Largely agree with this, although I think it would be interesting to formulate in terms of: "what is the mutual information between the fMRI scan and the stimul

Largely agree with this, although I think it would be interesting to formulate in terms of: "what is the mutual information between the fMRI scan and the stimulus". i.e) is there actually more information than a few bits encoding a crude object category, which stable diffusion then hallucinates the rest (/ uses to regurgitate an over-fit image)? Or are there many bits, corresponding spatially to different regions of the stimulus - allowing for some meaningful degree of generalization.