question about the input shape about cutblur HOT 1 OPEN

Train an SR model with a large patch size improves the performance, but it also drastically increases the training time and the memory consumption so that most of the methods use limited (48x48 or 64x64 for x4 scale) patch size.
However, from my experience, I have observed that performance improvement of using a large patch size is marginal when the patch size is larger than the 64x64. This may be because the SR model refers to the very adjacent neighbors when they reconstruct the pixel, but still, this is not clearly proven.

from cutblur.