[duckworthd]

The viewer biases towards assets closer to user's camera (otherwise you'd have to load the whole scene!). We tried training SMERF with a larger number of smaller submodels, but at some point, it becomes too onerous to train and quality begins to suffer.