Compact Neural Appearance Models for Efficient Gaussian Splatting
arXiv 2026

Florian Hahlbohm1, Jorge Condor2, Linus Franke3, Martin Eisemann1, Marcus Magnor1,4
1TU Braunschweig 2Università della Svizzera italiana 3University of Würzburg 4University of New Mexico
Appearance Framework
State-of-the-art appearance models, runtime-compiled into Faster-GS.
Spherical Harmonics Spherical Voronoi NASG NASGabor Neural
+ a WebGL viewer for all of them.
Compact Neural Appearance
28 vs. 192 bytes / Gaussian
1.3× faster training than SH
≥ SH in PSNR on average
Insights & Suggestions

Expressive appearance reduces geometric compensation

SH fakes the reflection on the sketch block with elongated white Gaussians, while NASGabor models it in the view-dependent residual.

… but it also fits non-static scene content

The topmost book moved during the capture of bonsai: SH relies on popping, while NASGabor partially fits this as a view-dependent effect.

The residual reveals the path where the camera operator stepped on the grass during the capture of garden.

Problem
Benchmark gains of expressive appearance models partly measure robustness to non-static content rather than better modeling of outgoing radiance. Advancing static scene reconstruction requires evaluations that account for this, e.g., through better metrics, masking, more controlled captures, or realistic synthetic scenes.
Opportunity
With expressive appearance models, the optimizer relies less on faking complex view-dependent effects through geometry. This should benefit more exact rendering methods that cannot rely on popping (e.g., HTGS) as well as representations using opaque surfaces (e.g., MeshSplatting).

Side note: color activation and opacity

Effect of color activation and opacity learning rate on a test view of the garden scene for SH, NASGabor, and the neural model

Unbounded colors let bright regions be fit by low-opacity Gaussians with colors above one. Bounded activations avoid this and can recover sharper geometry, though the best configuration varies per scene.

Abstract

Explicit primitive-based radiance fields such as 3D Gaussian Splatting typically model view-dependent appearance using low-order spherical harmonics (SH). Although efficient to evaluate, SH coefficients dominate per-primitive storage and memory traffic, while their band-limited basis restricts angular detail.
We present a thorough, end-to-end comparison of SH and recent spherical appearance models and introduce an implicit alternative that decodes compact per-primitive latent codes using a tiny shared MLP. We integrate all models into the same optimized pipeline, fusing their forward and backward passes into a differentiable CUDA rasterizer and provide a portable WebGL viewer for laptop and mobile GPUs.
Our evaluation across reconstruction quality, memory use, and optimization and rendering performance shows that recent spherical models offer the strongest overall quality–efficiency trade-off. Our neural representation is the most compact model evaluated and, compared to third-degree SH, reduces the per-primitive appearance footprint from 192 to 28 bytes, accelerates optimization by 1.3×, while improving reconstruction quality. We further analyze how appearance parametrization shapes optimization, identifying differences in recovered geometry and the tendency of expressive models to absorb non-static scene content. Together, our framework and analysis provide practical guidance for replacing SH beyond what image metrics alone can capture.

Citation

@misc{hahlbohm2026efficientgaussianappearance,
  title         = {Compact Neural Appearance Models for Efficient Gaussian Splatting},
  author        = {Florian Hahlbohm and Jorge Condor and Linus Franke and Martin Eisemann and Marcus Magnor},
  year          = {2026},
  eprint        = {2609.05255},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.05255},
}

Acknowledgements

We thank Timon Scholz for his contributions to the software infrastructure supporting this work and Jannis Möller for identifying and helping us fix a bug in tiny-cuda-nn. This work was partially funded by the DFG projects “Real-Action VR” (ID 523421583) and “Increasing Realism of Omnidirectional Videos in Virtual Reality” (ID 491805996).

The website template was adapted from Zip-NeRF, who borrowed from Michaël Gharbi and Ref-NeRF.