A Perfect Round Trip

A team trains an unpaired horse-to-zebra translator in the standard way. A forward generator $G$ turns horses into zebras, a reverse generator $F$ turns zebras back into horses, one discriminator judges whether $G$'s output belongs among real zebra photographs, another does the same for $F$'s output among horses, and a cycle loss requires the round trip to come back: $F(G(x)) \approx x$ for every horse photo $x$.

The run converges, and three things are true at the end.

1. The cycle loss is essentially zero. $F(G(x))$ reproduces each input horse almost pixel for pixel, down to the fence in the background and the photographer's shadow on the grass.
2. The zebra discriminator cannot separate $G$'s outputs from real zebra photographs.
3. Yet when a person puts $x$ and $G(x)$ side by side, they do not correspond. The stripes follow nothing in particular, and the fence that reappears perfectly in $F(G(x))$ is nowhere visible in $G(x)$ at all.

A reviewer says the third observation is impossible: a cycle loss this low proves that $G$'s output keeps the input's content, so somebody must have measured something wrong.

What best explains all three observations together?