Satra--- can you send the command line?
So a couple of possibilities outside of the random/weird library changes... and assuming you ran the same parameters--- (i.e. # of threads and same burn in, as well as sampling from both the f1 and f2 population).
The module uses a random seed to start off the Markov Chain module--- which can be set explicitly or more commonly left blank and is set by whatever random number generator they use---- the markv model generally converges after running 5000 or so simulations per voxel... i.e. if you run 10,000
-P,--nsamples Number of samples - default=5000
vs 5,000--- your output image will roughly have "twice" the intensity as running with 5000...... but the numbers (I don't think) are guaranteed to be exactly accurate...
--rseed Random seed
See if specifying a specific rseed and running it back to back produces the same results...
Just a guess.