Optimizing AIGC Services by Prompt Engineering and Edge Computing: A Generative Diffusion Model-Based Contract Theory Approach

Abstract

The development of Generative AI (GAI) and AI-generated content (AIGC) has been significantly improved by pretrained foundation models and prompt-based methods. To boost the quality and reduce the latency of AIGC generation, prompt engineering and edge computing are introduced, demanding a multi-dimensional resource allocation approach. Thus, we use the generative diffusion model (GDM) and contract theory to design a two-stage, multi-dimensional resource allocation framework. In the first stage, we employ an approximation approach to quantitatively assess the relationship between the level of prompt optimization, the number of diffusion denoising steps, and the quality of AIGC generation. Based on the quality function, we formulate models for the utilities of an AI-generated content Service Provider (ASP) and users, leading to a non-convex quality-based contract problem optimizing the level of prompt optimization and the number of diffusion denoising steps. To address the time-consuming process of solving the non-convex problem due to variable cost of the ASP and gain preferences of the users, a GDM-based scheme is proposed to optimize quality-based contract items. In the second stage, for each group of users who choose the same quality-based contract items, a non-convex latency-based contract problem optimizing the CPU cycle frequency and network transmission rate is formulated, then the GDM-based scheme is also applied to find the optimal latency-based contract items. Numerical results show that the proposed GDM-based contract generation scheme is very advantageous in improving the quality of AIGC generation and decreasing the latency of AIGC generation, compared to other standard schemes.