Indeed, I tend to agree with you, looking at these snapshots, I think also that EPI with no SDC has a closer fit (or at least looks surprisingly less deformed) to the anatomical image than the image with SDC. This follows the trend of reports showing this with FMRIPREP version > 20.2.7 and 2 magnitude and one phase difference image. See a similar report here and the two propositions to correct for this situation:
- move to fmriprep LTS v20.2.7 and use SDC
- skull strip the phase image (and not the magnitude images) before feeding them to fmriprep (whichever version).
You may try also the fieldmap-less option proposed by FMRIPREP to see if that works for your data if the SDC with provided fieldmap does not work well.