FID of FFHQ model does not match reported value

[georgestein]

I am attempting to use the FFHQ model provided to generate images that will give the same FID result as reported in the paper (3.39), but I instead get FID=4.36.

I am also generating images and calculating the ffhq 256 x256 FIDs of various other generative models (Stylegan-XL, StyleNAT, etc..) and found that my local version is consistent with what was reported, so I do not believe the error is on my end. Please let me know if it is.

Is the model that matches the reported FIDs available? Or is there a set of publicly available set of 100k generated ffhq256 images?

command and result below:

python calc_metrics.py --metrics=fid50k_full --network=https://s3.eu-central-1.amazonaws.com/avg-projects/projected_gan/models/ffhq.pkl --data=data/ffhq256/ 
Loading network from "https://s3.eu-central-1.amazonaws.com/avg-projects/projected_gan/models/ffhq.pkl"...

{"results": {"fid50k_full": 4.366911864743968}, "metric": "fid50k_full", "total_time": 1171.9112825393677, "total_time_str": "19m 32s", "num_gpus": 1, "snapshot_pkl": "https://s3.eu-central-1.amazonaws.com/avg-projects/projected_gan/models/ffhq.pkl", "timestamp": 1681159541.6151838}