Graph Transformer GANs With Graph Masked Modeling for Architectural Layout Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Architectural layout generation is complex, requiring modeling of intricate relationships between components.
• Existing methods struggle with end-to-end learning of graph-constrained layouts.

Purpose of the Study:

• To introduce a novel graph Transformer generative adversarial network (GTGAN) for effective graph node relation learning.
• To enhance architectural layout generation through improved modeling of local and global interactions.

Main Methods:

• Developed a graph Transformer encoder combining graph convolutions and self-attentions for local and global node interactions.
• Introduced Connected Node Attention (CNA) and Non-Connected Node Attention (NNA) to capture graph relations.
• Proposed a node classification discriminator and a graph-based cycle-consistency loss for spatial relationship preservation.
• Implemented a self-guided pre-training method with high-ratio node/edge masking for graph representation learning.

Main Results:

• Achieved state-of-the-art performance on house layout generation, house roof generation, and building layout generation tasks.
• Demonstrated significant improvements in both quantitative scores and visual realism across three public datasets.
• The proposed GTGAN effectively models complex graph node relations for challenging layout generation.

Conclusions:

• The novel GTGAN framework significantly advances graph-constrained architectural layout generation.
• The integrated attention mechanisms and pre-training strategy enhance model learning and generation quality.
• This approach sets a new benchmark for realistic and accurate architectural design generation.