I like to compare 2 subjects (case study) to a group of 20 subjects. Can I use "-1" in randomise (sign-flipping)?
No, that won't be the correct option. The -1 option is only for a one-sample t-test, and, for VBM data especially, never applicable.
For this design, you have 22-choose-2 or 231 possible permutations; all p-values will be multiples of the smallest possible p-value 1/231=0.0043. This seems like reasonable enough to conduct a test.
Should I add variance smoothing (-v) in addition to smoothing ? How can I estimate the appropriate value of -v?
Variance smoothing isn't really needed for fewer than 20 DF; you have exactly 20 here and I won't think it would do much. (When used, a variance smoothing equal to the original smoothing kernel is usually recommended).
-Tom