I’m very confident this doesn’t have anything to do with the phase encode direction, and is just a manifestation of the SDC regression that has been noted elsewhere. I am using the same technique: phase difference field map GRE acquisition with phase encoding i, and my bold acquisition has phase encoding j-. I have identical looking distortions with version 22.1.1 and 22.0.2, but the suggestion to try the LTS release (20.2.7) gave a good result. I posted screenshots showing a comparison of both versions on the same data here:
fmriprep-sdc-distort-the-image-and-there-is-error-for-some-subjects
I also tried manually changing the phase encode direction on the phase difference and magnitude fmap images from i to j-, this predictably had absolutely no effect (ran on version 22.1.1, still looked really bad).
For me this artifact is super consistent, I ran around 60 datasets and I would say about half had very severe artifacts. I’ll update when all 60 are done running on 20.2.7 but so far they look good.
I’m happy to help by sharing more examples or data if you need it.