We just initialize a random latent vector for each model, and then jointly train each of these unique latent vectors :)