Recent advancements in VGGT architectures have enabled users to generate 3D scene reconstruction from sequential visual dataset. However, the majority of recent successful VGGT architectures require millions of parameters, upto 10 billions. This makes them energy-, computation-, and memory-intensive. As a result, such architectures require long inference times and power-consuming computational resources (e.g., GPU). For these reasons, compression of VGGT architectures is essential for the training and deployment of such models in practice. In my research, I will be compressing VGGT architectures. In order to conduct the research, first I have to train and evaluate many state-of-the-art deep learning architectures, as well as new proposed architecture. I would like to use the supercomputer such as Arrhenius resources to perform various time-consuming simulations for my research.