Deblur trained on natural image does not do much

[balthazarneveu]

Several assumptions

• NafNET is too hard to train -> try UNET
• Bad hyper parameters?
• The task is too hard on natural dataset due to:
• many flat areas . mining? -> probably
• blurs are too big / too difficult to hold.

Results

-> Training on natural images gets "lazy", when the blur is too big, network does not take risk