Generative AI is transforming the way people accomplish tasks, reshaping numerous industries and workflows. In this study, we focus on one of the most popular applications of generative AI: text-to-image generation, a technology that gained significant attention in 2021. Despite its popularity, even after three years, text-to-image generation remains less efficient and slower on edge devices like mobile phones compared to its web-based counterparts. This research investigates various model variations and analyzes the factors influencing inference speed in image generation. While Mobile Diffusion, though not yet commercially available, claims to achieve an impressive inference speed of 0.02 seconds, we explore whether architectural modifications and sampling techniques can further enhance performance without compromising image quality. Our findings indicate that adjustments to sampling operations, such as switching the scheduler from PNDMS to DDIM, resulted in a 6.57% increase in inference speed, albeit with a slight degradation in FID and CLIP scores. In contrast, architectural changes yielded significant improvements, achieving up to a 15.24% increase in speed while maintaining favorable results in FID and CLIP scores.