The process of masking previously generated tokens in a diffusion model to allow regeneration and correction of that text.