The attached plot characterizes the variability in a 4Myr-old OB association. If you want to take the colours at different times, and want less than 10 percent of the stars to have varied by more than 0.05 mags the quick answer is that you want to take the two different colours within a night. In fact (see detailed arguments below), I think we should aim for 20 minutes.

Here are the details, and the ifs and buts.

(i) The data were taken in the I-band, in fact pre-main-sequence stars will be less variable in JHK.

(ii) The sample is photometrically selected, so will be contaminated by field stars, whose variability is presumably less than the PMS stars. I'd be surprised, though, if the contamination is 50 percent, so perhaps we should double the fraction that have varied by a given amount.

(iii) The 8 minute sample probably represents my photometric uncertainty, not any intrinsic variability, and so should really be subtracted from the other curves. My sample was chosen to have fractional single-point uncertainties of less than 0.02.

(iv) Younger stars will vary even more.

(v) I'm only speaking for "my" stars here, some other sorts of variable my change more.

And now my argument for 20 minutes.

The extra overhead for changing filters every 20 minutes is relatively small, and my 33 minute curve is little different from my 8 minute one. You'd have to have a very strong argument for the extra overheads, which in this case seem to give little gain. So, why not go to 2 hours? Well, what we don't know is how rare objects a science programme may be interested in. For example, if you expect 1 in a hundred stars to deviate from some colour relationship by 0.05 mags, a time separation of 2 hours completely stuffs you, as more than 1 percent of the stars will have varied by that amount in 2 hours. So, my pitch is for 20 minutes filter cycle time (assuming we are doing two filters).