Amazon

Optimization of Image Generation Models for Edge

EdgeDiffuse explores the deployment of Stable Diffusion–based image generation models on resource-constrained edge devices. As generative AI models grow increasingly powerful, their computational demands limit accessibility beyond cloud infrastructure. This project addressed that gap by applying mixed-precision quantization, structured pruning, and knowledge distillation to compress Stable Diffusion for deployment on ARM-based hardware (Orange Pi RK3588), with and without NPU acceleration. The goal was to achieve at least 20–25% model size reduction while maintaining acceptable image quality, enabling real-time, on-device generative AI without cloud dependency.