This paper investigates the application of Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) for generating synthetic urban layouts using Open Street Map (OSM) tile datasets from Gurugram, India, and California, USA. We trained the WGAN-GP model on a dataset of 1000-2000 RGB tiles (256x256 pixels) over 50-100 epochs, constrained by the 12-hour runtime limit of Google Colab’s free tier with a T4 GPU. Initial outputs exhibit noisy, pixilated patterns lacking distinct urban features such as roads, residential blocks, or green spaces, primarily due to limited dataset size, insufficient training duration, and hyper parameter instability. Training logs reveal fluctuating losses (e.g., Generator Loss: - 25.9219, Discriminator Loss: -16.3594 at Epoch 50), indicating training imbalance. Visualizations of generated layouts for Gurugram and California highlight slight diversity improvements in the latter but persistent abstractness. We propose enhancements including scaling the dataset to 5000 tiles, extending training to 100+ epochs using Colab Pro or local GPUs, and optimizing hyperparameters such as discriminator steps (disc_steps) and gradient penalty weight (gp. weight). This study establishes a foundation for scalable urban Layout synthesis, with potential applications in urban planning, traffic simulation, and generative AI research. It also underscores Google Colab’s computational constraints as a critical challenge for academic researchers, offering insights into low-cost deep learning workflows.