How to reproduce about class-to-image fid

[zqc214]

Hello author, this is a great job！

I want to reproduce the metrics of the fid in the class-to-image experiment.

May I download the imagenet1k dataset and then traverse 1000 categories and generate 50 images per category, for a total of 50000 images.

Finally, take these 50,000 images and the dataset of imagenet1k to calculate the fid indicator.

[Shenyi-Z]

Thank you for your interest in our work! It doesn't need to be so complicated. You just need to follow this link. The original DiT used it for evaluation. You’ll need to download the reference batch's npz file from here, then use it along with the npz file generated by your model (directly produced by sample_ddp.py) and evaluate them using the evaluation code evaluator.py provided here.