I am working on adding a task for DESC to generate nightly coadds for difference imaging. For a single patch with ~20 visits, it takes hours to run coadding and subtraction. The time is dominated by coaddPSF construction when measuring objects. There are two places in the pipeline where objects are measured, dynamic detection and when selecting objects for Alard Lupton.
I can reduce detection thresholds to limit the number of objects, but it would be useful here to use an approximate PSF that would be less accurate but fast. Is this possible in the future? I know cell-based PSFs will largely fix this, but that adds another layer of complication that is likely not needed in many cases.