Reproduced results

[pietz]

Did you achieve your posted results using the default parameters only varying the scale, training file and eval file? This would be interesting because I believe the results you posted from the paper were based on the imagenet dataset und not the T91 dataset.

Also, why did you use a custom RGB to Y function? Wouldn't a standard grayscale function do the trick?

[chensteven]

The paper did convert RGB to YCBCR and then use only Y channel. Grayscale would technically be different.

[543877815]

I think the result is not precise. The Set5 dataset from the download link seems to have a higher (BICUBIC) PSNR than other sources where pictures isRGB instead of y channel only.