Diffusion-Denoised Hyperspectral Gaussian Splatting

3DV 2026
Sunil Kumar Narayanan1, Lingjun Zhao1, Lu Gan1, Yongsheng Chen1
1Georgia Institute of Technology
Paper Supplementary Code (Coming soon) arXiv

Wavelength Sweep (Anacampserous)

Diffusion Effect


We present DD-HGS, a novel framework that combines 3D Gaussian Splatting with hyperspectral diffusion modeling for high-fidelity 3D hyperspectral reconstruction. Our approach achieves state-of-the-art performance on both spatial and spectral accuracy while maintaining real-time rendering capabilities.

Abstract

We present DD-HGS, a novel framework that combines 3D Gaussian Splatting with hyperspectral diffusion modeling for high-fidelity 3D hyperspectral reconstruction. Our approach achieves state-of-the-art performance on both spatial and spectral accuracy while maintaining real-time rendering capabilities. The framework leverages a wavelength encoder with positional encoding to capture spectral dependencies and employs a diffusion-based refinement module to enhance spectral fidelity. Extensive experiments on the BaySpec and Surface Optics datasets demonstrate significant improvements over existing NeRF-based and 3DGS-based methods in terms of PSNR, SSIM, SAM, and RMSE metrics.

Overview Architecture Diagram

Overview of our DD-HGS framework combining 3D Gaussian Splatting with hyperspectral diffusion modeling for high-fidelity 3D hyperspectral reconstruction.

Detailed Scene Results

We provide per-scene quantitative metrics for our hyperspectral reconstructions. Our method achieves state-of-the-art performance across multiple evaluation metrics including PSNR, SSIM, SAM, and RMSE.

BaySpec Dataset

Best results are highlighted in bold.

Method Pinecone Caladium Anacampserous FPS ↑
PSNR ↑ SSIM ↑ SAM ↓ RMSE ↓ PSNR ↑ SSIM ↑ SAM ↓ RMSE ↓ PSNR ↑ SSIM ↑ SAM ↓ RMSE ↓
NeRF 22.820.61130.041460.0728 23.120.583480.04910.0709 24.120.62200.03840.0628 0.13
MipNeRF 21.450.57080.04830.0865 23.400.59330.04650.0688 23.430.61600.04880.0699 0.09
TensoRF 24.120.65450.03930.0625 24.790.62440.06160.0677 25.070.66590.03940.0588 0.17
Nerfacto 15.360.48250.07750.0896 20.630.54180.07540.0835 20.930.61200.07080.0857 0.50
MipNeRF360 25.930.73780.03110.0510 26.190.71120.03830.0616 26.320.72610.03700.0524 0.01
Hyper-NeRF 20.070.58110.07250.1521 19.500.75050.08430.0893 20.390.72600.04930.0732 0.47
3DGS 22.650.63090.05160.0885 27.710.73110.02760.0525 27.920.75960.04310.0596 78.10
Hyper-GS 27.000.72990.03900.0474 27.700.83540.02710.0414 26.920.75450.01880.0469 2.31
DD-HGS (Ours) 25.110.93470.05720.0244 27.860.93620.02210.0417 28.570.94900.02170.0381 2.43

Surface Optics Dataset

Best results are highlighted in bold.

Method Rosemary Basil FPS ↑
PSNR ↑ SSIM ↑ SAM ↓ RMSE ↓ PSNR ↑ SSIM ↑ SAM ↓ RMSE ↓
NeRF 8.420.74610.02840.3560 9.910.55340.07960.5256 0.13
MipNeRF 13.640.56841.00000.2083 11.010.58780.07280.5334 0.09
TensoRF 12.100.73350.02120.2662 15.230.58110.04350.3628 0.20
Nerfacto 18.660.83660.07080.1025 16.540.79150.01760.1655 0.57
MipNeRF360 8.470.75180.08760.3825 13.920.85840.04970.2035 0.14
Hyper-NeRF 18.600.87000.00770.1187 16.910.77100.01700.1587 0.49
3DGS 25.560.96950.00280.0534 21.790.93850.01010.0897 79.00
Hyper-GS 26.770.98450.00210.0445 25.300.95030.00510.0569 3.56
DD-HGS (Ours) 28.540.91910.00430.0040 48.130.93400.00190.0018 2.95

Qualitative Results

We present qualitative comparisons of our method against baseline approaches on the BaySpec and Surface Optics datasets.

Spectral Curve Visualization
Spectral Curve Visualization

Spectral curve visualization comparing our DD-HGS method with baseline approaches.

BaySpec
BaySpec Qualitative Results

Qualitative results on the BaySpec dataset comparing our DD-HGS method with baseline approaches.

Surface Optics Dataset
Qualitative Results on Surface Optics Dataset

Qualitative results on the Surface Optics dataset comparing our DD-HGS method with baseline approaches.

BibTeX

@misc{narayanan2025hyperspectralgaussiansplatting,
      title={Hyperspectral Gaussian Splatting}, 
      author={Sunil Kumar Narayanan and Lingjun Zhao and Lu Gan and Yongsheng Chen},
      year={2025},
      eprint={2505.21890},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.21890}, 
}