Is Gaussian Splatting Becoming Neural Again? A Taxonomy and Controlled Study of Learned Parameterization

Summary

Three-dimensional Gaussian Splatting (3DGS) uses explicit Gaussian primitives and efficient rasterization, but recent systems increasingly add neural networks to generate or share Gaussian parameters. This paper organizes that trend across five axes: attribute decoding, spatial sharing, view-conditioned decoding, topology generation, and amortized inference. An analysis of 19 representative methods finds that these design choices address different limitations and cannot be meaningfully reduced to a binary distinction between neural and non-neural systems. The authors then isolate three forms of neural parameterization in a controlled study on mip-NeRF 360. Sharing appearance and opacity information improves reconstruction quality, whereas decoding geometric structure provides no additional gain. The results support selective neuralization: learned shared functions are useful when they capture reusable correlations while preserving the local geometric freedom of explicit splats.