Reinforcement Learning With LLMs Interaction for Distributed Diffusion Model Services.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Distributed Systems

Background:

• Distributed Artificial Intelligence-Generated Content (AIGC) faces challenges in Quality of Experience (QoE) and energy efficiency, especially in Generative Diffusion Model (GDM) image generation.
• Current GDM frameworks lack user-centric management for optimizing subjective experience and resource utilization.

Purpose of the Study:

• To propose a novel user-centric Interactive AI (IAI) approach for managing distributed GDM-based AIGC services.
• To enhance subjective Quality of Experience (QoE) and improve energy efficiency in AI-generated image services.
• To develop an adaptive resource allocation algorithm for dynamic wireless environments.

Main Methods:

• Restructured GDM inference allowing users with similar prompts to share denoising processes.
• Introduced Reinforcement Learning With Large Language Models Interaction (RLLI) for real-time, subjective QoE feedback using LLM-powered agents.
• Adapted the Deep Deterministic Policy Gradient (DDPG) algorithm into G-DDPG for effective resource allocation.

Main Results:

• The proposed IAI framework enables cooperative deployment and efficient GDM inference.
• RLLI effectively replicates user interactions, providing personalized QoE feedback.
• G-DDPG demonstrated a 15% improvement in total QoE compared to the standard DDPG algorithm.

Conclusions:

• The proposed IAI approach, coupled with RLLI and G-DDPG, significantly enhances QoE and resource efficiency in distributed AIGC services.
• This framework offers a promising direction for user-centric service management in generative AI.
• The findings highlight the potential of integrating LLMs and reinforcement learning for optimizing complex AI systems.