Some questions about your paper

[zinuoli]

Hi, sorry for the distrubance again, I got some ambiguous points after reading your paper:

1. Do you train CLIP Controller and Restoration Model separtely or train they at the same time?
2. I saw you introduce learnable prompt at this line, which is smart. However, I notice you incorporate prompt_embedding by t = t + prompt_embedding, my question is why you integrate degradation type into time step, instead of by cross attention like this x = attn(x, context=image_context).
3. For the NAFNet there's no time step, how did you integrate prompt_embedding into NAFNet?

Something I didn't find answer in your paper (or I missed), sorry for interrupting you. Thank you for your great work.

[Algolzw]

Hi,

1. I train the two models separately.
2. I directly add the prompt_embedding to time_embedding for model efficiency since we already have cross-attentions for the image_context.
3. For the modified NAFNet (with time embeddings), you can refer to the Refusion's code here.

[zinuoli]

Got it, thank you so much.😄