A Diffusion-based DRL Flow Scheduling for AIGC Services in Edge Network

Abstract

With the convergence of computing and networking, Artificial Intelligence Generated Content (AIGC) services are increasingly deployed at the network edge to support low-latency and bandwidth-intensive applications. However, constrained edge bandwidth and dynamically varying traffic demands present significant challenges for efficient flow scheduling. To address these issues, this paper proposes a diffusion-based deep reinforcement learning (DRL) scheduling framework that jointly accounts for flow latency sensitivity and bandwidth constraints. By leveraging the generative exploration capability of the diffusion model into the policy optimization process, the proposed method enhances exploration efficiency and mitigates the risk of convergence to local optima. Extensive experiments demonstrate that the proposed approach significantly reduces average flow latency and improves flow completion rates compared with baseline algorithms, validating its effectiveness for adaptive and intelligent flow scheduling in edge AIGC environments.