Inquiry about Data Diversification

[gilyoungCoder]

Hello,

I read your paper with great interest. In comparing RG and RGD, I noticed that for the CIFAR experiments, RGD constructs the remaining dataset (Dr) by sampling 50 images per class, whereas for Stable Diffusion, an LLM is used to generate a diverse set of prompts related to the concept, and then the concept is removed from these prompts to build Dr.

Could you please clarify how RG exactly constructs the prompts for Stable Diffusion, and how the experiments were conducted for CIFAR in this regard?

Thank you very much for your time and assistance.