Generalized Discrete Diffusion from Snapshots

TL;DR

GDDS introduces a flexible discrete diffusion framework supporting arbitrary noising processes, improving training efficiency and generation quality over existing methods, and surpassing autoregressive models at large scales.

Contribution

It presents a unified, flexible framework for discrete diffusion modeling with a novel ELBO for efficient training and improved performance.

Findings

Outperforms existing discrete diffusion methods in efficiency and quality.

Beats autoregressive models on large-scale discrete generation tasks.

Provides a simple, probabilistic ELBO based on snapshot latents.

Abstract

We introduce Generalized Discrete Diffusion from Snapshots (GDDS), a unified framework for discrete diffusion modeling that supports arbitrary noising processes over large discrete state spaces. Our formulation encompasses all existing discrete diffusion approaches, while allowing significantly greater flexibility in the choice of corruption dynamics. The forward noising process relies on uniformization and enables fast arbitrary corruption. For the reverse process, we derive a simple evidence lower bound (ELBO) based on snapshot latents, instead of the entire noising path, that allows efficient training of standard generative modeling architectures with clear probabilistic interpretation. Our experiments on large-vocabulary discrete generation tasks suggest that the proposed framework outperforms existing discrete diffusion methods in terms of training efficiency and generation quality, and beats autoregressive models for the first time at this scale. We provide the code along with a blog post on the project page : \href{https://oussamazekri.fr/gdds}{https://oussamazekri.fr/gdds}.