Intelligent Mobile AI-Generated Content Services via Interactive Prompt Engineering and Dynamic Service Provisioning

TL;DR

The paper tackles the infeasibility of on-device large AIGC by enabling edge-enabled, intelligent mobile AIGC through two core ideas: interactive prompt engineering to elevate generation quality and diffusion-augmented dynamic service provisioning to adapt resource allocation to heterogeneous tasks. It introduces an LLM-driven prompt corpus generator and IRL-based policy imitation to select optimal prompt strategies, paired with a diffusion-enhanced DRL agent ($D^3PG$) that optimizes the number of inferences $N_i$ and transmission power $P_i$ to maximize QoE under power and latency constraints. The approach yields significant gains, including a 6.3x improvement in single-round generation success and a 67.8% QoE boost, with D$^3$PG outperforming PPO and SAC baselines in converged utility and coverage. By tightly integrating prompt design, human-aligned quality assessment, and diffusion-based policy learning, the framework offers a unified mechanism to deliver high-quality, resource-efficient mobile AIGC across diverse edge scenarios.

Abstract

Due to massive computational demands of large generative models, AI-Generated Content (AIGC) can organize collaborative Mobile AIGC Service Providers (MASPs) at network edges to provide ubiquitous and customized content generation for resource-constrained users. However, such a paradigm faces two significant challenges: 1) raw prompts (i.e., the task description from users) often lead to poor generation quality due to users' lack of experience with specific AIGC models, and 2) static service provisioning fails to efficiently utilize computational and communication resources given the heterogeneity of AIGC tasks. To address these challenges, we propose an intelligent mobile AIGC service scheme. Firstly, we develop an interactive prompt engineering mechanism that leverages a Large Language Model (LLM) to generate customized prompt corpora and employs Inverse Reinforcement Learning (IRL) for policy imitation through small-scale expert demonstrations. Secondly, we formulate a dynamic mobile AIGC service provisioning problem that jointly optimizes the number of inference trials and transmission power allocation. Then, we propose the Diffusion-Enhanced Deep Deterministic Policy Gradient (D3PG) algorithm to solve the problem. By incorporating the diffusion process into Deep Reinforcement Learning (DRL) architecture, the environment exploration capability can be improved, thus adapting to varying mobile AIGC scenarios. Extensive experimental results demonstrate that our prompt engineering approach improves single-round generation success probability by 6.3 times, while D3PG increases the user service experience by 67.8% compared to baseline DRL approaches.

Figures (15)

• Figure 1: The heterogeneity of AIGC tasks. We can obverse that generating an image of a city is much more difficult than that of an apple since more complicated objects and compositions are required. Therefore, more inference trials should be allocated. Moreover, complex images accommodate more information (e.g., edges and visual signals) Complexity. Hence, they are more sensitive to transmission loss and require more transmission power.
• Figure 2: Top: A typical mobile AIGC service scheme (e.g., Stable Diffusion). Bottom: The proposed intelligent mobile AIGC scheme. Note that the orange and blue lines correspond to service configuration and operation stages, respectively.
• Figure 3: The illustration of prompt engineering strategy and policy. We can observe that for one raw prompt, different prompt engineering strategies lead to diverse optimized prompts and generated images. Therefore, the prompt engineering policy $\pi_\omega^{(p)}$ aims to select the optimal prompt engineering strategy dynamically.
• Figure 4: The workflow for training prompt engineering policy $\pi_\omega^{(p)}$. First, the prompt corpus corresponding to each demonstration prompt is generated by an LLM. Then, different prompt engineering strategies are performed, and the demonstration dataset is constructed. From the demonstration dataset, the expert policy can be acquired (The expert policy is the one that always selects the strategy that leads to the optimal generation quality). Finally, an IRL framework is utilized for policy imitation.
• Figure 5: The impact of two decision variables of dynamic mobile AIGC service provisioning on user received images.