Conflicting Instructions for computing FID

[Kumbong]

The paper mentions using CFG of 2.0 to compute the FID and inception scores reported on the paper, while the codebase mentions "For FID evaluation, use var.autoregressive_infer_cfg(..., cfg=1.5, top_p=0.96, top_k=900, more_smooth=False)" Wondering which is the correct one ?