Autoencoders have been rapidly evolving as a powerful tool for reduced order modelling in the fluid mechanics community. Convolutional neural networks (CNNs) and graphics processing units (GPUs) have made it much easier to immediately model the flowfield data by reducing the computational cost associated with training these networks. Despite the huge boost in the number of studies in the field, many fundamental questions remain unanswered. Despite the fact that a simple multi-layered perceptron (MLP) can theoretically be used to simulate any type of complex function, in practice people have started to employ more complex architectures, such as multi-scale CNNs, to model fluid flow problems. In the light of this, the question arises: Does complexity of the model architecture matter? Is there a limit to how complicated and intricate the network can be built before it becomes ineffective or does it always lead to more accurate reconstruction. These questions were explored in this study by taking into account a sophisticated CNN autoencoder model called multiscale CNNAE and comparing it to the standard baseline model by training and analysing the models on two distinct flowfield datasets: a simple flow past a stationary cylinder and a case of an airfoil that is plunging at low Reynolds number. It was found that across both the datasets, the basic CNNAE performed on par with or even better than the complicated multiscale CNNAE.