There is no difference. In fact, any quasigroup operation will do.
Specifically, the only property we really need is that, for every ciphertext symbol $C$ and every plaintext symbol $P$, there exists one (and only one) key symbol $K$ such that encrypting $P$ with $K$ yields $C$. This implies that, as long as the actual key symbols are chosen uniformly at random (so that each of them is equally likely to be correct), knowing the ciphertext symbol $C$ does not reveal any information about the plaintext symbol $P$ (since each $K$, and therefore each $P$, is equally likely).
In fact, we can generalize further and allow several key symbols to produce the same encryption, as long as the number of key symbols that yield $C$ is the same for every plaintext symbol $P$.
There are several common operations that can satisfy this property. For example, we can use modular addition for encryption, and modular subtraction for decryption, or vice versa. Or we could use e.g. bitwise XOR, which is its own inverse, and can thus be used for both encryption and decryption in the same system. Or, if we wanted, we could use e.g. multiplication in a Galois field for encryption, and multiplication by the group inverse for decryption (or, again, vice versa).