Position Embedding Operations on Latent

[Litoda]

Hello, may I ask if operations like farthest point sampling, KNN construction, and position encoding can be applied in the latent space? Are these calculations performed in high-dimensional Euclidean space? For example, calculating distances in a 64-dimensional Euclidean space for features with dimensions (512, 64)?

If the trilinear interpolation can operate on latent space, is there any trainable weight of this?