NoisyTwins Augmentation

[nuclearsugar]

I was reading this paper and the augmentation strategy looks quite interesting. The new method preserves the intraclass diversity with class-consistency in image generation and it outperforms other methods by ∼19% FID. Would you considering implementing it into this repo?

https://rangwani-harsh.github.io/NoisyTwins/
https://github.com/val-iisc/NoisyTwins
https://arxiv.org/abs/2304.05866

[PDillis]

Hi, sorry for the late response.

It certainly looks interesting, and the code is available. I'll definitely leave this as a nice addition, but I have other projects lined up and I have left this repo a bit in the background. I hope to change this in the coming weeks, but for now I don't have any major additions planned tbh