A close-in voltage divider, referred tightly to the input-signal-
of-interest reference ground, is the best option I've ever seen.
At 20mV you will be challenged to test in any but the most ideal
lab bench conditions, or have to play teicks like putting the
assembly inside an integrator loop looking for Vdd/2 output
and filtering out the chatter.
The only thing you have to fear in setting up a large ratio
divider, is any unanticipated shunt conductances about the
high valued resistors. That's basically board fab & handling
hygeine, conformal coating and moisture purge (for low temp)
etc. along with keeping the shunt value of the divider well
below Zin(comparator).