What is behind using BCE loss for Variational Autoencoder for images?

Question

I\'m just implementing Variational Autoencoder and wondering what is behind using BCE Loss for RGB images.

In my understanding, for the MNIST dataset, standard VAE is