To answer your question Francois I actually binned 3x3 (after calibrating and combining the images un-binned).
This gives a sampling factor of only 2.1 which is not really enough. Also I had to leave optimum binning turned off in ISIS.
Here is the same data processed with 2x2 binnnig and with optimum binning turned on. Sampling is 3.2.
And here is a detailed comparison.
I'm not sure if the gain in S/N for 3x3 binning is justified for very faint objects due to the marginal sampling.
Tim