Backpropagating through Fréchet Inception Distance

Published in arXiv, 2021

The Fréchet Inception Distance (FID) has been used to evaluate thousands of generative models. We present a novel algorithm, FastFID, which allows fast computation and backpropagation for FID. FastFID computes the true FID value but is faster than previous algorithms; in theory and practice. It is even fast enough to train GANs directly with FID as a loss.

For a presentation on the paper, see here.