Directly From Alpha to Omega: Controllable End-to-End Vector Floor Plan Generation.

Area of Science:

• Artificial Intelligence
• Computer-Aided Design (CAD)
• Computational Geometry

Background:

• Automated floor plan generation seeks to create residential layouts by optimizing topological, geometric, and aesthetic factors within defined boundaries.
• Current multi-step AI methods often lack flexibility, leading to suboptimal or unreasonable outputs when encountering data outside predefined workflows, unlike human designers.

Purpose of the Study:

• To introduce CE2EPlan, a controllable, end-to-end diffusion model designed to overcome the limitations of existing multi-step approaches in automated floor plan generation.
• To enable AI models to learn floor plan design directly from data, capturing diverse solution paths without predefined restrictions.

Main Methods:

• Developed CE2EPlan, a controllable, end-to-end topology- and geometry-enhanced diffusion model.
• The model learns to generate complete floor plan layouts directly from input boundaries, bypassing intermediate representations and multi-step pipelines.

Main Results:

• CE2EPlan demonstrated superior performance compared to existing multi-step pipeline methods in extensive experiments.
• The proposed method achieved higher-quality results, enhanced user control, and greater output diversity.

Conclusions:

• CE2EPlan offers a more versatile and adaptable approach to AI-driven floor plan generation, mimicking the flexibility of human designers.
• The end-to-end diffusion model represents a significant advancement in creating realistic and diverse residential layouts with improved AI design tools.