Comparative Performance Analysis of GAN Architectures for Grayscale Image Generation: Stability, Efficiency, and Image Quality

This study presents a comparative analysis of four Generative Adversarial Network (GAN) architectures Vanilla GAN, Deep Convolutional GAN (DCGAN), Wasserstein GAN with Gradient Penalty (WGAN-GP), and Least Squares GAN (LSGAN) for image generation on grayscale datasets with varying complexity: MNIST, Fashion-MNIST, KMNIST, and EMNIST Digit. The evaluation focuses on training stability, computational efficiency, and image quality, measured by the Inception Score (IS). Experimental results demonstrate significant performance differences across datasets and hyperparameter settings. On the EMNIST dataset, LSGAN achieved the highest IS of 6.346 at epoch 950, while WGAN-GP exhibited the shortest training time (1041.46 seconds). Conversely, DCGAN showed superior visual quality for MNIST-generated images despite requiring longer training (3175.45 seconds). For Fashion-MNIST, LSGAN again outperformed others with an IS of 3.769, whereas WGAN-GP maintained the fastest training speed (446.24 seconds). Notably, hyperparameter tuning yielded the highest Inception Scores is β = 0.5 and β = 0.9 for the MNIST, KMNIST, and EMNIST datasets, indicating enhanced stability across those domains. However, for the Fashion-MNIST dataset, the optimal configuration was found at β = 0.5 and β = 0.999, suggesting that this variant of the momentum term is better suited to its specific data characteristics. These results highlight the importance of tailoring optimizer settings to the underlying dataset, especially when balancing generative quality and training consistency across diverse low-resolution image domains.