search

lbc12345 / SeD

Semantic-Aware Discriminator for Image Super-Resolution
94 stars 5 forks source link

metrics in Table 1 #9

Closed llmpass closed 2 months ago

llmpass commented 2 months ago

Hi,

I found that the metrics in ESRGAN (RRDB+P) are way better than what reported in Table 1 of your paper, why? For example, Set 5 psnr is 32.73db in ESRGAN, but only 30.67db in your paper.

llmpass commented 2 months ago

Never mind, I found that your esrgan has 2 pretrained models...

lbc12345 commented 2 months ago

Hi~ The 32.73dB you mentioned is the PSNR-oriented weight from ESRGAN, which not includes GAN training. Our reported RRDB+P denotes the RRDB+PatchGAN training.

llmpass commented 2 months ago

Thank you for your reply. Do you use this PSNR-oriented model to start fine-tuning in your method?

lbc12345 commented 2 months ago

Yes, as this is a common operation in GAN-based image SR training schemes. Performing PSNR-oriented training first and then introducing discriminator to start GAN training.