This paper conducts a large-scale empirical study to evaluate the performance of various Generative Adversarial Network (GAN) models and their evaluation metrics. The authors find that most models can achieve similar scores with sufficient hyperparameter optimization and random restarts, suggesting that improvements may come from increased computational budgets rather than fundamental algorithmic changes. They propose new datasets to compute precision and recall, which are more robust to mode dropping and encoding network choices. The study also highlights the limitations of current metrics, such as their inability to detect overfitting. The main contributions include a fair comparison of state-of-the-art GANs, evidence that a summary of results distribution is necessary for meaningful comparisons, and the introduction of a series of tasks for which precision and recall can be computed. The authors conclude that future GAN research should focus on more systematic and objective evaluation procedures.This paper conducts a large-scale empirical study to evaluate the performance of various Generative Adversarial Network (GAN) models and their evaluation metrics. The authors find that most models can achieve similar scores with sufficient hyperparameter optimization and random restarts, suggesting that improvements may come from increased computational budgets rather than fundamental algorithmic changes. They propose new datasets to compute precision and recall, which are more robust to mode dropping and encoding network choices. The study also highlights the limitations of current metrics, such as their inability to detect overfitting. The main contributions include a fair comparison of state-of-the-art GANs, evidence that a summary of results distribution is necessary for meaningful comparisons, and the introduction of a series of tasks for which precision and recall can be computed. The authors conclude that future GAN research should focus on more systematic and objective evaluation procedures.