Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mobile Edge Generation: A New Era to 6G

Ruikang Zhong, Xidong Mu, Yimeng Zhang, Mona Jabor, Yuanwei Liu

TL;DR

MEG tackles the growing traffic and latency of centralized GAI in next-generation networks by distributing generative AI workloads across edge servers and user devices, forming both single-ES and multi-ES deployment frameworks. It introduces seed-based and sketch-based generation schemes and four operating protocols (UIEG, EIUG, CIAG, ESUC) for single-ES, plus parallel (UIDG, DIUG, DSUC) and cooperative (UIDCG, DCSUC) protocols for multi-ES, enabling efficient transmission via seeds or sketches and allowing joint training and model management considerations. A case study using a Latent Diffusion Model demonstrates that seeds/ sketches-based MEG can achieve high-quality generation under very low signal-to-noise ratios and substantially reduce network bandwidth versus centralized GAI. The findings suggest MEG as a practical pathway to scale GAI services in 6G networks, with open challenges in distributed model management, security/privacy, and resource coordination.

Abstract

A conception of mobile edge generation (MEG) is proposed, where generative artificial intelligence (GAI) models are distributed at edge servers (ESs) and user equipment (UE), enabling joint execution of generation tasks. Various distributed deployment schemes of the GAI model are proposed to alleviate the immense network load and long user queuing times for accessing GAI models. Two MEG frameworks are proposed, namely the single-ES framework and the multi-ESs framework. 1) A one-to-one joint generation framework between an ES and a UE is proposed, including four specific single-ES MEG protocols. These protocols allow distributed GAI models to transmit seeds or sketches for delivering information efficiently. 2) Several protocols are proposed for multi-ESs MEG, which enable multiple ESs to perform the generation task cooperatively or in parallel. Finally, a case study of a text-guided-image-to-image generation is provided, where a latent diffusion model is distributed at an ES and a UE. The simulation results demonstrate that the proposed protocols are able to generate high-quality images at extremely low signal-to-noise ratios. The proposed protocols can significantly reduce the communication overhead compared to the centralized model.

Mobile Edge Generation: A New Era to 6G

TL;DR

MEG tackles the growing traffic and latency of centralized GAI in next-generation networks by distributing generative AI workloads across edge servers and user devices, forming both single-ES and multi-ES deployment frameworks. It introduces seed-based and sketch-based generation schemes and four operating protocols (UIEG, EIUG, CIAG, ESUC) for single-ES, plus parallel (UIDG, DIUG, DSUC) and cooperative (UIDCG, DCSUC) protocols for multi-ES, enabling efficient transmission via seeds or sketches and allowing joint training and model management considerations. A case study using a Latent Diffusion Model demonstrates that seeds/ sketches-based MEG can achieve high-quality generation under very low signal-to-noise ratios and substantially reduce network bandwidth versus centralized GAI. The findings suggest MEG as a practical pathway to scale GAI services in 6G networks, with open challenges in distributed model management, security/privacy, and resource coordination.

Abstract

A conception of mobile edge generation (MEG) is proposed, where generative artificial intelligence (GAI) models are distributed at edge servers (ESs) and user equipment (UE), enabling joint execution of generation tasks. Various distributed deployment schemes of the GAI model are proposed to alleviate the immense network load and long user queuing times for accessing GAI models. Two MEG frameworks are proposed, namely the single-ES framework and the multi-ESs framework. 1) A one-to-one joint generation framework between an ES and a UE is proposed, including four specific single-ES MEG protocols. These protocols allow distributed GAI models to transmit seeds or sketches for delivering information efficiently. 2) Several protocols are proposed for multi-ESs MEG, which enable multiple ESs to perform the generation task cooperatively or in parallel. Finally, a case study of a text-guided-image-to-image generation is provided, where a latent diffusion model is distributed at an ES and a UE. The simulation results demonstrate that the proposed protocols are able to generate high-quality images at extremely low signal-to-noise ratios. The proposed protocols can significantly reduce the communication overhead compared to the centralized model.
Paper Structure (30 sections, 6 figures)

This paper contains 30 sections, 6 figures.

Table of Contents

  1. Introduction
  2. Background of Generative AI
  3. Motivation of MEG
  4. Contributions
  5. Single-ES MEG Model
  6. Network Model
  7. Communication Schemes for MEG
  8. Seed-based Generation Scheme
  9. Sketch-based Generation Scheme
  10. Operating Protocols
  11. UE inference and Edge generation (UIEG)
  12. Edge inference and UE generation (EIUG)
  13. Cooperative inference and Generation (CIAG)
  14. Edge sketch and UE Complete (ESUC)
  15. Multi-ES MEG
  16. ...and 15 more sections

Figures (6)

  • Figure 1: The working flow of the seed-based and the sketch-based MEG.
  • Figure 2: MEG Protocols. (a) Local generation; (b) Centralized generation; (c) UIBG protocol; (d) BIUG protocol; (e) CIAG protocol; (f) ESUC protocol.
  • Figure 3: Multi-ES MEG Model (a) UIDG protocol; (b) DIUG protocol; (c) DSUC protocol; (d) UIDCG protocol; (e) DCSUC protocol.
  • Figure 4: Work flow of a distributed LDM model in MEG.
  • Figure 5: Visible results text-guided image-to-image generation.
  • ...and 1 more figures