Beyond Spherical Harmonics: Rethinking Appearance Models for Radiance Reconstruction

This paper takes aim at a familiar bottleneck in neural rendering: spherical harmonics are convenient, but they get blurry fast when you need sharp angular detail. The authors systematically compare a broad set of spherical functions for scene reconstruction, including some they say are new to graphics and vision.

Their main result is a new representation called the Normalized Anisotropic Spherical Gabor function. It is designed to model high-frequency view-dependent appearance more compactly than higher-order SH, which usually means more coefficients, more memory, and more runtime cost.

For developers, the practical angle is straightforward: if your reconstruction or neural rendering pipeline needs specular highlights, glints, or other tight reflection behavior, this could reduce the usual quality-vs-budget tradeoff. The paper claims up to 5x better memory efficiency and a more efficient evaluation path, which matters for training scale, inference cost, and potentially real-time deployment.

The work is still in the research stage, but it fits a broader trend away from treating SH as the default answer for appearance encoding. If the results hold up outside the paper’s benchmarks, this could be a useful building block for radiance-field systems that need richer material response without ballooning model size.