This study presents a deep learning framework designed to accelerate Large Eddy Simulation (LES) of airflow in urban environments. The framework leverages physics constrained conditional Generative Adversarial Networks (GANs) trained on instantaneous velocity snapshots from a synthetically generated dataset comprising 130 high-fidelity CFD simulations of simple building configurations. By learning the mapping from early-stage flow fields to their statistically steady-state counterparts, the framework allows the simulation to bypass the lengthy transient averaging phase and predict the final time-averaged fields directly. Two GAN-based architectures are explored: a conventional convolutional model operating on structured uniform grids (Grid-GAN), and a graph-based model (Graph-GAN) that utilizes Graph Neural Networks (GNNs), specifically Graph Attention Networks (GATs), to process unstructured CFD mesh data while preserving native spatial connectivity. Both approaches are integrated into a fully automated pipeline built exclusively on open-source tools, including OpenFOAM for CFD simulations, FreeCAD and ParaView for preprocessing, and PyTorch for deep learning model development and training. Results demonstrate that the proposed models can significantly reduce LES computational costs while retaining accuracy in predicting turbulent flow characteristics. The Graph-GAN, in particular, shows enhanced adaptability and physical consistency due to its ability to exploit mesh refinements in critical regions. This work lays the foundation for the development of robust, physics-informed surrogate models and supports the growing integration of deep learning with scientific simulations in fluid mechanics.