DAG-Diff: Diffusion Models for Directed Acyclic Graph Generation

Md Faizal Karim¹    Vignesh Vembar¹    Keshab Patra¹    Gaurav Singh²    K Madhava Krishna¹   

¹ Robotics Research Center, IIIT Hyderabad, India    ² Brown University   

Diffusion models have demonstrated strong performance in generative modeling but are primarily designed for continuous domains. In this work, we propose DAG-Diff, a diffusion-based framework for generating structured data in the form of Directed Acyclic Graphs (DAGs). Our approach enforces structural constraints while maintaining expressive generative capabilities. DAG-Diff enables efficient learning of complex dependencies and produces valid graph structures during sampling. We evaluate our method on structured prediction tasks and demonstrate improvements over existing graph generative models in both quality and scalability.