Heterogeneous Image GNN: Graph-Conditioned Diffusion for Image Synthesis

TL;DR

This paper introduces Heterogeneous Image Graphs (HIG), a graph-based conditioning method for diffusion models that effectively handles complex, relational, and unstructured graph data for improved image synthesis.

Contribution

It presents a novel graph representation (HIG) and a magnitude-preserving GNN that integrate with diffusion models, enabling flexible conditioning on complex graph data.

Findings

Outperforms state-of-the-art on COCO-stuff and Visual Genome datasets.

Effectively conditions on graph attributes and relationships.

Handles variable-length and complex relational conditioning variables.

Abstract

We introduce a novel method for conditioning diffusion-based image synthesis models with heterogeneous graph data. Existing approaches typically incorporate conditioning variables directly into model architectures, either through cross-attention layers that attend to text latents or image concatenation that spatially restrict generation. However, these methods struggle to handle complex scenarios involving diverse, relational conditioning variables, which are more naturally represented as unstructured graphs. This paper presents Heterogeneous Image Graphs (HIG), a novel representation that models conditioning variables and target images as two interconnected graphs, enabling efficient handling of variable-length conditioning inputs and their relationships. We also propose a magnitude-preserving GNN that integrates the HIG into the existing EDM2 diffusion model using a ControlNet approach. Our approach improves upon the SOTA on a variety of conditioning inputs for the COCO-stuff and Visual Genome datasets, and showcases the ability to condition on graph attributes and relationships represented by edges in the HIG.