Figure 2: Motivation. CKA (Kornblith et al.,
2019) (left) reveals weak feature correspondence
between DBSR and DFSR, while residual and
zoom-in comparisons (right) show local discrep-
ancies between DBSR outputs and paired HR
targets.