I have test it for custom images(include human and dog , low resolution).It, the upsample image, has no difference with the resize image, sometimes even wrong than resize馃槀.
But for the example image, it does help, so it is not "an off-the-shelf superresolution model" which conflict with the paper ("Figure 3. Our GAN-based upsampler, similar to Figure 2, can also be used as an off-the-shelf superresolution model for real images with a large scaling factor by providing an appropriate description of the image")
In the situation of upsampling image, we does not have super resolution image (if we have, we do not need it).
I have test it for custom images(include human and dog , low resolution).It, the upsample image, has no difference with the resize image, sometimes even wrong than resize馃槀.
But for the example image, it does help, so it is not "an off-the-shelf superresolution model" which conflict with the paper ("Figure 3. Our GAN-based upsampler, similar to Figure 2, can also be used as an off-the-shelf superresolution model for real images with a large scaling factor by providing an appropriate description of the image")
In the situation of upsampling image, we does not have super resolution image (if we have, we do not need it).