Gaussian Splatting: Real-Time Photorealistic 3D Makes Its Way into Industry

3D representation of the real world has always been a challenge for stakeholders in Industry 4.0, simulation, and digital twins. For decades, available solutions (photogrammetry, polygon meshes, LiDAR) have required compromises between visual quality, accuracy, and real-time performance.

An emerging technology is changing the game: Gaussian Splatting, or 3D Gaussian Splatting (3DGS). Developed in research labs in 2023, it is gradually establishing itself as a major breakthrough in how three-dimensional environments are captured, represented, and visualized.

1. What is Gaussian Splatting? Back to Basics

To understand Gaussian Splatting, one must first understand its predecessors.

Photogrammetry, the oldest of the three techniques, reconstructs a 3D model from a series of photographs taken from different angles. It produces geometrically accurate, measurable models that are easy to share. However, it struggles to accurately reproduce transparent, reflective, or textureless surfaces, and the results can suffer from visual artifacts on complex geometries (Teleport/Varjo, 2024).

Neural Radiance Fields (NeRF), introduced in 2020, represented a major breakthrough. They enable the generation of photorealistic environments from multi-angle photos by modeling light as it travels through space via a neural network. The rendering is of remarkable visual quality, even on complex surfaces. Their main limitation remains computational cost: training a NeRF model can take several hours, and rendering is generally not real-time, making it impractical for many professional applications (Clarté / augmented-reality.fr, 2025).

Gaussian Splatting thus emerges as a hybrid and more accessible alternative. Presented in August 2023 in a research paper by INRIA (the French National Institute for Research in Computer Science and Control) and the University of Tübingen, it is based on a fundamentally different idea: rather than modeling the scene using polygons or a continuous neural network, the scene is represented as a dense cloud of “Gaussian blurs” (semi-transparent ellipsoids, each defined by its position, size, rotation, color, and opacity (Kerbl et al., ACM Transactions on Graphics, 2023)). Thanks to this technique, we obtain a continuous, fluid, and photorealistic volumetric rendering, without meshing or traditional texturing, and in real time.

The name itself is evocative: “splatting” refers to the act of “projecting” these ellipsoids onto the screen, like splashes of light that overlap to reconstruct perceived reality. Gaussian Splatting does not reconstruct an object: it reconstructs the perception of light (Widdim, 2025).

2. How does it work? The three-step pipeline

The process of generating a Gaussian Splat follows a relatively straightforward pipeline, even though the underlying calculations are complex.

Step 1 - Capture. A set of images or videos is captured around the target scene or object. This can be done using a simple smartphone, a camera, a drone, or a LiDAR scanner for applications requiring greater precision.

Step 2 - Initialization and Optimization. Using the images, a Structure-from-Motion (SfM) algorithm, typically COLMAP, first generates a sparse 3D point cloud. Each point in this cloud serves as the starting point for a Gaussian. A machine learning optimization process then adjusts the position, shape, color, and opacity of each Gaussian so that their 2D projection best matches the input images. This is known as differentiable Gaussian rasterization (Escadrone, 2025).

Step 3 - Real-time rendering. Once optimized, the millions of Gaussians are projected and merged directly on the GPU to produce an image. This process is extremely fast: whereas a NeRF required hours of computation for rendering, a Gaussian Splatting model can be displayed and explored in a matter of seconds, at high frame rates (Widdim, 2025; Artlight, 2026).

3. Gaussian Splatting vs. NeRF vs. Photogrammetry: Which One Should You Choose?

These three technologies are not in direct competition; they address distinct needs. In summary: photogrammetry remains the gold standard for applications requiring certified metrological accuracy (quality control, industrial metrology). NeRF excels at very high-quality rendering where computation time is not a constraint. Gaussian Splatting stands out as the best compromise for real-time photorealistic rendering, with the ability to capture complex surfaces that photogrammetry struggles to handle (Escadrone, 2025; Synima, 2025).

