In this project, we focus on studying the scaling of latent novel view synthesis models. For a while, novel view synthesis solutions have involved explicit 3D representations, but recent work has shown that this inductive-bias is unnecessary. However, there is yet to be a thorough analysis of these sorts of models, which we aim to do across several architectures and several scales of compute and data. We’ll then use this analysis to introduce several refinements to existing latent novel view synthesis architectures and training schemes.

I am participating in SuperUROP in order to strengthen my general research and communication skills while deepening my computer vision knowledge through my particular project. With a background in machine learning research across both academic and industry settings, I’m excited to apply my experiences in a scaling focused project that can offer valuable insights for other academic researchers.