FID-infinite comparison and discussion

[clovaai]

FID-infinite comparison and discussion

[coallaoh]

Thank you too for the interest in our work :)
And thanks for pointing out the interesting work. We are indeed aware of the paper, but have not managed to include a discussion in the first version of our paper.

Please note that our claim is not to replace FID and Inception scores with our density and coverage metrics. They serve different purposes. FID and Inception scores provide single-value assessments that are useful for ranking generative models by their performances. On the other hand, density and coverage are fidelity-diversity metrics (as are precision and recall metrics) that serve more diagnostic purposes. Our argument is that for fidelity-diversity metrics, density and coverage are better than precision and recall metrics. We believe it makes a lot of sense to use our density and coverage metrics together with the unbiased FID or IS in the future.

We will include this discussion in the next version :)

[ogencoglu]

Thank you for the timely and informative response Seong!