SRGAN not the best thing for high resolution input, due to full size convolution. Only thing that how this can be processed multi-patch approach, when you process not full image but some regions and then combine them together, minimizing error on edges.