X-Gaussian: Radiance meets Radiation

Traditionally, when we discuss radiance fields, we focus on radiance itself. However, today there's a subtle yet significant shift towards radiation—yes, that kind of radiation.

Traditionally, when we discuss radiance fields, we focus on radiance itself. However, today there's a subtle yet significant shift towards radiation—yes, that kind of radiation.

Michael Rubloff

Michael Rubloff

Michael is the Founder and Managing Editor of Radiancefields.com

X Gaussian

The array of use cases for radiance fields is quite staggering. Traditionally, when we discuss radiance fields, we focus on radiance itself. However, today there’s a subtle yet significant shift towards radiation—yes, that kind of radiation. It’s not surprising that we’re venturing into an industry that we’ve only begun to explore: healthcare.

While previous radiance methods have ventured into applications like surgical reconstruction, our current focus is on X-rays, specifically through an innovation called “X-Gaussian.” This approach utilizes 3D Gaussian splatting to significantly enhance the efficiency and quality of X-ray novel view synthesis.

By enhancing the tools available to physicians, we invariably increase the likelihood of receiving better medical care.

X-Gaussian transcends traditional limitations by utilizing a radiative Gaussian point cloud model, which uniquely ignores the viewing direction when predicting radiation intensity. This model, inspired by the isotropic nature of X-ray imaging, represents a departure from methods not well-suited to X-ray’s distinctive penetration and attenuation characteristics.

This model incorporates a Radiation Intensity Response Function (RIRF)—a novel concept that replaces the spherical harmonics typically used in RGB domain 3DGS. The RIRF facilitates the prediction of radiation intensity at 3D points using a set of learnable parameters, effectively simulating the X-ray interaction with object materials independent of viewing direction. This strategic shift not only aligns with X-ray imaging’s physical properties but also streamlines the computational process, resulting in faster inference and reduced training times.

A standout innovation within X-Gaussian is the Differentiable Radiative Rasterization (DRR) with CUDA implementation, enabling the efficient rendering of novel X-ray projections. DRR begins by assigning radiation intensity values to each point in the cloud, based on the novel Radiation Intensity Response Function (RIRF). These points are then meticulously rasterized onto a 2D plane, ensuring that the essential characteristics of X-ray penetration and attenuation are preserved.

What sets DRR apart is its ability to perform this complex rendering in a differentiable manner, allowing for the seamless back-propagation of gradients during the optimization phase. This critical feature enables X-Gaussian to learn and adapt with unprecedented efficiency, refining its understanding of the object’s radiative landscape with each iteration.

Additionally, the X-Gausssian k introduces an Angle-pose Cuboid Uniform Initialization (ACUI) strategy for camera calibration. Leveraging the geometric parameters of the X-ray scanner, ACUI accurately calculates the camera’s intrinsic and extrinsic matrices, bypassing conventional structure-from-motion methods like COLMAP. This method not only optimizes the initialization process but also markedly decreases training duration.

The ACUI strategy employs a cuboid encompassing the object of interest, within which it uniformly samples initial points for the Gaussian point clouds. This method provides a more structured and efficient starting point for the optimization process, significantly cutting down the time required for training without sacrificing the quality of the synthesized views.

X-Gaussian has been evaluated against state-of-the-art methods, showcasing its superior performance with a notable improvement of 6.5 dB in image quality while achieving more than 73× faster inference speed and requiring less than 15% of the training time. Additionally, the framework’s application in sparse-view CT reconstruction has highlighted its practical utility in medical imaging, promising a future where comprehensive imaging is possible with minimal X-ray exposure.

What this all means in practice is 100+ fps, with training times under 10 minutes, but still outputting high fidelity models.

This groundbreaking study not only advances the field of X-ray imaging but also sets a new benchmark for efficiency and quality in novel view synthesis. X-Gaussian represents a significant step forward in harnessing the power of 3D Gaussian Splatting for medical imaging, offering a promising avenue for future research and application in clinical settings.

The potential of radiance fields is bound only by the speed at which various industries recognize and leverage this technology. In all scenarios, it is us, the end-users, who stand to gain from access to superior innovations. The code for X-Gaussian can be found here.

Thumbnail for Gaussian Splatting News episode 13

13:12

Watch · New every Monday

The Biggest Week Yet for Gaussian Splatting: Blender + Unreal

Episode 13 · Nov 16, 2026

Blender 5.3 is out, and Gaussian splats are now a native part of it: PLY and SPZ files import as their own point cloud type and render in EEVEE, Cycles and Workbench, with no add-on.

Featured

Recents

Platforms

Esri Makes ArcGIS Reality Server Generally Available

ArcGIS Reality Server is generally available with ArcGIS Enterprise 12.2, adding self-hosted, distributed processing of drone imagery into Gaussian splats.

Michael Rubloff

Nov 18, 2026

Platforms

Esri Makes ArcGIS Reality Server Generally Available

ArcGIS Reality Server is generally available with ArcGIS Enterprise 12.2, adding self-hosted, distributed processing of drone imagery into Gaussian splats.

Michael Rubloff

Platforms

SuperSplat 3.9 Imports GSAF 4D Splats as Timeline Sequences

SuperSplat v3.9.0 imports GSAF 4D Gaussian splat files as timeline sequences, plays several sequences at once and adds a review table for multi-file imports.

