Correct me if I am wrong, but this is not a VAE. This is a normal AE. For a VAE you would need to use tf probability or another mechanism to get uncertainty into each of the experts.
So as a result I woulf assume that this model will always output the same image, because the weights are fixed and not drawn from a learned distribution