Diffusion models have transformed generative modelling for images and text, built on a simple principle: progressively destroy structure by adding noise, then train a neural network to reverse the process. Starting from pure noise, the learned reverse process gradually constructs coherent outputs. Here we present a framework for networks that shares this core idea.                       

Instead of adding Gaussian noise to pixels, we remove edges uniformly at random until the network is empty. The reverse process then learns to reconstruct edges one at a time. Given a collection of networks (molecular structures, social networks, supply chains), the model learns their distribution and can generate new instances. This enables data augmentation, hypothesis testing, and exploration of plausible structures.