This paper explores the evolution of generative artificial intelligence, focusing on four major models: Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Convolutional Neural Networks (CNNs), and Diffusion Models. These models have significantly impacted fields such as anomaly detection, image and video synthesis, and content creation across industries like augmented reality, entertainment, and healthcare. GANs, through adversarial training, have achieved high visual realism, with Inception Scores (IS) over 8.5 and Fréchet Inception Distances (FID) as low as 4.59 on datasets like CIFAR-10 and CelebA. VAEs, while generating less sharp outputs (typically FID between 35–40), excel at learning interpretable latent spaces and smooth data interpolation. CNNs, initially developed for classification tasks, have proven effective in generative domains like style transfer and image super-resolution, with models like SRGAN achieving PSNR scores above 30 dB and SSIM over 0.9 in enhancing image quality. Diffusion models, the most recent advancement, generate photorealistic and diverse outputs by reversing a noise process and have achieved state-of-the-art FID scores below 3.0 on datasets like ImageNet-64, though they require longer inference times with hundreds to thousands of sampling steps. Based on an analysis of over 30 studies from 2019 to 2024, this paper investigates these models’ theoretical foundations, architectural innovations, and comparative performance, while also highlighting emerging trends like hybrid models and quantum-enhanced frameworks, thus emphasizing their transformative role in next-generation AI applications.