Radiance Meshes for Volumetric Reconstruction

Research Date: 2025-12-03 Source URL: https://arxiv.org/abs/2512.04076 Project Page: https://half-potato.gitlab.io/rm

Reference URLs

Authors and Affiliations

Author	Affiliation
Alexander Mai	UC San Diego
Trevor Hedstrom	UC San Diego
George Kopanas	Google
Janne Kontkanen	Google
Falko Kuester	UC San Diego
Jonathan T. Barron	Google

Summary

Radiance Meshes replace the ellipsoidal Gaussian primitives of 3DGS with constant-density tetrahedral cells generated through Delaunay tetrahedralization. Because tetrahedra decompose into ordinary triangles, this representation plugs directly into existing GPU rasterization and ray-tracing hardware without custom CUDA kernels or splatting shaders. The method achieves rendering speeds 32% higher than 3DGS at 1440p resolution and eliminates the popping and sorting artifacts inherent to Gaussian splatting. Unlike 3DGS, the tetrahedral structure produces watertight meshes by design, meaning surface extraction requires only a density threshold rather than external marching-cubes or neural thresholding post-processing. Training code, a Vulkan desktop renderer, and a WebGPU browser viewer are all publicly available.

The quality tradeoff is modest. On MipNeRF-360 outdoor scenes, Radiance Meshes score 24.38 dB PSNR against 3DGS’s 24.63 dB, a 0.25 dB gap that is difficult to distinguish visually. Indoor scenes show a similar margin (30.61 vs 31.05 dB). Where Radiance Meshes pull ahead is in rendering consistency: exact volume rendering through sorted tetrahedra avoids the view-dependent popping and photometric inconsistencies that splatting introduces through approximate sorting.

Main Analysis

Practical Value Over 3DGS

The central practical advantage of Radiance Meshes is that the output is a standard triangle mesh rather than a proprietary point-cloud format. This distinction has immediate consequences for production workflows:

No mesh extraction step. With 3DGS, converting a splat scene to a usable mesh requires external tools such as SuGaR, 2DGS, or TSDF-based marching cubes. Each introduces its own quality loss, parameter tuning, and compute cost. Radiance Meshes produce a watertight surface by applying a density threshold directly to the tetrahedral cells. The resulting mesh is immediately compatible with standard 3D software (Blender, Houdini, Unity, Unreal) without format conversion.

Hardware rasterizer compatibility. 3DGS relies on custom CUDA-based splatting kernels that limit deployment to NVIDIA GPUs with specific driver versions. Radiance Meshes render through the standard triangle rasterization pipeline, which runs on any GPU with Vulkan, DirectX, or WebGPU support. This broadens deployment from dedicated workstations to laptops, mobile devices, and web browsers.

Elimination of sorting artifacts. 3DGS approximates front-to-back ordering of semi-transparent Gaussians, which produces view-dependent popping when the sort order changes between frames. Radiance Meshes sort tetrahedra using the power of the circumsphere relative to the camera origin, a mathematically exact ordering that produces stable, flicker-free rendering across all viewpoints.

Rendering Performance

Benchmarks on MipNeRF-360 scenes at 1440p resolution demonstrate practical speed gains:

Method	Outdoor PSNR	Indoor PSNR	Outdoor FPS	Indoor FPS
Radiance Mesh	24.38	30.61	240	384
3DGS	24.63	31.05	145	251
RadFoam	23.90	30.66	234	162
EVER	24.73	30.98	36	66

Radiance Meshes achieve 65% higher FPS than 3DGS on outdoor scenes and 53% higher on indoor scenes. The speed advantage comes from two factors: native hardware rasterization avoids the overhead of custom splatting kernels, and the mesh shader implementation reduces redundant memory loads through warp shuffles.

For ray-tracing workloads, Radiance Meshes achieve a 17% speed increase over RadFoam on indoor scenes by using hardware triangle intersection units (RT cores on NVIDIA GPUs) rather than custom ray-primitive intersection tests.

Web and Cross-Platform Deployment

Three rendering implementations ship with the project, each targeting a different deployment scenario:

Vulkan desktop renderer (vkrm). The fastest option, using mesh shaders for maximum throughput on modern desktop GPUs. Suitable for workstation viewing and integration into desktop applications.

WebGPU browser viewer (webrm). Runs in any browser with WebGPU support (Chrome 113+, Edge 113+, Firefox Nightly). Interactive demos for all MipNeRF-360 scenes (bicycle, garden, flowers, treehill, room, kitchen, counter, bonsai, corsair) are available through the project page. This is a direct path to embedding radiance field scenes in web pages without plugins or WASM splatting viewers.

Instanced rendering fallback. For older hardware that lacks mesh shader support, an instanced rendering path provides compatibility at reduced performance.

The WebGPU path is particularly relevant for production use cases. Current 3DGS web viewers (GaussianSplats3D, antimatter15/splat, PlayCanvas SuperSplat) require custom WebGL-based splatting implementations. Radiance Meshes render through WebGPU’s standard triangle pipeline, potentially simplifying maintenance and improving compatibility as WebGPU matures.

