Discrete Diffusion

Discrete diffusion models are a class of generative models designed to create samples from discrete data, such as text, graphs, or sequences of actions, by reversing a noise diffusion process. Current research focuses on improving sampling efficiency by addressing limitations like slow convergence and the inability to capture inter-variable dependencies, often through incorporating copula models or employing absorbing processes. These advancements are enabling applications across diverse fields, including natural language processing, computer vision (e.g., image inpainting, layout generation), and robotics, where they offer improved sample quality and controllability compared to previous methods.