FID in one dimension

Real images summarised by one Gaussian, generated ones by another. In 1-D the Fréchet distance is just the squared difference of means plus the squared difference of spreads, so the whole metric is visible at once.

Try this

Match the means exactly and leave the spreads different. FID stays stubbornly above zero — generating the right average is not generating the right variety.

FID = (μr − μg)² + (σr − σg)² · lower is better

Where this lab comes from

Image Generation Metrics: FID, IS, KID, LPIPS, and CLIPScore

How do you score images with no ground truth? You compare distributions and use learned perceptual models: Inception Score, Fréchet Inception Distance, KID, LPIPS, and CLIPScore for prompt fidelity, on one running cat-generator example, with the equations, the catches, and torchmetrics.

More in metrics