I was at SUNY Stony Brook to give the astro seminar. I took the opportunity to spend some quality time with Pizagno, who is converting multi-band SDSS images into stellar-mass images by various methods, for the purposes of making stellar-mass-based structural properties and comparisons with kinematics measurements. The big issue is how to work in the low signal-to-noise parts of the images; everything would be fine if the conversion from CCD counts to stellar mass was a linear operation, but it is not (because components of stellar mass must add non-negatively and also because there is dust absorption). Pizagno's very nice solution—which he is testing against the other ideas you might have—is to pull the low s/n pixels towards a
prior color determined by averaging the very low s/n pixels. His method is correct even if the low s/n parts of galaxies are a different color from the bright parts (which they are, in general), but it also makes use of the pixel-to-pixel information you do have at low s/n, and it correctly implements a gaussian prior in each pixel. I love the method technically, of course, but it also seems to give much better results than, say, doing adaptive smoothing or medianing at the outer parts.
BTW—for the nerds—non-negative linear optimization is awesome these days. Read Blanton & Roweis on K corrections for a non-trivial application in astronomy.