Two Clouds, One Score

You are scoring a generator the way FID does — not image against image, but the whole cloud of generated samples against the whole cloud of real ones.

Here is that idea stripped down to one dimension. A feature network has already reduced every image to a single number. Five real photographs and five generated samples come out as

The score compares the two clouds through their centre and their spread only:

$\mathrm{FD} = (\mu_r - \mu_g)^2 + \sigma_r^2 + \sigma_g^2 - 2\,\sigma_r\,\sigma_g$

where $\mu_r, \sigma_r$ summarise the real set and $\mu_g, \sigma_g$ the generated one. Use the population standard deviation — divide the squared deviations by $n$, not by $n - 1$. Lower is better, as always.

What score does this generator get?

Round the answer to 2 decimal places.