Hi Matilde,
I'm not an expert, but I can tell you that this isn't going to work the way you want it to. The problem with trying to use 3dGroupInCorr with each subject in his/her own space is not just that the seeds will be in different places for different subjects, but that each subject's correlation map will be in that subject's space as well. 3dGroupInCorr needs to be able to run a t-test on (transformed versions of) these correlation maps, and for that to be meaningful, the maps all need to be defined over the same space.
On the upside, transforming all the subjects' data to the same space is cheap and easy, so why not do it?
ijs