visit:

This shows very little difference when using the masked objects.

There appear to be a couple of issues going on:

• Cases where there are too few stars to constrain the psf model. Things get especially bad when the PSF is interpolated into regions where there are no stars. This will be problematic in HSC for the some of the outer CCDs that are heavily vignetted. We could consider requiring a minimum number of PSF stars for an image to be used.
• There are artifacts in the PSF models due to the attempt to mask out neighbors. Below is an example of the full postage stamps on the left and the masked image fed to psfex on the right. We could potentially avoid these by using the deblended images or by removing them out. Some preliminary numbers from HSC suggest that 75% of stars are isolated.

Here is an example PSF image of the problems that are being seen.