I explored using a relative model throughout building the DcrModel, but found it increased the code complexity and computation time. I was able to get the same results by using a smoothly-varying relative model in frequency regularization, which is comparatively quick to calculate. The new regularization significantly decreases chromatic artifacts in the wings of sources, and for the first time allows sensible color images to be made directly from the DcrModel subfilter images (see below).

To produce the above image, I generated 12 simulated images using StarFast with airmasses ranging from 1.0 to 2.0, each including main sequence stars with Kurucz SEDs and quasars using the Vanden Berk quasar spectrum with redshifts from 0 to 4. I ran dcrAssembleCoadd on the ingested images, and loaded the resulting three DcrCoadd subfilter exposures into DS9, using the longest wavelength subfilter for the red channel, the center subfilter for the green channel, and the shortest wavelength subfilter for the blue channel. Bright objects are clipped and white for this color scale, but for the fainter objects those that are very red are quasars and while those that are blue are mostly cool stars with a steep spectrum in g band.

Please note that this ticket includes code in both ip_diffim as well as pipe_tasks, though due to a long-standing bug the pipe_tasks pull request does not show up in Jira.

Code looks all right, as far as I understand. A few documentation related minor comments:

• I'd suggest adding a few lines of docstring to regularizeModelFreq describing that you use a combination of 2 different scale smoothing on the relative model, plus a note on the hard-wired factors of 3. (Where are they from?)
• May not be related to the present ticket, I could not find any docs from dcrModel.py in pipelines.lsst.io docs builds (e.g. DcrModel). Should classes be imported into the package top level _init_.py ?
• Not strictly related to the present ticket, I think in the description of getReferenceImage we might want to avoid calling the "mean model" as "template" in any context. It may be also worth noting somewhere that in the present implementation the "reference" is the "mean model" in general, and not something like another standalone image.