I’ll do that. I was erring on the side of more smoothing only because I assumed it mattered how closely the test and reference images were identically aligned.
They should be cropped the same.
If I understand correctly, I could shuffle the pixels of both images into a random order and get the same result.
Exactly this.