Contract-Inspired Contest Theory for Controllable Image Generation in Mobile Edge Metaverse

TL;DR

A novel framework that integrates contract-inspired contest theory, Deep Reinforcement Learning (DRL), and GDMs to optimize image generation in these resource-constrained environments is proposed, offering enhanced performance and efficiency in creating immersive virtual environments.

Abstract

The rapid advancement of immersive technologies has propelled the development of the Metaverse, where the convergence of virtual and physical realities necessitates the generation of high-quality, photorealistic images to enhance user experience. However, generating these images, especially through Generative Diffusion Models (GDMs), in mobile edge computing environments presents significant challenges due to the limited computing resources of edge devices and the dynamic nature of wireless networks. This paper proposes a novel framework that integrates contract-inspired contest theory, Deep Reinforcement Learning (DRL), and GDMs to optimize image generation in these resource-constrained environments. The framework addresses the critical challenges of resource allocation and semantic data transmission quality by incentivizing edge devices to efficiently transmit high-quality semantic data, which is essential for creating realistic and immersive images. The use of contest and contract theory ensures that edge devices are motivated to allocate resources effectively, while DRL dynamically adjusts to network conditions, optimizing the overall image generation process. Experimental results demonstrate that the proposed approach not only improves the quality of generated images but also achieves superior convergence speed and stability compared to traditional methods. This makes the framework particularly effective for optimizing complex resource allocation tasks in mobile edge Metaverse applications, offering enhanced performance and efficiency in creating immersive virtual environments.

Figures (10)

• Figure 1: System components and data flow for the mobile edge immersive Metaverse image generation. The data flow in the system model is in the following steps: (a) User devices capture input images and upload them to an edge server for semantic extraction. (b) The edge server extracts semantics for different tasks and transmits the semantics to a generation server. Before transmission, semantics are compressed in different level to be transmitted at the same time. (c) The generation server receives and recovers the semantics, generates the images, and then sends them back to the users.
• Figure 2: Illustration of the relationship between compression level and the corresponding compression loss. As the compression level $Z\left(D\left(P_i\right)\right)$ increases, the width $w$ and height $h$ of the semantic data are downscaled, leading to a decrease in the amount of transmitted information. However, this results in a loss of detail in the semantic data.
• Figure 3: System model for the proposed mobile edge immersive Metaverse image generation framework, detailing the flow of information and relationships between each component. User devices capture input images, which are transmitted to the edge server for semantic extraction. The edge server extracts various types of semantic information and forwards them to the generation server, which then generates high-quality images. The edge server runs a contest-based mechanism to incentivize the semantic transfer tasks by allocating transmit power based on their contributions. Meanwhile, the generation server sets a payment plan to effectively adjust the overall performance.
• Figure 4: Illustration of the forward and reverse diffusion processes: The forward diffusion process introduces gaussian noise to the current training data. In contrast, the reverse diffusion process, known as "denoising," focuses on reconstructing the original data or target data under conditions.
• Figure 5: Image quality as a function of different levels of semantic compression for various types of semantic inputs (depth map, segmentation, pose estimation, and Canny edge detection). The results indicate a general decrease in image quality with increased compression, with depth maps showing an anomaly at certain levels.