Discrete MeanFlow: One-Step Generation via Conditional Transition Kernels

Discrete MeanFlow extends the MeanFlow one-step generative framework to discrete state spaces by replacing continuous velocity fields with conditional transition kernels of continuous-time Markov chains. This enables one-step generation for discrete data like text or categorical sequences, a meaningful advance over diffusion-based discrete models requiring many sampling steps.