Maybe finding a shot/frame with as many colors as possible (I like the falcon cockpit wide shots for this) from the print-based LDs, you could average/median them and see what results you get. Might help remove the inconsistencies.

I still think one of our scripting/programming geniuses should look into writing a mode average plugin for avisynth or matlab, as the most commonly occurring pixel is more likely to be correct as opposed to the one in the middle or a combination of those available.