<![CDATA[The advancement of artificial intelligence technology, particularly deep learning, presents significant potential in facial image processing. Generative Adversarial Networks (GANs), a type of deep learning model, have demonstrated remarkable capabilities in generating high-quality synthetic images through a competitive training process between a generator, which creates new data, and a discriminator, which evaluates its authenticity. However, the use of public facial datasets such as CelebA and FFHQ faces limitations in representing global demographic diversity and raises privacy concerns. This study aims to generate realistic synthetic facial datasets using the StyleGAN2-ADA architecture, a specialized variant of GAN, with two training approaches: training from scratch on two types of datasets (private and public), each containing 480 images. The public dataset used is FFHQ (Flickr-Faces-HQ), known for its broader facial variation and high-quality images. Evaluation is conducted using the Frechet Inception Distance (FID), a metric that assesses image quality by comparing the feature distributions of real and generated images. Results indicate that training from scratch with the public dataset (FFHQ) using a batch size of 16 and a learning rate of 0.0025 achieves an FID score of 85.67 and performance of 86.46% at Tick 100, whereas the private dataset, under the same conditions, results in an FID score of 98.59 with a performance of 18.54%.. The training from scratch approach with the public dataset proves more effective in generating high-quality synthetic facial images compared to the private dataset. In conclusion, this approach supports the optimal generation of realistic synthetic facial data.]]>
