Diffusion-LM (Li et al., 2022), a representative work in the continuous diffusion approach, mitigates discretization errors through end-to-end training of the embedding layer and the introduction of a clamping technique.