Closed alexhe101 closed 5 months ago
Thanks for your careful reading. The code should prevail, which the prompt is added in decoder. I will revise the paper.
According to our observation, for U-like structures, the key is to provide different prompt features for each U scale. Adding to an encoder or decoder or both has little effect.
Thank you for your reply, really helpful!
According to the paper, it is stated that the prompt features would be injected solely into the encoder part of the U-Net-like structure. However, upon inspecting the restormer_ep code, it is observed that the prompt features are being injected into the decoder part instead. Why is that? Does injecting into the decoder part bring higher performance?