Great, thanks!
I looked and the unflipped was indeed better. But would be really nice if you could check if there is a bug or implementation fault from my side!
I also noticed that in the suggested proc blurring was used with a 4 mm setting. We kind of have a debate here regarding blurring/smoothing.
In a sense the data is already blurred biologically, MR scanning and then normalization. And another thing is that in AFNI we correct for false positives via FWHMx+clustsim where spatial smoothness/autocorrelation is measured and converted into tables of p-values vs cluster-sizes. If we use blur we usually get stronger stats and perhaps bigger clusters but we are then penalized with more conservative clust-tables since we have introduced smoothness. Is it a net gain using blur compared to penalty we get in the per-voxel p-value vs cluster size in the clust tables (clustsim output)?
Since blurring is in the current recommended afni-proc example. (I know this is a difficult question and might be per case basis and depend on the task and MR resolution at hand -but generally...).