IceClear
Training dataset question (whether crop small patches)
May I ask, for the DF2K and DIV8K datasets, did you crop them into small patches in advance during training? Because this will affect the proportion of various types of data in the final training set. For example, if cropped them in advance, there will be many patches from DF2K and DIV8K in the training set. Also, how did you treat other datasets, OST, FFHQ, do you crop them?
We follow the same settings as RealESRGAN. We did not process the data in advance.
We follow the same settings as RealESRGAN. We did not process the data in advance.
Hi, this is still a bit confusing. You mentioned following RealESRGAN, but RealESRGAN by default crops the images in DF2K into small patches. To clarify, did you not preprocess the image and read it as a complete image?
We follow the same settings as RealESRGAN. We did not process the data in advance.
Hi, this is still a bit confusing. You mentioned following RealESRGAN, but RealESRGAN by default crops the images in DF2K into small patches. To clarify, did you not preprocess the image and read it as a complete image?
Yes. Pre-cropping as RealESRGAN and directly reading the whole image are not very different from my view. Finally, the GT should be a fixed resolution and cropping is a necessary step. For details, you can refer to the related code of this repo.