I am glad for the perspective that separate residuals in this case would probably be fine. The plan is to do representational similarity analysis following on from a behavioral study looking at an interesting subset of ~700 images from the full set. So deriving one "representation" per image is the goal.

I would have been comfortable modeling each session separately if each session contained the same conditions---that would feel like a good fit for mixed effects analysis. But since the conditions are different across all sessions and runs, I was reluctant to circumvent the memory limitations by splitting the data by session.Thanks again.