QuantumGS: Quantum Encoding Framework for Gaussian Splatting

Grzegorz Wilczyński^1,2, Rafał Tobiasz^1,2, Paweł Gora¹ Marcin Mazur¹ Przemysław Spurek^1,2

¹ Jagiellonian University ² IDEAS Research Institute

Abstract

Recent advances in neural rendering, particularly 3D Gaussian Splatting (3DGS), have enabled real-time rendering of complex scenes. However, standard 3DGS relies on spherical harmonics, which often struggle to accurately capture high-frequency view-dependent effects such as sharp reflections and transparency. While hybrid approaches like Viewing Direction Gaussian Splatting (VDGS) mitigate this limitation using classical Multi-Layer Perceptrons (MLPs), they remain limited by the expressivity of classical networks in low-parameter regimes. In this paper, we introduce QuantumGS, a novel hybrid framework that integrates Variational Quantum Circuits (VQC) into the Gaussian Splatting pipeline. We propose a unique encoding strategy that maps the viewing direction directly onto the Bloch sphere, leveraging the natural geometry of qubits to represent 3D directional data. By replacing classical color-modulating networks with quantum circuits generated via a hypernetwork or conditioning mechanism, we achieve higher expressivity and better generalization.

Method

Our framework integrates quantum processing into 3D Gaussian Splatting for view-dependent color and opacity residuals. Top: Two interchangeable pipelines. Pipeline I (Hyper-Quantum) generates per-Gaussian VQC parameters via hypernetwork from spatial hash encoding. Pipeline II (Joint-Hash Global) feeds spatial and directional hash features to a shared quantum network. Bottom: Hybrid QMLP maps viewing directions to Bloch sphere via rotation gates (Ry, Rz), processes through VQC with circular entanglement, and decodes measurements via classical MLP to guide rendering.

Quantitative Results

NeRF Synthetic Dataset

Table 1. Quantitative comparison on the NeRF Synthetic dataset. We report PSNR, SSIM, LPIPS, and rendering speed (FPS). Our method achieves better results while maintaining real-time rendering capabilities.

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FPS ↑
NeRF	31.01	0.947	0.081	0.023
VolSDF	27.96	0.932	0.096	---
Ref-NeRF	31.29	0.947	0.058	---
ENVIDR	28.13	0.956	0.067	---
QRF	32.65	0.960	0.029	47.26
GS	33.30	0.969	0.030	733.00
VDGS	33.37	0.969	0.032	284.29
QuantumGS (Ours)	33.98	0.970	0.030	12.64

Large-Scale Scenes (Mip-NeRF 360, Deep Blending, Tanks & Temples)

Table 2. Quantitative results on large-scale real-world datasets. Our method achieves comparable results on Mip-NeRF 360, and best results on Deep Blending and Tanks & Temples datasets. Notably, we maintain real-time rendering speeds despite utilizing a quantum circuit simulator.

Dataset	Method	PSNR ↑	SSIM ↑	LPIPS ↓	FPS ↑
Mip-NeRF 360	MiPNeRF360	27.69	0.792	0.237	0.06
	GS-30K	27.21	0.815	0.214	134
	VDGS	27.64	0.813	0.220	41.35
	QuantumGS	27.27	0.793	0.244	10.78
Deep Blending	MiPNeRF360	29.40	0.901	0.245	0.09
	GS-30K	29.41	0.903	0.243	137
	VDGS	29.54	0.906	0.243	44.72
	QuantumGS	30.15	0.916	0.163	10.76
Tanks & Temples	MiPNeRF360	22.22	0.759	0.257	0.14
	GS-30K	23.14	0.841	0.183	154
	VDGS	24.02	0.851	0.176	28.53
	QuantumGS	24.70	0.888	0.118	16.15

Qualitative Comparisons

Synthetic Scenes

Object-centric scenes. In the Drums scene, VDGS blurs the reflection on the drum surface, losing the geometric definition. QuantumGS preserves the distinct shape. In the LEGO scene, VDGS exhibits floater artifacts near the roof. QuantumGS recovers occlusion shadows on the chassis.

Real-World Scenes

Comparisons on real-world datasets. In the Truck scene (Tanks and Temples), standard 3DGS fails to capture high-frequency reflections on the windshield. QuantumGS recovers sharp specular details. In the Kitchen scene (Mip-NeRF 360), standard 3DGS renders the LEGO truck with unnatural foggy appearance. QuantumGS produces clear boundaries.