Computing FID for Celebrity Experiments

[VMS-6511]

Hi,

I am looking to replicate the results from the celebrity experiments in the paper using this repo. I'm a bit confused as to what dataset is used to compare to the distribution of sampled images from the fine-tuned models? A random sample of 30K images from MS-COCO is mentioned but that doesn't seem right for celebrity images? If so why was that chosen?

Thanks!

[Shilin-LU]

Hi, thanks for your interest.

Yes, the dataset used to evaluate the FID is MSCOCO.

Indeed, this may not be an optimal choice.

We aim to see if the generation quality of the fine-tuned model is affected or not, and using MSCOCO to evaluate FID can reflect the quality of a wide range of natural objects and scenes.