Reference-based Restoration

[dolphin0104]

Hi! Thanks for your wonderful work! I have a couple of questions:

1. Could you please share the details of the test dataset you used in the paper to evaluate the reference-based restoration?
2. I wonder why the example code on GitHub differs from the parameters described in the paper?
   for example..
   1. In the paper, the default setting is cosine similarity loss, while in the GitHub code, 'Choice 1: MSE Loss (default)' is set. How can I reproduce the results reported in the paper?
   2. As far as I know, the 'ref_weight' is set to 10 in the paper. But, in the GitHub example commands, it's set to 25 for MSE loss and 1e-4 for cosine similarity loss.

I'm looking forward to hearing from you soon!! Thank you for your time!!

[lincoln0211]

Hi, I'm also curious about the answer to the question above.