Traditional methods of evaluating stress in mechanical engineering and materials science often involve computationally intensive processes or face limitations in applicability due to their reliance on differential equations and discretization methods. Consequently, these approaches may not be well-suited for fast or real-time applications. To overcome these challenges, Deep Learning techniques have arisen as a promising alternative, delivering fast results. This work introduces StressGNN, a Graph Neural Network (GNN) model designed to predict the 2D von Mises stress distribution in solid mechanics. This model, named StressGNN, is trained to generate stress distributions based on input geometries, loads, and boundary conditions. Results indicate that, across various and complex cases of geometries, loads, and boundary conditions, our model consistently predicts stress distributions more accurately than a baseline conditional generative adversarial model. The StressGNN, with its accurate and efficient stress prediction capabilities using GNNs, stands as a viable alternative to the Finite Element Method in structural analysis. This approach has the potential to greatly enhance the design and optimization processes for structures and materials, thereby influencing a diverse array of industries such as aerospace, automotive, and civil engineering.