Open AlphaNext opened 10 months ago
Thank you for your interests and attention.
To achieve degrdation transfer, we need these images: an input image 1 with deg type A, a reference image 2 with deg type B, a corresponding GT image with content 1 and deg type B. In the code implementation, the original clean images (img_HQ1 and img_HQ2) are first transferred to img_LQ1 (as target LQ, the corresponding GT image with content 1 and deg type B) and img_LQ2 (as reference). Then, we add another degradtion type to img_HQ1 to get the input (img_HQ1).
input_noise_flag and input_blur_flag are not supposed to be used in the training process. They are prepared for the degradation transfer test, which is not included in the currrent codes.
Thanks for your work and released code. There are some questions about DDG_train_Pretrain_Complex_dataset.py file.
https://github.com/lyh-18/DegAE_DegradationAutoencoder/blob/0e9611301aafb9d17752acac05f56db76a9f8ff0/codes/data/DDG_train_Pretrain_Complex_dataset.py#L100-L114
https://github.com/lyh-18/DegAE_DegradationAutoencoder/blob/0e9611301aafb9d17752acac05f56db76a9f8ff0/codes/data/DDG_train_Pretrain_Complex_dataset.py#L123-L137
img_LQ_list and img_HQ_list will be same output, is this right ??
and where to use input_noise_flag and input_blur_flag ?