*4.1. Datasets*

Because there are no large-scale unpaired image deblurring datasets yet to be disclosed, it is meant for comparison with algorithms on paired datasets. Therefore, we still used paired datasets for training but used unpaired training methods. We used two datasets: the GOPRO dataset and a small unpaired dataset built in this work. All models were trained on the GOPRO dataset. The self-built unpaired dataset was only used as a test dataset due to its small number. These two datasets are described separately.
