This paper presents Discrete Spatial Diffusion, a generative model for discrete-state data that ensures mass conservation, enabling applications in scientific domains like materials science, also demonstrating results on popular image benchmarks.