We introduce GMMag, a workflow for generating crystal structures together with their site-resolved magnetic moments, built on a conditional variational autoencoder with a graph neural network encoder and a transformer decoder. Gaussian-mixture sampling redirects generation towards selected chemical and magnetic targets without retraining. We systematically investigate how encoder and decoder design, the representation of magnetic-moment magnitudes and relative signs, and sampling choices affect reconstruction, generation quality and diversity. Because a higher screening pass rate does not necessarily yield more qualifying candidates with distinct chemical formulas absent from the reference dataset, we model how repeated generation of the same chemical formulas limits their accumulation and test the model in two separate ten-million-draw runs. As an application, the workflow identifies Nd-containing candidates with chemical formulas absent from the reference dataset, negative formation energies and high estimated saturation magnetization.