EvoDiagram: Agentic Editable Diagram Creation via Design Expertise Evolution

A research team led by Tianfu Wang and 15 collaborators has developed EvoDiagram, a novel framework that tackles the complex challenge of automated diagram creation. Unlike existing methods that either produce pixel-based outputs lacking precise control or code-based syntheses limiting intuitive flexibility, EvoDiagram introduces an intermediate canvas schema. This approach enables the generation of object-level editable diagrams, meaning users can modify individual elements after creation. The system employs a coordinated multi-agent architecture that separates semantic intent from rendering logic, effectively resolving conflicts across different design layers.

A key innovation is the design knowledge evolution mechanism, which distills execution traces from the multi-agent system into a hierarchical memory of domain guidelines. This allows the AI agents to retrieve context-aware expertise adaptively, essentially learning and applying design principles over time. The researchers also released CanvasBench, a comprehensive benchmark with both data and metrics specifically for evaluating canvas-based diagramming systems. Extensive experiments demonstrate that EvoDiagram achieves excellent performance in generating diagrams that are not only editable but also structurally consistent and aesthetically coherent, striking a balance that previous methods have struggled to achieve.