About the FID reported in the RAR paper

[jeasinema]

Hi,

Thanks for open-sourcing the code and models, it's been an pleasant read:) I have a QQ on the FID reported in the paper, which can go as low as 1.48 on IN-1k 256^2. However, as indicated here, the tokenizer (MaskGIT VQ) used by RAR can only attain 2.28 reconstruction FID on the same setting. I am wondering if you have some insights on this or maybe I missed anything. Thank you.

Best,
XM

[cornettoyu]

Hi,

As discussed in #46 , rFID and gFID are computed differently and thus are not directly comparable (if you look at MAR's table2 they also report a gFID < rFID).

[jeasinema]

Thank you!