Automating Software Diagram Generation with Large Language Models

Recent advances in Large Language Models (LLMs) have created new opportunities for automated and efficient diagram generation. This study introduces DiagramAI, a tool designed for AI-driven diagram generation and editing. To evaluate its feasibility, we assess the capabilities of seven LLMs in generating five types of software diagrams: use case, class, architectural, sequence, and C4 context diagrams. We compare two-generation approaches, direct Mermaid code output, and structured YAML-based representations, analyzing their syntactic correctness, structural consistency, and adherence to diagramming conventions. The results reveal significant variations in model performance, with some LLMs demonstrating strong syntax adherence, while others struggle with diagram structure and semantic accuracy.