In literature, when VAEs are used, the latent variables are samples with both a learned mean and covariance (usually diagonal) matrices. Can we use VAEs with sampling latents only with learned mean, but a constant variance across all samples/classes?