Physics Simulation Compatibility

Because the representation is a volumetric tetrahedral mesh, it is directly compatible with physics engines that operate on tetrahedral grids. The paper demonstrates integration with Extended Position Based Dynamics (XPBD), a simulation framework used in game engines and VFX tools for soft-body dynamics, cloth, and deformable objects.

This is a qualitative difference from 3DGS. Gaussian splats have no inherent spatial connectivity: each Gaussian is an independent primitive with no neighbor relationships. Simulating physics on splats requires first extracting a mesh (losing information) or building adjacency structures externally. Radiance Meshes carry connectivity information natively through the Delaunay tetrahedralization, so physics simulation can operate directly on the scene representation without format conversion.

Potential applications include:

Deformable object simulation in reconstructed scenes
Collision detection using the watertight surface mesh
Real-time interaction with volumetric scene elements in VR/AR
Training data generation for robotics (similar to the SplatSim approach, but without the mesh extraction bottleneck)

Fisheye and Non-Standard Camera Support

The circumsphere-based sorting algorithm works with any camera projection model, not just the pinhole cameras that 3DGS assumes. The paper demonstrates rendering with fisheye lens distortion, which is relevant for:

Autonomous driving datasets captured with wide-angle cameras
VR headset rendering with barrel distortion
Surveillance and security camera footage with lens distortion
Panoramic and 360-degree capture workflows

3DGS sorting relies on depth ordering relative to a planar image, which breaks down with non-planar projections. The power-of-circumsphere metric used by Radiance Meshes is projection-independent.

Watertight Mesh Extraction

Surface extraction from Radiance Meshes applies a density threshold to tetrahedra and outputs the boundary faces of high-density cells. The resulting mesh is watertight by construction because every internal face between a high-density and low-density tetrahedron becomes a surface triangle, and the Delaunay tetrahedralization guarantees that these faces form a closed manifold.

This contrasts with mesh extraction from 3DGS, which typically involves:

Converting Gaussians to a density field (lossy)
Running marching cubes or Poisson reconstruction (parameter-dependent)
Post-processing to close holes and remove artifacts

The direct extraction path from Radiance Meshes removes two of these three steps and eliminates the associated quality degradation.

Training and Memory Characteristics

Training runs on a single RTX 4090. The Delaunay recomputation step (updating the tetrahedralization after vertex movement) takes approximately 143ms per invocation on scenes with 1.2 million tetrahedra. This recomputation runs every 10 optimization iterations. The method supports approximately 15 million tetrahedra on an RTX 4090, compared to RadFoam’s limit of approximately 3.3 million Voronoi cells on the same hardware.

Densification uses two criteria: SSIM-based splitting targets high-error regions, while total variance splitting detects under-represented thin structures. New points are inserted at the approximate intersection of rays from the two highest-error views rather than by uniform sampling within existing tetrahedra.

Limitations

Quality ceiling. Peak PSNR on MipNeRF-360 is approximately 0.25 dB below 3DGS on outdoor scenes and 0.44 dB below on indoor scenes. For applications where visual fidelity is the primary metric (film VFX, archival documentation), this gap may matter. For real-time applications where rendering speed and mesh compatibility are more important, the tradeoff favors Radiance Meshes.

Training speed. The paper does not report training time comparisons, but the periodic Delaunay recomputation adds overhead relative to 3DGS training, which requires no such global topology update.

Ecosystem maturity. 3DGS has a two-year head start in tooling: commercial services (Luma AI, Polycam, Kiri Engine), Houdini plugins (GSOPs), game engine integrations (Unity, Unreal), and extensive community documentation. Radiance Meshes are newly released with three GitHub repositories and no third-party tooling yet.

Key Findings

Radiance Meshes render 32% faster than 3DGS at 1440p by using native hardware triangle rasterization instead of custom splatting kernels
The tetrahedral representation produces watertight meshes without post-processing, eliminating the mesh extraction pipeline required by 3DGS
Three open-source rendering implementations cover desktop (Vulkan), web (WebGPU), and legacy (instanced rendering) deployment
Direct compatibility with physics simulation frameworks (XPBD) enables deformable object interaction without format conversion
Quality tradeoff is modest: 0.25-0.44 dB lower PSNR than 3DGS, offset by elimination of sorting artifacts and popping
The circumsphere sorting algorithm supports non-pinhole cameras (fisheye, panoramic) that 3DGS cannot handle correctly

References

Radiance Meshes for Volumetric Reconstruction (arXiv 2512.04076) - December 2025
Project Page - December 2025
Training Code Repository - December 2025
Vulkan Renderer Repository - December 2025
WebGPU Viewer Repository - December 2025
Radiance Fields Blog Post - December 2025
3D Gaussian Splatting (Kerbl et al., SIGGRAPH 2023) - August 2023