Mobile AI-generated content (AIGC) services (mobile ChatGPT)

Abstract

This project undertakes the task of deploying a Stable Diffusion, a text-to-image generation model, on edge devices such as Android phones to enable on-device image generation without requiring active internet connection. Furthermore, this project also explores the techniques of utilizing large language models to improve the visual accuracy and aesthetic quality of generated images. To achieve this, an Android application was developed, integrating a diffusion image generation pipeline in Android native code. As an optional component of the image generation pipeline, this project also utilizes edge networks, such as a local area network (LAN), to run a lightweight large language model (LLM), enhancing prompt processing without requiring cloud based resources, thus reducing latency and ensuring greater data privacy. Due to the unavailability of a physical Android device during development, the project was solely developed on Android emulator. This precluded the exploration of mobile GPU-based inferencing, as Android AI inferencing libraries do not support GPU acceleration on Android Emulators. Consequently, all inference tasks were restricted to the CPU. This report aims to elaborate on the design and development of Stable Diffusion image generation pipeline in Android native code, including model conversion for compatibility with Android and edge devices. It also covers the integration of a large language model into the pipeline for enhancing image generation. Additionally, the report discusses the limitations encountered due to the early stage of the Edge AI development ecosystem and proposes potential solutions for future work to optimize deployment efficiency and performance.