Is it still relevant to measure IS and FID if not using ImageNet?

edit

Yes, it is still relevant to measure IS and FID for evaluating the quality of generative models even if not using ImageNet. These metrics provide a quantitative measure of how similar the generated samples are to the real data in terms of their distribution and visual appearance, which can be useful for assessing the performance of a generative model in various applications. Additionally, there are alternative datasets that can be used for evaluating generative models outside of ImageNet, such as CIFAR-10, STL-10, and MNIST, among others.