I do not believe that the pixel values in coadds correspond to nJy, pending DM-17018. John Parejko may wish to correct me.

A few things to note here:

A calexp is not calibrated. The pixel values are still instrumental fluxes (with units of counts), not fluxes (with units of nJy). Were you referring to an Exposure that you had calibrated to flux using photoCalib.calibrateImage()? This is one of several reasons that I don't like the name "calexp".

Second, as John Swinbank points out, our coadds are not in nJy either, but rather a "counts-like" unit that is a constant multiple of nJy such that the magnitude at zero instFlux is some pre-set value (typically 27). Note the various arguments made in the comments of RFC-545.

The warps that are produced by makeCoaddTempExp (can we finally get that renamed to makeWarp, please?} should now be in the same "counts-like" units, with the same magnitude at zero instFlux as the final coadds (I think, but I'd have to double check the current code).

In short: nothing to see here. Thanks John!

Following discussions with John Swinbank and John Parejko:

• As we have a nice, spatially varying photometric calibration, we do not want the PSF matching in ip_diffim to do a photometric scale matching job, in general. Ideally, the PSF matching kernel solution should be sum 1.
• DM-17018 The coadds should be in nJy. Otherwise tracing the zeropoint for cutting, warping, etc. is cumbersome and error-prone.
• Currently, coadds do not have error on their calibration and they are not re-calibrated. It is assumed that all of the coadd input images were calibrated and during coaddition only the standard deviation would go down as sqrt(N), but there is no need for recalibration.
• Another argument against recalibration is that coadds may be too deep compared to the reference catalogs unbiased brightness range. Though it would be an argument for re-calibration that the S/N of all sources are generally better.
• Once templates are in nJy, ip_diffim may use the calexp photocalib instance to convert calexps to nJy prior to PSF matching. Also, difference images then will be in nJy. They currenlty inherit the scale of the science calexp, whatever it is in the template convolution case. In the science convolution case, they are rescaled by the sum of the kernel solution evaluated at 0,0. There are no guarantee that difference images are on the same photometric scale.

One point on this:

Though it would be an argument for re-calibration that the S/N of all sources are generally better.

Since jointcal and FGCM already use all measurements of each source when computing the calibration, we already get most of the benefit of this anyway, so I don't think there's much to be gained.