Eliminating Rasterization: Direct Vector Floor Plan Generation With DiffPlanner.

Area of Science:

• Computer Science
• Artificial Intelligence
• Computational Geometry

Background:

• Current learning-based floor plan generation methods convert vector data to raster images, leading to complexity, information loss, and scaling issues.
• Existing approaches struggle with precision and detail when generating scalable architectural designs.

Purpose of the Study:

• To introduce DiffPlanner, a novel deep learning framework for boundary-constrained floor plan generation operating entirely in vector space.
• To overcome the limitations of raster-based methods by maintaining data integrity and precision throughout the generation process.
• To enable user-controllable and high-quality floor plan and bubble diagram generation.

Main Methods:

• Developed a Transformer-based conditional diffusion model, DiffPlanner, that processes data directly in vector space.
• Integrated an alignment mechanism during training to synchronize the model's optimization with iterative design processes.
• Conducted quantitative comparisons, qualitative evaluations, ablation experiments, and perceptual studies to validate the framework.

Main Results:

• DiffPlanner successfully generates complex vector data for floor plans and bubble diagrams, maintaining precision and scalability.
• The framework demonstrates superior performance compared to state-of-the-art methods in fitting target distributions and achieving layout design.
• User-controllable generation capabilities were achieved, producing results closely matching ground truths.

Conclusions:

• DiffPlanner offers a significant advancement in floor plan generation by operating natively in vector space.
• The proposed method enhances controllability and quality, addressing key limitations of previous raster-based approaches.
• This framework provides a more efficient and precise tool for architectural design and creative stages.