GO-WGAN and Graph Transformer-Based Framework for Real-Time Mobile UI Layout Generation and Estimation

Abstract

Designing user interface (UI) layouts for mobile applications is a labor-intensive task that demands considerable manual effort and domain expertise. To address this, we present FastLayout, a deep learning framework that fuses generative and relational modeling for dynamic UI layout generation and estimation. At its core, FastLayout integrates WGAN-GP (adapted here as GO-WGAN) to generate diverse, high-quality synthetic layout samples, alleviating data scarcity. The Branching Hybrid Attention Mechanism (BHAM) enhances the CNN backbone by improving feature extraction and reducing task conflicts, while the Graph Transformer explicitly models spatial and relational dependencies between UI components. Our experiments use an internal dataset of 3,275 annotated mobile UI screens across 15 layout categories and the public RICO benchmark, with training conducted in two phases using AdamW optimization and early stopping to ensure fairness and stability. Experimental evaluations show that FastLayout achieves an Intersection over Union (IoU) of 83.52%, a Pixel-wise Error (PE) of 4.97%, an Edge Error (EE) of 4.91 units, and a Root Mean Square Error (RMSE) of 0.2753, outperforming state-of-the-art baselines in both accuracy and efficiency. solution for intelligent UI automation in modern software development.