You said that we have used encoder (q(z|x)) in regularizer, which is KL Divergence, so we do not need to include it into negative log likelihood function. But if we look at the ELBO function, we see q(z|x) in both parts. Don’t we need to include q(z|x) into negative log likelihood? Regularizer is an additive function, so we cannot expect it to be part of negative log likelihood.