DiPhon: Diffusion on Graphons for Scalable Graph Generation

2026-07-08 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

DiPhon is a novel diffusion framework for size-scalable graph generation, addressing the challenge of scaling diffusion models to large graphs, especially in dense-graph settings. It utilizes graphons, size-agnostic limit objects of dense graph sequences, to study structural graph statistics across node-size scales. DiPhon formulates a continuous diffusion process on the graphon space via a Jacobi stochastic differential equation (SDE), then proposes a discretized graph-level process mimicking these dynamics on finite graphs. Its reverse-time process uses a tractable marginal score, estimated via graph denoising. DiPhon is proven to exactly match the first moment and approximate the second moment of marginal distributions from the continuous graphon process. This ensures DiPhon inherits key size-agnostic statistical properties, enabling training on small graphs and generating progressively larger graphs at inference time without retraining, while preserving core topological properties. The paper was published on 2026-07-08.

Key takeaway

For Machine Learning Engineers scaling graph generation models for molecular design or large-scale network analysis, DiPhon offers a critical advancement. You can now train models on smaller, more manageable datasets and reliably generate significantly larger graphs without retraining. This approach provides a principled route to overcome current scalability limitations. It allows you to focus resources on initial model training rather than iterative scaling for different graph sizes.

Key insights

DiPhon enables scalable graph generation by applying diffusion models to graphons, preserving topological properties across graph sizes.

Principles

• Graphons provide a size-agnostic limit for dense graph sequences.
• Continuous diffusion on graphon space can be discretized for finite graphs.
• Tractable marginal scores enable reverse-time processes for graph generation.

Method

DiPhon formulates a continuous diffusion process on graphon space via a Jacobi SDE, then discretizes it. The reverse-time process estimates a tractable marginal score from data via graph denoising to generate samples.

In practice

• Train on small graphs to generate larger graphs without retraining.
• Preserve core topological properties during size-scalable generation.

Topics

• DiPhon
• Graph Generation
• Diffusion Models
• Graphons
• Jacobi SDE
• Scalable AI

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Diffusion enabled Optimal Transport distances for graph matching — DsrFGW unifies node features and structural connectivity via optimal transport, enhanced by diffusion for robust graph comparison.
• From Uniform to Learned Graph Priors: Diffusion for Structure Discovery — Diffusion-parameterized adaptive prior calibrates latent graph distributions for reliable structure discovery.
• Zero-Shot Size Transfer for Neural ODEs on Sparse Random Graphs: Graphon Limits and Adjoint Convergence — Graph Neural Differential Equations (GNDEs) exhibit zero-shot size transfer, enabling training on small graphs for accurate deployment on larger, similar graphs.
• Generative Diffusion Models of Stochastic Graph Signals — A U-GNN-powered denoising diffusion framework unifies conditional graph signal generative modeling for diverse applications, sampling from conditional laws.
• Information-geometric adaptive sampling for graph diffusion — DVS adaptively adjusts graph diffusion sampling steps based on information-geometric curvature for improved fidelity and efficiency.
• Nemotron-TwoTower: Diffusion Language Modeling with Pretrained Autoregressive Context — Decoupling context and denoising in diffusion models significantly boosts generation throughput while maintaining quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.