search

ShuhangGu / DASR

Training and Testing codes for our paper "Real-world Image Super-resolution via Domain-distance Aware Training"
235 stars 21 forks source link

About DSN #2

Open qibao77 opened 4 years ago

qibao77 commented 4 years ago

For the design of DSN, how do you prevent DSN from completely becoming a "bicubic" down sampler, because when DSN is a "bicubic" down sampler, the loss will be very small.

ShuhangGu commented 4 years ago

We use a combination of different losses to train our DSN, the loss between generated image and bicubicly downsampled image helps to maintain the content in the HR image. The adversarial loss helps the generated LR image capture the character of real-world LR images.

qibao77 commented 4 years ago

Thank you for your reply! I used the combination loss introduced in your paper, but my DSN became a "bicubic" down sampler. Can you give some advice?