About FID

[liyueying233]

Hi! Thank you for your excellent work.

I noticed a significant discrepancy between the FID results I obtained and the FID reported in the paper. I suspect it might be due to issues with how the dataset is constructed. Could you please let me know how you obtained the 128128 LR (low-resolution) images from the DIV2K dataset? I followed the method described in the readme to generate the LR-HR (low-resolution - high-resolution) pairs for training, but the LR images I got are all 512512. I am wondering if the difference in resolution could be causing the large gap in FID scores?

[IceClear]

Hi. Sry that I do not quite get your question here. The images you get are already 4x upsampled because our model takes the LR images with the same resolution as the output. You may check the LR images in our test set for reference on the huggingface. The link is provided in the readme. You can also check the data generation details in this repo. It is almost the same as RealESRGAN.

[liyueying233]

Thank you for your response. I suspect that the low number of samples in my generated dataset, which is only 1000, might be affecting the FID value. I appreciate the dataset you provided on Huggingface, and I will try it as soon as possible.

Additionally, I would like to ask if it is necessary to ensure that the test dataset has more than 3000 samples to obtain a stable FID.