DIV2K reproduction

[sobieskibj]

Hi,
congratulations on a very nice work. I'm having problems when reproducing the FID results of your method. Are you using the official DIV2K validation dataset for evaluation? What is the actual of your method there? Do you just take the original high resolution image (e.g., 2000x3000), downsample it and upsample to match the original resolution? Or is there some e.g. center cropping applied to first obtain the same resolution as during training? The FID values that I'm getting are around ~140 for 100 images from the original validation split.

[Algolzw]

Hi, thanks! I use the official DIV2K validation dataset. The image is bicubicly upscaled with this code. You can check the visual results which should be better than LQ images (obviously). If not, there must be some wrong with the upscaling.