HybridSplat: Fast Reflection-baked Gaussian Tracing using Hybrid Splatting

Chang Liu1 Sichao Wang1 Jianwei Guo1 Shi-Sheng Huang1
1 Beijing Normal University
Paper coming soon Code coming soon
HybridSplat Introduction Figure

Figure 1: The novel view rendering results of our HybridSplat (left) and performance comparison with different 3D Gaussian splatting approaches (right). As ray-tracing based 3DGS(marked green), our HybridSplat outperforms latest Gaussian tracing such as EnvGS, Ref-Gaussian with much faster rendering speed and fewer Gaussian primitives, while preserving similar rendering quality which is consistently better than original 3DGS and other general splatting pipeline acceleration approaches (marked blue). The circle size represents the corresponding Gaussian primitive numbers.

HybridSplat Teaser Figure

Figure 2: The visual comparison results between our HybridSplat and similar ray-tracing based Gaussian splatting like EnvGS along with the 3DGS. The rendering speed (FPS), quality (PSNR) and Gaussian primitive number (GS) are also demonstrated respectively, where 'M' is short for million.

Abstract

Rendering complex reflection of real-world scenes using 3D Gaussian splatting has been a quite promising solution for photorealistic novel view synthesis, but still faces bottlenecks especially in rendering speed and memory storage. This paper proposes a new Hybrid Splatting (HybridSplat) mechanism for much faster complex reflection scene rendering with significantly reduced Gaussian primitives, while preserving similar level of high fidelity rendering quality compared with state-of-the-art Gaussian tracing. One novel idea within our HybridSplat is a new reflection-baked Gaussian tracing, which bakes the view-dependent reflection within each Gaussian primitive while rendering the reflection using tile-based Gaussian splatting. Thus we can hybridly integrate the reflective Gaussian primitives with base Gaussian primitives together using the efficient tile-based Gaussian splatting framework for high-fidelity complex scene rendering. Moreover, we further introduce a differentiable reflection-sensitive Gaussian pruning to reduce the Gaussian primitives for lower memory storage while preserving the reflection rendering quality. By extensive evaluation, our HybridSplat accelerates about 7x rendering speed across complex reflective scenes from Ref-NeRF, NeRF-Casting with 4x fewer Gaussian primitives than similar ray-tracing based Gaussian splatting baselines. Besides, our HybridSplat can be seamlessly integrated to pioneering general splatting pipeline acceleration frameworks for further rendering speed acceleration, providing a useful tool for complex reflective scenes rendering with state-of-the-art rendering speed and memory storage.

Demo Video

This video demonstrates our pipeline and comparisons with SOTA methods.

Methodology

Pipeline

Figure 3: Pipeline of HybridSplat. We reconstruct complex reflective scenes with base Gaussians \(G_{base}\) and reflective Gaussians \(G_{ref}\), where the latter bake the reflection using a novel reflection-baked Gaussian tracing (left). These Gaussians are efficiently rendered using our hybrid splatting including a base splatting branch for base Gaussians and a reflective branch for reflective Gaussians in a hybrid manner (middle). We further perform a differentiable reflection-sensitive Gaussian pruning and splatting pipeline acceleration, to adaptively reduce the Gaussian primitives and improve the rendering speed.

Comparative Results

Qualitative comparison on Ref-NeRF and NeRF-Casting Shiny Scenes datasets.

Comparison on datasets with complex reflections

Evaluation on grinder scene of NeRF-Casting dataset

Evaluation on toaster scene of NeRF-Casting dataset

On datasets with complex reflections, our method achieves much faster rendering speed and fewer Gaussian primitives than EnvGS, while maintaining high image quality that is better than other 3DGS baselines.

Qualitative comparison on Mip NeRF 360, Tanks & Temples and Deepblending datasets.

Comparison on datasets without complex reflections 1 Comparison on datasets without complex reflections 2

Evaluation on bonsai scene of Mip NeRF 360 dataset

On datasets without complex reflections, our method achieves much faster rendering speed and fewer Gaussian primitives than EnvGS, with competitive image quality compared with other 3DGS baselines.

Gaussian density comparison with EnvGS.

Our method uses significantly fewer Gaussian primitives than EnvGS, demonstrating the effectiveness of our reflection-sensitive Gaussian pruning.

Quantitative comparison on Ref-NeRF and NeRF-Casting Shiny Scenes datasets.

Methods Ref-NeRF Real Scenes NeRF-Casting Shiny Scenes Rendering Performance
PSNR↑ SSIM↑ LPIPS↓ PSNR↑ SSIM↑ LPIPS↓ FPS↑ speedup GS number↓
3DGS 23.94 0.641 0.265 28.05 0.847 0.191 227.78 1x 3033253
MetaSapiens 24.28 0.629 0.330 28.43 0.845 0.205 1358.98 5.97x 212662
FlashGS 22.93 0.535 0.317 27.10 0.818 0.240 714.97 3.14x 3033253
PUP 3D-GS 23.78 0.644 0.289 28.33 0.858 0.180 653.56 2.87x 606650
EnvGS 24.39 0.660 0.301 30.21 0.869 0.158 15.05 1x 1408937
Ref-Gaussian 23.57 0.638 0.324 25.85 0.810 0.274 65.77 4.37x 592099
Ours 24.40 0.664 0.280 30.28 0.871 0.163 106.98 7.11x 385722

Our method achieves high-quality rendering while being the fastest among tracing methods. More results can be found in our paper.

Citation

This paper has been accepted by ACM Multimedia (ACM MM). The official citation will be updated when the publication details become available.

@inproceedings{liu2026hybridsplat,
  title={HybridSplat: Fast Reflection-baked Gaussian Tracing using Hybrid Splatting},
  author={Chang, Liu and Sichao, Wang and Jianwei, Guo and Shi-Sheng, Huang},
  note={Accepted by ACM Multimedia (ACM MM); official citation information forthcoming}
}