What considerations are necessary for deploying a real-time generative model on edge devices?