NASGabor Offers 5× Memory Savings Over Spherical Harmonics in Novel View Synthesis

Researchers from EPFL and Università della Svizzera italiana have introduced NASGabor (Normalized Anisotropic Spherical Gabor), a new spherical kernel function designed to replace Spherical Harmonics (SH) in modeling view-dependent appearance within radiance-field pipelines. Accepted to Computer Graphics Forum (EGSR 2026), the work benchmarks a broad family of parametric spherical functions under identical scene-reconstruction conditions and distills three properties that drive quality: anisotropy, coupled multi-modality, and closed-form integration.

NASGabor combines a normalized anisotropic Spherical Gaussian envelope with a controllable cosine carrier, sits on the Pareto frontier of reconstruction quality (PSNR) versus memory footprint, and can match or surpass third-degree SH while using roughly five times fewer appearance parameters per primitive. The authors implement the model as custom forward/backward CUDA kernels inside gsplat, with analytic gradients derived for both NASGabor and its unmodulated NASG variant. The result is particularly effective for high-frequency view-dependent phenomena such as glints that low-order SH representations tend to smooth away.