Volumetrically Consistent 3D Gaussian Rasterization

University of California San Diego
CVPR 2025 (Highlight)

Abstract

Recently, 3D Gaussian Splatting (3DGS) has enabled photorealistic view synthesis at high inference speeds. However, its splatting-based rendering model makes several approximations to the rendering equation, reducing physical accuracy. We show that the core approximations in splatting are unnecessary, even within a rasterizer; we instead volumetrically integrate 3D Gaussians directly to compute the transmittance across them analytically. We use this analytic transmittance to derive more physically-accurate alpha values than 3DGS, which can directly be used within their framework. The result is a method that more closely follows the volume rendering equation (similar to ray-tracing) while enjoying the speed benefits of rasterization. Our method represents opaque surfaces with higher accuracy and fewer points than 3DGS. This enables it to outperform 3DGS for view synthesis (measured in SSIM and LPIPS). Being volumetrically consistent also enables our method to work out of the box for tomography. We match the state-of-the-art 3DGS-based tomography method with fewer points.

Intuition behind our rasterizer

The core idea of our work, is that instead of splatting the 3D Gaussians to 2D, we analytically compute the transmittance of a 3D Gaussian along the camera ray, and use that to compute alpha values for blending Gaussians in screen space. (Left)
Our approach alleviates many of the approximations in the splatting process, making the 3DGS forward rendering process more consistent with the volume rendering equation. This can be easily incorporated into a rasterizer, thus retaining the training and inference speed of 3DGS. Visually, our approach allows rendering opaque surfaces better than a splatted 3D Gaussian, as shown below for a 2D example (Right). See figure 3 in the paper for more details
Method overview image

Interactive 1D Density Plot

We show 1D cross sections of a single 3D Gaussian placed \(z_0\) distance from the camera along the camera ray. Increasing the scale \(s_z\) along the camera ray and density \(k\), both increase the flat region where the opacity \(\alpha=1\). This allows our method represent opaque surfaces better than 3DGS, which always has a Gaussian-like opacity fall off.

Visual Comparison

Compared to 3DGS, our method improves rendering opaque surfaces and sharp details like edges.
Ours
3DGS
Ours
3DGS
Ours
3DGS
Ours
3DGS

Related Links

A lot of concurrent work has been done in this space, exploring different 3D scene primitives, as well as studying the effect of approximations in 3DGS in simple settings.

Does 3D Gaussian Splatting Need Accurate Volumetric Rendering? studies the ray-marching 3D Gaussians in comparison with 3DGS in simple settings without densification.

EVER: Exact Volumetric Ellipsoid Rendering for Real-time View Synthesis uses constant density ellipsoids as primitives and ray-marches through them for rendering.

Many more works might have surfaced by the time you read this.

BibTeX

@article{talegaonkar2025vol3dgs,
  author    = {Chinmay Talegaonkar and Yash Belhe and Ravi Ramamoorthi and Nicholas Antipa},
  title     = {Volumetrically Consistent 3D Gaussian Rasterization},
  journal   = {CVPR},
  year      = {2025},
}