Image Generation Performance Analysis: A Comparative Study of GAN Models on MNIST Dataset

Innovations continue in the field of image processing and image synthesis, as in many other fields, with the advancement of artificial intelligence. With the use of various neural networks such as Generative Adversarial Networks (GANs) in the field of image synthesis, the development of this field has begun to accelerate. In this study, four different GAN models-Simple GAN, Deep Convolutional GAN (DCGAN), Spectrally Normalized GAN (SNGAN), and Wasserstein GAN with Gradient Penalty (WGAN-GP)-were applied to MNIST dataset and a comparative analysis of their performance in image synthesis was conducted. While SNGANs provided the most stable training process compared to other models, the Simple GAN model showed the most stable performance. The fact that the WGAN-GP model exhibited the most variable performance showed that it could potentially produce more creative results. Despite the difficulty of the discriminator, the DCGAN draws attention with its low generative loss values. As a result, the model that produces the closest images to reality is the SNGAN model with a generative loss of 0.69342 and a discriminator loss of 0.69321.