Comment by markisus

Yes, the normal case uses 2D input, but it can take hours to create the scene. Using the depth channel allows me to create the scene in 33 milliseconds, from scratch, every frame. You could conceptualize this as a compromise between raw pointcloud rendering and fully precomputed Gaussian splat rendering. With pointclouds, you have a lot visual artifacts due to sparsity (low texture information, seeing "through" objects"). With Gaussian splatting, you can transfer a lot more of the 2D texture information into 3D space and render occlusion and view-dependent effects better.