Position embedding is added inside each block?

[LinB203]

DiT is to add positional embedding once before the transformer block. And according to Figure 3 (c) of FiT's paper, it seems that the position embedding is added inside each block. Is this an unfair comparison?

[WZDTHU]

This is a common implementation of relative positional embedding.