Beyond Spherical Harmonics: Rethinking Appearance Models for Radiance Reconstruction

Computer Graphics Forum [EGSR'26]
Teaser: Normalized Anisotropic Spherical Gabor kernel for radiance reconstruction We introduce the Normalized Anisotropic Spherical Gabor (NASGabor): an anisotropic, multi-modal spherical kernel with a closed-form integral expression. The proposed representation is both compact and highly expressive, and is particularly well-suited for modeling view-dependent effects in novel view synthesis, outperforming commonly used approaches such as Spherical Harmonics.


View-dependent appearance modeling remains a challenging problem in novel-view synthesis and reconstruction. Accurately representing complex angular effects often requires substantial memory and computational resources. For learning-based methods, a common approach is to rely on low-order Spherical Harmonics (SH), which limits the ability to model complex view-dependent effects and tends to produce overly smooth or diffuse representations. To address these limitations, we systematically evaluate a wide range of spherical functions in the context of scene reconstruction — many of them introduced to graphics and computer vision for the first time in this paper. Guided by the resulting insights, we develop the Normalized Anisotropic Spherical Gabor (NASGabor) function: a compact, anisotropic, multi-modal kernel with a closed-form integral that achieves higher-quality reconstruction of view-dependent phenomena such as glints, while being up to five times more memory-efficient and more efficient to evaluate than commonly used SH expansions.



A systematic study of spherical functions

We benchmark a wide family of parametric spherical functions — from classical Spherical Harmonics, isotropic and anisotropic Spherical Gaussians, Normalized ASGs and Spherical Betas, all the way to less explored families coming from statistics, like the Fisher-Bingham families. All of them are plugged into the same radiance-field pipeline and evaluated under identical conditions.



Visualization of the spherical functions evaluated in our study A subset of the spherical functions covered in our study.

From this analysis we distill three properties that govern effectiveness in radiance-field reconstruction: anisotropy (anisotropic kernels consistently outperform isotropic ones), coupled multi-modality (jointly parameterized lobes are easier to optimize than independent ones), and closed-form integration (learning normalized kernels improves the results, but integrals must be practical to compute during training). A single spherical lobe combined with diffuse color can match or surpass third-degree SH while using roughly five times fewer appearance parameters.



The Quality vs. Memory Trade-off

The chart below summarizes reconstruction quality (PSNR) versus per-primitive memory for the spherical functions in our study on Mip-NeRF360, grouped by indoor and outdoor scene type and varying lobe count. Motivated by this experiment, we developed a new kernel: a normalized anisotropic spherical Gabor function or NASGabor, which outperforms all other kernels in the task.



PSNR vs. memory for spherical functions PSNR vs. memory footprint, averaged over indoor and outdoor scenes. NASGabor sits on the Pareto frontier, matching or surpassing higher-degree SH at a fraction of the size.

The NASGabor Kernel

NASGabor is a normalized anisotropic spherical Gaussian (NASG) envelope multiplied by a positive cosine carrier that introduces a controllable number of ripples. We derive analytic gradients for NASGabor and un-modulated NASG, and implement the appearance model as forward/backward CUDA kernels inside gsplat.



fNASGabor(v) = e2λ(K1 − 1) K0 1 + cos(k xv) 2

Orange: NASG (Gaussian) envelope Blue: shifted cosine carrier (harmonic)



Drag on the sphere to orbit, scroll to zoom. Sliders control kernel parameters directly.

Unnormalized preview: the plotted value is the product of the two factors above, clipped for display (not divided by the analytic normalization constant).

Comparison to Prior Work

We compare NASGabor against recent appearance models for primitive-based radiance fields, all implemented in gsplat and evaluated on Mip-NeRF360, Deep Blending, and Tanks & Temples under identical training budgets (learning rates are set per method: manually tuned baselines or our camera-spacing heuristic for NASGabor). Overall, NASGabor matches or outperforms prior methods in quality at a fraction of the memory cost, with faster training and rendering.



