Accelerating Mobile Edge Generation (MEG) by Constrained Learning

TL;DR

This work tackles on‑device, high‑resolution image generation by distributing a large latent diffusion model (LDM) across an edge server (ES) and user equipment (UE) in a mobile edge generation (MEG) setting. It introduces a low‑complexity protocol that combines offline backbone distillation of few‑step diffusion with a length‑adaptive feature merging scheme, enabling online decisions governed by a constrained Markov decision process (MDP). To enforce feasibility, the authors develop MEG‑CVPO, a constrained variational policy optimization algorithm that alternates an E‑step (finding a feasible auxiliary policy) with an M‑step (updating the main policy within a trust region). Numerical results on SDXL demonstrate over 40% end‑to‑end latency reduction for 1024×1024 outputs, while MEG‑CVPO provides constraint guarantees and controllable quality–cost tradeoffs, highlighting the practical impact for latency‑ and energy‑constrained mobile AIGC. The approach offers a scalable path to cost‑efficient, privacy‑preserving on‑device generation across edge networks, with potential extensions to multi‑user and parallelized MEG deployments.

Abstract

A novel accelerated mobile edge generation (MEG) framework is proposed for generating high-resolution images on mobile devices. Exploiting a large-scale latent diffusion model (LDM) distributed across edge server (ES) and user equipment (UE), cost-efficient artificial intelligence generated content (AIGC) is achieved by transmitting low-dimensional features between ES and UE. To reduce overheads of both distributed computations and transmissions, a dynamic diffusion and feature merging scheme is conceived. By jointly optimizing the denoising steps and feature merging ratio, the image generation quality is maximized subject to latency and energy consumption constraints. To address this problem and tailor LDM sub-models, a low-complexity MEG acceleration protocol is developed. Particularly, a backbone meta-architecture is trained via offline distillation. Then, dynamic diffusion and feature merging are determined in online channel environment, which can be viewed as a constrained Markov Decision Process (MDP). A constrained variational policy optimization (CVPO) based MEG algorithm is further proposed for constraint-guaranteed learning, namely MEG-CVPO. Numerical results verify that: 1) The proposed framework can generate 1024$\times$1024 high-quality images over noisy channels while reducing over $40\%$ latency compared to conventional generation schemes. 2) The developed MEG-CVPO effectively mitigates constraint violations, thus flexibly controlling the trade-off between image distortion and generation costs.

Figures (7)

• Figure 1: The proposed dynamic MEG acceleration framework.
• Figure 2: A pipeline of the proposed MEG for ES-UE co-inference.
• Figure 3: Latency and energy consumption evaluation of the proposed MEG framework.
• Figure 4: Images created by different generation methods. The text prompt is given by "Two people standing in the snow with skis".
• Figure 5: Comparisons of average MSE achieved by different generation schemes.