You are tasked with developing an input pipeline for a machine learning model that processes a large volume of images from various sources, requiring low latency. Upon initial analysis, you realize that the input data exceeds the memory capacity of your current setup. Considering Google's best practices for handling large datasets that do not fit in memory, which of the following approaches should you implement to optimize the dataset creation and training process? Choose the best option.