This could be related to a previous discussion re the alignment process in fmriprep.
You can see more details here Is it possible to specify a target EPI volume to realign to (motion correct to) in fmriprep?.
The current thinking in our lab is that the fmriprep pipeline may not be optimal when requiring very consistent voxel alignment across sessions for each subject. This is required (for example) for single voxel encoding models (and I assume RSA too).
Fmriprep may be more suited for more “standard” group level analysis where data is analyzed after transforming to a common template and some smoothing is applied. In this case maybe slight voxel misalignments at the single subject level do not matter.
This thinking is based on some small and simple comparisons between data processed using our in-house pipeline and fmriprep, so could well turn out to be wrong ;).
Btw, you will notice at the end of that linked thread that I am supposed to check something to help drill down where the issue may be. I have not yet done this but I will report back on that original thread once I do.