Open forever208 opened 11 months ago
Hi @unixpickle , thanks for the work and code.
with the pre-trained model, I tested the conditional ADM-G and ADM on ImageNet64 with 100 sampling steps and dropout 0.1. ADM gets FID 3.25 (looks normal)
but the FID for ADM-G is strange: ADM-G scale 0.3 = 4.14 FID ADM-G scale 1 = 7.22 FID ADM-G scale 3 = 13.67 FID
But ADM-G achieves higher IS than ADM.
Is this normal in your experiments?
Thank you very much!
Hi, can anybody share the FID of ADM-G on ImageNet 64? I get a worse FID than the unconditional ADM