Currently the arrow lengths are normalized according to the conventional sum of the eigenvector components (probably a mass-weighted normalization).
This, however, means that the more atoms the unit cell has, the smaller the arrows are.
@giovannipizzi proposed to normalize the arrow length of each mode instead such that the longest arrow is always the same. This can help visualize modes for systems with more atoms, but potentially it might also have drawbacks. One potential drawback is that it's harder to visually distinguish localized modes from delocalized ones (as both have similar length arrows).