Creating Realities: An In-Depth Study of AI-Driven Image Generation with Generative Adversarial Networks
Raman Dhand, Sagar Sidana, Anurag Kumar Sinha, Biresh Kumar, Shreya Kumari, G. Madhukar Rao, Petheru Raj, Ahmed Hussein Alkhayyat, S Maheswaran · 2024
The field of artificial intelligence has witnessed significant advancements with the advent of Generative Adversarial Networks, revolutionizing image generation techniques. The literature review covers seminal works and recent advancements, highlighting the evolution of image generation techniques. Taxonomy is developed to classify methods based on their underlying principles and applications across various domains including art, fashion, healthcare, and entertainment. Comparative analysis evaluates the effectiveness, efficiency, and applicability of different models, utilizing metrics corresponding Inception Score and Frechet Inception Distance. By conducting a thorough review of existing literature and technologies, we explore the evolution of GAN architectures, from their inception to the latest innovations. We examine various GAN models, including DCGAN, StyleGAN, and CycleGAN, assessing their effectiveness in generating high-fidelity images across diverse applications. In addition to this comprehensive analysis, we present our own implementation of a GAN model tailored for specific image generation tasks. Our model is tested on multiple datasets to evaluate its performance, highlighting improvements in training stability and image quality. We discuss the architectural choices, optimization techniques, and training procedures employed in our implementation. Furthermore, we address the challenges faced in the field, such as mode collapse, training instability, and the need for largescale datasets. Our findings offer valuable insights into optimizing GAN performance and propose potential directions for future research. This comprehensive analysis serves as a foundational reference for researchers and practitioners aiming to leverage GANs for advanced image generation tasks, pushing the boundaries of what is visually conceivable through artificial intelligence.