GRAFT: Relational Node Fields for Latent Graph Generation

Graph generation is challenging because structural relations are inherently pairwise, leading many diffusion and flow models to maintain $\mathcal{O}(n^2)$ adjacency, edge, or pair-level states throughout generation. Latent graph generation offers a more compact alternative, but existing approaches often lack relational representations rich enough to preserve graph structure, limiting their ability to realize the scalability benefits of node-only transport as graph size grows. We propose GRAFT, a latent graph generation framework that constructs relational node fields and transports them using Flow Matching. GRAFT folds pairwise and global relational information into node representations while keeping the entire generative trajectory node-only, with self-conditioning used to improve node-field transport. GRAFT achieves competitive generation quality across diverse graph domains while substantially improving sampling efficiency. On Protein, GRAFT samples a graph in $0.290$ s with $110.6$ MiB of peak GPU memory under each method's native inference configuration, i.e. $2.1\times$ to $148.7\times$ faster and $1.9\times$ to $5.7\times$ lower in peak memory than the evaluated methods.