Joint Communication and Computation Scheduling for MEC-Enabled AIGC Services Based on Generative Diffusion Model

Artificial Intelligence-Generated Content (AIGC) based on Generative Diffusion Model (GDM) has emerged as a promising paradigm of content generation, revolutionizing the creation of diverse contents and driving significant technological advancements. Due to the low latency requirements of AIGC services, mobile edge computing (MEC) has become a crucial enabling technology for these services. In this work, we consider an MEC-enabled GDM-based AIGC network, which consists of multiple GDMs with varying sizes and capabilities deployed on edge computing servers (ES), and multiple mobile users (UEs) with diverse latency and accuracy requirements requesting AIGC services from ES through wireless access points (APs). In such a scenario, we are interested in the joint communication and computation scheduling problem for UEs, which involves selecting the appropriate APs (along with the communication bandwidth allocation) and the appropriate ES (together with the computation resource allocation and model inference optimization) for UEs, considering both the UEs' heterogeneous requirements and the GDMs' heterogeneous capabilities. To address the problem in a practical scenario with decentralized, autonomous, and self-interested UEs, we formulate a non-cooperative game, called the Joint User Association and Computation Offloading (JUACO) game, where each UE acts as a game player, selecting the best AP (for communication) as well as the best ES and the best GDM model inference step (for computation), aiming to minimize the inference time while meeting the specified inference accuracy requirement. We prove that the proposed JUACO game is a potential game, thus guaranteeing the existence of Nash equilibrium (NE) and the convergence of simple best response-based distributed algorithms to NE. Simulation results demonstrate that the proposed JDACO game approach significantly reduces the inference time and meets the required accuracy compared to traditional methods, validating the effectiveness and practicality of the game-theoretic approach in real-world scenarios.