In the DM-24371 implementation, we throw away the difference image variance plane as it cannot be afterburned due to unaccounted covariances. However the variance plane calculation from the originals does not consider the effect of the overall (afterburned) convolution of (the original) pixels at the moment which it could do.

For more accurate variance plane calculation, re-do the variance plane operations separately on the science and template image original variance planes using the idea of ZOGY eq. (26)-(27) then add them. This is expected to increase the CPU cost by x4 of the afterburner. Consider keeping this approach optional, especially not doing on the spatially varying case just once for the whole image.