Diffusion Models for Graphs (GDSS/DiGress) apply denoising diffusion probabilistic modeling to discrete graph structures — gradually corrupting a graph into noise (random edge flips, node type randomization) in the forward process, then training a GNN to reverse the corruption step by step — producing high-quality molecular and general graph samples that outperform VAE and GAN-based generators in both sample quality and diversity.

What Are Diffusion Models for Graphs?

• Definition: Graph diffusion models adapt the DDPM (Denoising Diffusion) framework to discrete graph data. The forward process gradually destroys graph structure by independently flipping edges and randomizing node types over $T$ timesteps until the graph becomes an Erdős-Rényi random graph (pure noise). The reverse process trains a graph neural network $epsilon_ heta(G_t, t)$ to predict the clean graph $G_0$ from the noisy graph $G_t$, enabling iterative denoising from random noise to a valid graph.
• GDSS (Graph Diffusion via the System of SDEs): Operates in continuous state space — node positions and features are continuous variables that undergo Gaussian diffusion, and the score function $\nabla_G log p_t(G_t)$ is learned via a GNN. GDSS handles both the adjacency structure and node features through a coupled system of stochastic differential equations.
• DiGress (Discrete Denoising Diffusion): Operates in discrete state space — edges have discrete types (no bond, single, double, triple) and nodes have discrete atom types. The forward process replaces edge/node types with random categories according to a transition matrix, and the reverse process predicts the clean categorical distributions. DiGress achieves state-of-the-art molecular generation quality.

Why Graph Diffusion Models Matter

• Superior Sample Quality: Diffusion models consistently produce higher-quality molecular graphs than VAEs (which suffer from posterior collapse and blurry outputs) and GANs (which suffer from mode collapse and training instability). The iterative refinement process allows the model to correct errors gradually, producing molecules with better validity, uniqueness, and novelty metrics.
• No Mode Collapse: Unlike GANs, diffusion models do not suffer from mode collapse — the training objective (denoising score matching) is a simple regression loss that covers the full data distribution uniformly. This means diffusion-generated molecules exhibit high diversity, covering many structural families rather than repeatedly producing a few high-reward scaffolds.
• Conditional Generation: Graph diffusion models support flexible conditioning — generating molecules with specific properties by guiding the reverse diffusion process using a property predictor (classifier guidance) or by training a conditional denoising network (classifier-free guidance). This enables property-targeted molecular design without modifying the base architecture.
• scalability: DiGress and related methods scale to graphs with hundreds of nodes — significantly larger than GraphVAE (~40 nodes) or MolGAN (~9 atoms), making them applicable to drug-sized molecules, polymers, and material structures that one-shot generation methods cannot handle.

Graph Diffusion Model Variants

Diffusion Models for Graphs are structural denoising — sculpting valid molecular and network structures from random noise through iterative refinement, achieving the same quality revolution in graph generation that diffusion models brought to image synthesis.