About the performances with iCT.

[limsanky]

Hi @Kinyugo !

Thank you so much for your great work! I had a question about this implementation.

Are you able to reproduce the improved performances using the Consistency Training scheme described in the "Improved Techniques for Training Consistency Models" paper?

I am running a different codebase but am unfortunately only get around 10~12 FID on Cifar10.
I would appreciate any responses and/or help!

Thank you in advance!

[Kinyugo]

Hello. Unfortunately this is an issue that others have been facing as well. While the implementation reflects what's in the paper, I think I might have missed something thus the results aren't that good.