Michael Rubloff

Nov 17, 2026

Platforms

SuperSplat 3.9 Imports GSAF 4D Splats as Timeline Sequences

SuperSplat v3.9.0 imports GSAF 4D Gaussian splat files as timeline sequences, plays several sequences at once and adds a review table for multi-file imports.

Michael Rubloff

Platforms

CARLA 1.0 Adds NVIDIA NuRec 26.04 Support

CARLA 1.0 adds NVIDIA NuRec 26.04 support, turning recorded real-world drives into simulated scenes with neural lidar, asset swapping and actor compositing.

Michael Rubloff

Nov 17, 2026

Platforms

CARLA 1.0 Adds NVIDIA NuRec 26.04 Support

CARLA 1.0 adds NVIDIA NuRec 26.04 support, turning recorded real-world drives into simulated scenes with neural lidar, asset swapping and actor compositing.

Michael Rubloff

News

Volumetric Media Town Hall Turns to 4D Gaussian Splats

The Metaverse Standards Forum holds its third Volumetric Media Town Hall on November 24, with 4D Gaussian splatting talks by Dymensium, Chronospace and Gracia.

Michael Rubloff

Nov 17, 2026

News

Volumetric Media Town Hall Turns to 4D Gaussian Splats

The Metaverse Standards Forum holds its third Volumetric Media Town Hall on November 24, with 4D Gaussian splatting talks by Dymensium, Chronospace and Gracia.

Michael Rubloff

Trending Articles
  1. TRENDING
    Loading...
  2. TRENDING
    Loading...
  3. TRENDING
    Loading...
Michael Rubloff

Written by Michael Rubloff

Michael is the Founder and Managing Editor of Radiancefields.com

More from Michael Rubloff

More from Michael Rubloff

Platforms

Esri Makes ArcGIS Reality Server Generally Available

ArcGIS Reality Server is generally available with ArcGIS Enterprise 12.2, adding self-hosted, distributed processing of drone imagery into Gaussian splats.

Michael Rubloff

Nov 18, 2026

An Esri graphic showing a 3D reality-mapped city with a diagram of ArcGIS Pro sending processing from ArcGIS Enterprise to ArcGIS Reality Server

Platforms

Esri Makes ArcGIS Reality Server Generally Available

Michael Rubloff

Nov 18, 2026

An Esri graphic showing a 3D reality-mapped city with a diagram of ArcGIS Pro sending processing from ArcGIS Enterprise to ArcGIS Reality Server

Platforms

Esri Makes ArcGIS Reality Server Generally Available

ArcGIS Reality Server is generally available with ArcGIS Enterprise 12.2, adding self-hosted, distributed processing of drone imagery into Gaussian splats.

Michael Rubloff

Nov 18, 2026

An Esri graphic showing a 3D reality-mapped city with a diagram of ArcGIS Pro sending processing from ArcGIS Enterprise to ArcGIS Reality Server

Platforms

SuperSplat 3.9 Imports GSAF 4D Splats as Timeline Sequences

SuperSplat v3.9.0 imports GSAF 4D Gaussian splat files as timeline sequences, plays several sequences at once and adds a review table for multi-file imports.

Michael Rubloff

Nov 17, 2026

The SuperSplat editor in a browser window, editing a Gaussian splat of a snow-covered cabin in a pine forest

Platforms

SuperSplat 3.9 Imports GSAF 4D Splats as Timeline Sequences

Michael Rubloff

Nov 17, 2026

The SuperSplat editor in a browser window, editing a Gaussian splat of a snow-covered cabin in a pine forest

Platforms

SuperSplat 3.9 Imports GSAF 4D Splats as Timeline Sequences

SuperSplat v3.9.0 imports GSAF 4D Gaussian splat files as timeline sequences, plays several sequences at once and adds a review table for multi-file imports.

Michael Rubloff

Nov 17, 2026

The SuperSplat editor in a browser window, editing a Gaussian splat of a snow-covered cabin in a pine forest

Platforms

CARLA 1.0 Adds NVIDIA NuRec 26.04 Support

CARLA 1.0 adds NVIDIA NuRec 26.04 support, turning recorded real-world drives into simulated scenes with neural lidar, asset swapping and actor compositing.

Michael Rubloff

Nov 17, 2026

CARLA

Platforms

CARLA 1.0 Adds NVIDIA NuRec 26.04 Support

Michael Rubloff

Nov 17, 2026

CARLA

Platforms

CARLA 1.0 Adds NVIDIA NuRec 26.04 Support

CARLA 1.0 adds NVIDIA NuRec 26.04 support, turning recorded real-world drives into simulated scenes with neural lidar, asset swapping and actor compositing.

Michael Rubloff

Nov 17, 2026

CARLA

New to Gaussian Splatting? Start here

Radiance fields, explained from the ground up: what a Gaussian splat is, how to capture one, and which tools and gear to start with.

Newsletter

Keep up with the field.

New research, tools and captures from the Radiance Fields newsroom, in your inbox. Free, no spam.

Gaussian Splatting News on YouTube

The week in radiance fields, every Monday.

Gaussian Splatting News on YouTube

The week in radiance fields, every Monday.

Gaussian Splatting News on YouTube

The week in radiance fields, every Monday.

Gaussian Splatting News on YouTube

The week in radiance fields, every Monday.