Figure 1. Three representative types of SISR regions generated by
ESRGAN [40]. For each example, the left is an LR patch and the
right is its GAN-SR result. Type A patches represent regions that
are easy to super-resolve, e.g., smooth and large-scale structural
areas, where the main structures are preserved in the LR input. In
contrast, patches of type B and type C are with fine-scale details,
which are hard to be faithfully restored due to the signal alias-
ing in the LR inputs. The results of texture-like type B patches
are perceptually realistic despite the pixel-wise differences to the
ground-truth, since the patterns are naturally irregular with weak
priorsforobservers. However,theresultsoftypeCpatchesexhibit
perceptually unpleasant visual artifacts since the overshoot pixels
and distorted structures are sensitive to human perception.