Quantitative comparison on Mip-NeRF360, Deep Blending, and Tanks & Temples. Params denotes appearance parameters per primitive. Best results per column are highlighted (1st / 2nd / 3rd).
Method Params Mip-NeRF360 Deep Blending Tanks & Temples
PSNR ↑SSIM ↑LPIPS ↓ PSNR ↑SSIM ↑LPIPS ↓ PSNR ↑SSIM ↑LPIPS ↓
2DGS + SH 48 27.220.8040.275 29.560.9040.325 22.850.8270.244
3DGS-MCMC + SH 48 27.990.8300.229 29.490.9120.306 24.460.8660.174
3DGUT-MCMC + SH 48 27.820.8260.233 29.870.9130.309 24.200.8610.180
Beta Splatting + SH 48 28.000.8300.226 29.790.9110.294 24.340.8680.173
Beta Splatting + SB 15 28.090.8300.227 29.240.9080.301 24.640.8700.173
Spherical Voronoi (12 params) 12 28.190.8300.230 29.560.9070.301 24.640.8690.197
Spherical Voronoi (48 params) 48 28.460.8320.226 29.900.9120.301 24.670.8700.172
NASGabor — 1 lobe 12 28.400.8290.228 29.910.9110.290 24.640.8670.173
NASGabor — 2 lobes 21 28.460.8300.227 30.240.9150.287 24.680.8690.173
NASGabor — 4 lobes 39 28.460.8300.227 30.370.9140.286 24.790.8690.172


Training and rendering efficiency on Mip-NeRF360 (appearance parameters per primitive).
Method Params Storage ↓ Train time ↓ FPS ↑
Spherical Harmonics 48 747.68 MB 15m50s 128
Spherical Beta 15 356.04 MB 15m08s 143
Spherical Voronoi 48 747.68 MB 17m32s 130
NASGabor — 1 lobe 12 320.44 MB 13m38s 146
NASGabor — 2 lobes 21 427.25 MB 14m42s 143
NASGabor — 4 lobes 39 640.87 MB 16m36s 136


Sharper Highlights, Fewer Parameters

Replacing SH with NASGabor inside an otherwise standard 3DGS-style pipeline yields measurably better view-dependent reconstructions. FLIP error maps below show that NASGabor more accurately depicts sheens, highlights, and other high-frequency effects than spherical harmonics.



FLIP comparison on MipNeRF360 FLIP maps comparing SH and NASGabor on MipNeRF360 scenes. Brighter regions indicate higher error.

Qualitative MipNeRF360 reconstructions Qualitative comparison on Mip-NeRF360. NASGabor achieves comparable or superior appearance reconstruction; insets highlight view-dependent effects captured by our model.

Free Disentanglement of Diffuse and Specular

An interesting byproduct of mixing diffuse color with spherical functions as an appearance model is its capacity to naturally disentangle diffuse albedo from other view-dependent effects related to material properties or lighting, with no extra losses, regularizers, or supervision. This basic intrinsic decomposition could be useful for simple relighting, material estimation, or reflection removal.



Diffuse-specular disentanglement Full reconstruction (a), and its diffuse (b) and specular (c) components. Diffuse appearance is modeled explicitly; view-dependent effects are captured by NASGabor lobes.

Normalization Matters

One the most interesting findings, and one that can be extended to other proposed appearance models for 3DGS (e.g. Spherical Beta kernels[Liu et al.'25]). Learning normalized kernels (individually) consistently improves quality across scenes. Kernels where integration is analytic or easy to compute becomes critical for maximizing the potential of the model. We derived closed-form integrals for Spherical Beta kernels and NASGabor, which we detail in the paper.



Normalized vs. unnormalized PSNR Per-scene PSNR with and without analytic normalization of the spherical functions.

Citation

Ewa Miazga, Jorge Condor, and Piotr Didyk. Beyond Spherical Harmonics: Rethinking Appearance Models for Radiance Reconstruction, Computer Graphics Forum (Proc. EGSR 2026).

@article{Miazga2026BeyondSH,
  author = {Miazga, Ewa and Condor, Jorge and Didyk, Piotr},
  title = {{Beyond Spherical Harmonics: Rethinking Appearance Models for Radiance Reconstruction}},
  journal = {Computer Graphics Forum},
  note = {Proc. EGSR 2026},
  year = {2026},
  eprint = {2606.09794},
  archivePrefix = {arXiv},
}


Acknowledgements

We thank Wenzel Jakob for supporting the project, Nicolai Hermann for interesting discussions, and Sophie Kergaßner for help with designing the figures. This project has received funding from the Swiss National Science Foundation (SNSF, Grant 200502). We acknowledge access to Alps at the Swiss National Supercomputing Centre, Switzerland under USI's share (project ID u6).


Affiliations

EPFL USI