Hello, congratulations for your work, I have the impression that the L1 loss gives very good results on the metric of the PSNR without the good visual quality. Have you studied the possibility of using a loss funciton such as DISTS or LPIPS mixed with an L1 loss function ?
Have a nice day.