Generative AI encompasses a range of model architectures, each with its own design philosophy, strengths, and limitations. Choosing the right architecture is a critical step in building systems that meet the specific needs of a given application. This chapter examines four of the most influential approaches: transformer-based models for handling complex sequences, generative adversarial networks for high-fidelity synthesis through adversarial training, variational autoencoders for learning structured probabilistic representations, and diffusion models for gradual denoising and high-quality sample generation. By exploring the principles, advantages, and trade-offs of each, this chapter provides the knowledge needed to align architectural choice with project objectives, performance requirements, and resource constraints.