Discrete Diffusion Models: A Unified Framework from Tokenization to Generation

This paper proposes a unified view of discrete denoising diffusion models (DDMs) centered on how their discrete state space is constructed. It treats transition-matrix, masking or absorbing-state, and score/ratio-based formulations as different points in one design space. The framework connects tokenization, vocabulary topology, and domain-specific alphabets to trade-offs in training objectives, inference algorithms, scaling behavior, systems optimization, and evaluation. The supplied abstract emphasizes conceptual unification and future research directions; it does not report specific benchmark results or new quantitative experiments.