A Model Segmentation Method for Personalized Mobile Edge Generation

Abstract

A personalized mobile edge generation (P-MEG) stable diffusion structure is proposed. The stable diffusion model is deployed on the edge server, enabling users to train personalized weights for customized generation tasks. To mitigate oscillations and accelerate convergence speed during user-personalized training, an effective constant scaling connection (CSC) with random model segmentation method is introduced. In addition, the stability of forward propagation in conditioned stable diffusion generation is explored. Furthermore, we theoretically analyze the benefits of CSC in enhancing the stability of user-personalized model training. The numerical results demonstrate that the proposed CSC method effectively assists the user in training personalized weights. Additionally, the CSC significantly stabilizes hidden feature oscillations and accelerates convergence speed during the training of personalized stable diffusion model.