why deep-adversarial-decomposition worked in Rain800, DID, ISTD, SRD dataset ?

[zjupx]

As mentioned in your paper, gt2 uses ‘zero image’ in these datasets. Does this problem degenerate to standard pixel-wise l1 loss and discriminator loss? So why is your method better than other methods in this case? Which one do you think is working for this case? I really want to ask you to discuss this question. I hope to hear from you, thank you ^.^