.png){kind=icon}
.png){kind=icon}
Return
Integrating Plan-VAE, genetic algorithm, and diffusion transformers for automated layout optimization and synthesis
.png){kind=icon}
PRE
AI
DOI：10.1007/s10586-026-06428-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The conceptual design phase of architectural projects is characterized by high impact but minimal data availability, making it a critical yet challenging candidate for automation. While generative design tools are increasingly prevalent, they often falter during early-stage planning due to a deficiency in precise technical inputs. This research introduces a multi-stage hybrid framework designed to bridge this gap by integrating a Transformer-based Plan-Variational Autoencoder (Plan-VAE), a multi-objective Genetic Algorithm (GA), and a Diffusion Transformer (DiT). In the proposed pipeline, the GA explores a regularized 24-dimensional latent space within the Plan-VAE to identify optimal topological configurations based on objective constraints, including circulation efficiency, programmatic adjacency, and area ratios. These optimized spatial plans are subsequently passed to the DiT as conditioning inputs, which employs iterative denoising to synthesize architecturally realistic and functionally coherent geometries. Experimental results show that the DiT visualization component achieves strong distributional fidelity on the curated test set (FID 1.4873), while the full pipeline improves the normalized Fitness Delta ( $$\Delta F$$ ), a composite spatial optimization metric aggregating penalties for adjacency, area fidelity, and spatial separation, by 24.43% relative to the human-designed baseline. This study offers a robust methodology for automating high-fidelity floor plan synthesis during the nascent stages of architectural development.
Keywords:
Architecture
Design management
Artificial intelligence
Computer vision
Genetic algorithm
Journal
C
IF:
4.1
Papers:
5.0K
Citations:
7.5K