IoT Thrust Seminar | Reinforcement Learning with LLMs Interaction for Edge Network-enabled Distributed Diffusion Model

In the rapidly evolving field of generative artificial intelligence (GenAI) and AI-generated content (AIGC) services, generative diffusion models (GDMs) have garnered widespread attention due to their extensive application and exceptional performance, supporting a series of applications such as Stable Diffusion, Sora, and others. Despite their success, the deployment of GDMs faces significant challenges, particularly in aligning generated content with individual user preferences and ensuring production efficiency. This presentation introduces a novel user-centric interactive AI approach for edge network-enabled distributed GDM-based AIGC service framework, prioritizing efficient and collaborative GDM deployment. Specifically, we restructure the GDM’s inference process, i.e., the denoising chain, to enable users’ semantically similar prompts to share a portion of diffusion steps. Furthermore, to maximize the users’ subjective quality-of-experience (QoE), we present a reinforcement learning with large language models interaction (RLLI) approach, which utilizes large language model (LLM)-empowered generative agents to simulate user feedback, providing real-time and subjective QoE feedback that reflects the spectrum of user personalities. In conclusion, this presentation seeks to explore the reciprocal relationship between "GenAI for Network" and "Network for GenAI," aiming to achieve more efficient, intelligent, and sustainable next-generation network services.