NOMA-Assisted Mobile Edge Generation (MEG): Enabling Mobile Access to Large Models

The popularity of artificial intelligence generated content (AIGC) is prompting the deployment of large language model (LLM) from cloud to edge networks, leading to mobile edge generation (MEG). Due to high latency and limited computational capabilities of mobile devices, personalized image generation for mobile healthcare and education requires edge-mobile generation paradigm. In this paper, a novel non-orthogonal multiple access (NOMA) assisted multi-user MEG framework is proposed for text-guided mobile image generation. NOMA enables concurrent access from multiple user equipments (UEs) to the edge-deployed large model, facilitating adjustable generation splitting. Specifically, the edge server (ES) partially generates the image and transmits it via downlink NOMA, while UEs complete the remaining parts using lightweight models. Both unlimited and limited energy budget scenarios are considered. 1) For unlimited energy budget, a joint generation splitting ratio and NOMA power allocation optimization problem is formulated, which minimizes the maximum (min-max) latency of UEs to ensure fairness. The closed-form globally optimal solutions based on Karush-Kuhn-Tucker (KKT) and Lambert-W theory are derived. Moreover, the superiority of MEG-NOMA over conventional MEG-orthogonal multiple access (OMA) is mathematically proved. 2) For limited energy budget, a multi-objective programming problem is formulated to minimize the latency of each UE, which leads to a user-centric latency minimization problem. The closed-form solutions of generation splitting ratio and power allocation are derived. Simulation results illustrate that the proposed MEG-NOMA outperforms the MEG-OMA in both two-user and multi-user cases. Compared to conventional MEG-OMA, the MEG-NOMA framework reduces the min-max latency and the user-centric latency by 33.01% and 9.86%, respectively.