Cross entropy penalizes wrong class significantly better. In terms of KL and BCE difference, you can change the KL divergence to mean as well. Different losses might have different learning rates, so empirical testing should tell you the ratio of the losses