# Ordered elements of feature vectors for autoencoders?

Here is a newbie question; when one trains an autoencoder or a variational autoencoder, does the order of the objects in the training vector $x$ matter?

Suppose I take an MNIST image image $(28\times28)$ and turn it into a feature vector of size $x \in \mathbb{R}^{1\times784}$. Then does it matter if I e.g. flatten the whole image vertically, or horizontally, or some other fancy way? Or if I were to scramble the order of the elements in the feature vector, would that make the VAE or AE mess up?