From Probabilistic Models to Transformers: The Technological Trajectory of Generative AI in Vision Tasks

With the rapid advancements in generative artificial intelligence (GAI), visual computing has witnessed transformative changes across a range of applications such as image synthesis, restoration, super-resolution, 3D reconstruction, and medical imaging. This review systematically examines the evolution of generative models, from early statistical approaches to state-of-the-art transformer-based architectures. Key models including variational autoencoders (VAE), generative adversarial networks (GAN), and diffusion models are compared in terms of their structure, training stability, image quality, and suitability for various visual tasks. In addition to technical progress, the review highlights the ethical, explainability, and safety challenges associated with GAI deployment, especially in high-stakes fields like healthcare and manufacturing. While GAI enables highly realistic and semantically meaningful image generation, challenges remain in balancing innovation with interpretability, computational efficiency, and social responsibility. The paper also acknowledges the limitations of static literature reviews in a rapidly evolving domain and calls for ongoing comparative studies and interdisciplinary collaboration to shape a responsible and sustainable future for generative AI in visual computing.