How to calculate intra-modal uniformity?

[Leg-end]

As mention in the paper that the intra-modal uniformity was computed using data sampling from image and text data distribution.
Does that mean norm distance from samples of same class also contribute to the metric when number of shot > 1 ?
Or instead, using prototype(mean) feature of different classes so only distance from different classes will be considered ?

For example, provided with a k-shot data, k > 1, it has embedding of shape like [NK, d].
Should I average it to shape of [N, d], then compute the intra-modal, or directly do it on unaveraged shape ?