This project investigates the spectral bias of vision-based generative modeling of fluid dynamics data from recorded physical simulations. We seek to establish a fair and reproducible benchmark of generative models for fluid dynamics data, focusing on comparing and evaluating their capacity to model high-frequency spatial information. We will systematically analyze the trade-offs associated with various architectures, training data sizes, and parameter budgets. Our evaluation metrics will monitor high-frequency modeling and computational cost in terms of training and inference. Furthermore, we will assess the effect of spectral vs spatial representation on modeling high-frequency information. Establishing this benchmark is critical because conventional image-generation metrics often fail to capture the high-frequency characteristics of data with heavy-tailed spectral content, which are essential in modeling fluid dynamics. This benchmark will be a foundation in identifying the associated tradeoffs between modeling high-frequency information and other performance metrics, and in helping to highlight research gaps for future work. This work will serve as the baseline for our subsequent research on conditional generative modeling of fluid dynamics. The requested compute resources are necessary to ensure fair comparisons through hyperparameter optimization and evaluation in a controlled setting. We will rely on repeated experiments to obtain confidence intervals for reliable comparisons. This project is conducted by KTH research intern Shaghayegh Mirjalili (mirjali@kth.se), under the close supervision of Computer Science PhD student Amir Mehrpanah (amirme@kth.se) and Fluid Mechanics Postdoc Cristiano Pimenta (cristpi@kth.se), with general supervision from Prof. Hossein Azizpour (azizpour@kth.se) and Prof. Seyedshahabaddin Mirjalili (msey@kth.se).