Generative AI for Image Synthesis - A Study on CPUs

With the immense growth of Artificial Intelligence (AI), especially in image processing, many people have taken advantage of its numerous capabilities.It is used in different domains, including generative imaging (GenI).Some of the crucial applications of GenI include content creation, shadow synthesis, and in-painting.However, image synthesis comes with flaws.Due to the complexity of generating images, the process can have very high processing times.This paper analyzes the shortcomings by profiling LaMa, an in-painting model run on CPUs.We provide insights by highlighting the bottlenecks that hinder performance.Our main finding is that the mkldnn_convolution layer takes up most of the time, indicating that speeding up this layer's computations could drastically improve overall processing time.