How can you imporve the accuracy in last submittion, please?

[huang-yuefeng]

Dear Sir,

Thanks so much to your work, we are trying some new idea based your work.
We have seen that you have submitted Feb 10 a new picture result with great improvement to previous one.
Would you like to share your experience to achieve that, please?

[leehomyc]

Which one are you talking about?

[huang-yuefeng]

Thank you for reply.
I am referring this Photo-Realistic-Super-Resoluton/pics/_s.png, the effect is very nice.
More over, in our training on Helen dataset, our generator error falls below to 0.0100, but increases to 0.0200 again after discriminator error falls below to 0. I don't know why.

We are just replaying your result in Photo-Realistic-Super-Resoluton/pics folder, I think this is perfect. So if you can share more details, that will be very great.
Thanks and Best Regards.

[leehomyc]

I think that image(_s.png) is ground truth. I have had the same problem but did not figure out why. Current result shown is given by early stopping. I am not sure if adding tv-loss would help.