augmentations for SPARNetHD
t-martyniuk opened this issue · 1 comments
t-martyniuk commented
why don't you use any of the augmentations you mentioned for SPARNet training except for the "random_gray"?
chaofengc commented
SPARNet and SPARNetHD are trained for different purpose. The former is to compare with previous x8 bicubic downsample methods, and the latter is designed to generate high resolution results for real world LQ images. (I believe this has been clearly claimed in the paper)
If you ask for the augmentations of SPARNetHD. The augmentations are implemented here
Face-SPARNet/data/ffhq_dataset.py
Line 46 in 5204283