Different score observed in the website demo

[NaughtyGentalman]

This is a very nice project. However, I have some questions. Why is there a difference between the same image and the one on the personal website (https://baidut.github.io/PaQ-2-PiQ/)? Does it need further training?

[NaughtyGentalman]

I have noticed an interesting phenomenon: a clear watermark on a blurred image can have a significant impact on the results. When I crop the watermark and input it into an image with a smaller size but unchanged content, sometimes the final score will increase.