The runtime-complexity of attention layers scales quadratically with the number of tokens, and thus triangles in our case. As a result, we limit the total number of triangles in our scenes to 4,096;