A Comparative Study of Generative AI Models in Semantic Communication Networks

Abstract

Communication networks are gradually shifting from traditional bit-oriented transmission toward semantic communication, where the emphasis is placed on delivering meaningful and context-aware information rather than transmitting complete data packets. This transition has encouraged the adoption of Generative Artificial Intelligence (Generative AI) models to improve semantic understanding, intelligent encoding, and adaptive information delivery. The present study provides a comparative examination of prominent generative models, including Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), Transformer architectures, Large Language Models (LLMs), and Diffusion Models within semantic communication networks. The comparison focuses on important performance aspects such as semantic accuracy, computational efficiency, latency, bandwidth optimization, contextual understanding, and adaptability in dynamic communication environments. In addition, the study discusses the applicability of these models in emerging technologies such as 5G, 6G, Internet of Things (IoT), edge intelligence, and autonomous communication systems. Particular attention is given to practical concerns, including privacy preservation, semantic ambiguity, computational cost, scalability, and security challenges associated with generative AI integration. Based on the comparative analysis, Transformer-based and large language models demonstrate stronger capabilities in semantic representation and intelligent communication, whereas hybrid generative frameworks show promising potential for improving efficiency and contextual accuracy in future communication ecosystems. The study concludes that Generative AI can significantly contribute to the development of intelligent, adaptive, and semantically aware communication infrastructures.