4. Practical Applications: How Gaussian Splatting Is Transforming the Industry

Digital twins and facility documentation

Gaussian Splatting ushers in a new era for the creation of photorealistic digital twins. By capturing an industrial site (a factory, a warehouse, an energy infrastructure) with a camera or a drone, it becomes possible to generate a navigable visual twin in real time, with unprecedented visual fidelity. Where photogrammetry may leave surfaces incomplete or textures rough, Gaussian Splatting renders vegetation, fine structures, and reflective materials with remarkable perceptual accuracy (Escadrone, 2025). Major players have already integrated the technology into their tools.

Industrial Inspection and Maintenance

Gaussian Splatting is particularly well-suited for inspecting complex infrastructure: electrical substations, refineries, civil engineering structures, and wind turbines. When paired with drones, it enables the rapid generation of navigable representations of a site that can be used directly for anomaly detection or maintenance planning. Bentley Systems has notably demonstrated the use of Gaussian Splatting to visualize electrical substations and geospatial sites in CesiumJS, with a level of immersion and fluidity that traditional meshes could not achieve (Khronos Group / Bentley Systems, 2026).

Virtual Reality and Immersive Training

For training applications in VR environments, Gaussian Splatting offers the ability to recreate real industrial environments—rather than just modeled ones—with an unprecedented level of photorealism. Operators can train in a faithful reproduction of their actual work site, which enhances the effectiveness of the training.

Robotics and Autonomous Navigation

Robotics research is also embracing Gaussian Splatting for dense environment representation. Its ability to provide rich, actionable visual maps in real time makes it a strong candidate for autonomous robot navigation, path planning, and scene recognition (Zhu et al., “3D Gaussian Splatting in Robotics: A Survey,” arXiv, 2024).

5. Toward Standardization: The Pivotal Moment of 2026

One of the main obstacles to the widespread adoption of Gaussian Splatting has long been the lack of a standard exchange format. Each tool used its own proprietary formats (PLY, SPZ, etc.), making it difficult to transfer scenes between 3D engines, web platforms, and industrial pipelines.

This barrier is now being broken down. In February 2026, the Khronos Group (the consortium behind the OpenGL, Vulkan, and glTF standards) announced a release candidate for the KHR_gaussian_splatting extension for glTF 2.0, the most widely adopted 3D delivery format on the web and in real-time engines. This extension defines how the attributes of a Gaussian Splat (position, orientation, scale, color, opacity) are stored in a standard glTF file (Khronos Group, February 2026).

Final ratification is expected in the second quarter of 2026, with support from Google, NVIDIA, Apple, Adobe, Autodesk, Bentley Systems, and Esri. Neil Trevett, president of the Khronos Group, called this extension a “major milestone for the 3D community” (Khronos Group, 2026).

For industry professionals, this is a game-changer: a photorealistic scene captured from images will soon be able to be encapsulated in a standard glTF file, manipulable by the same tools as traditional 3D models. The built-in “fallback” mechanism also ensures minimum compatibility: when software does not yet support the extension, it can display the content as a point cloud (Leaxea, 2026).

6. Limitations to Be Aware Of

Gaussian Splatting is a powerful technology, but there are certain limitations to consider before integrating it into an industrial pipeline:

Limited geometric accuracy. The technology is optimized for visual quality, not metrological accuracy. Average geometric measurements have an error of around 7.8 cm, which is insufficient for strict dimensional inspection applications. For these uses, LiDAR remains essential as a complementary technology (The Future 3D, 2026).
File size. A Gaussian Splatting model can be hundreds of megabytes, or even several gigabytes, which poses challenges for storage, streaming, and mobile compatibility. Compression efforts (notably Niantic Spatial’s SPZ format, which reduces file size by 90% compared to PLY) are currently being standardized (Khronos/OGC, 2025).
Artifacts on certain geometries. Long linear structures (cables, antennas, rails) tend to produce visual artifacts that are difficult to eliminate without significantly increasing file sizes (Khronos/OGC, 2025).
Editing and physical interactivity. Unlike a mesh, a Gaussian Splat is not naturally “editable” or subject to physics. Recent work is exploring the integration of physical dynamics to enable interactions in VR, but these approaches remain experimental.
Expertise and tools are still evolving. The processing workflow (capture, optimization, visualization) still requires specialized skills and dedicated software (Nerfstudio, PostShot, SuperSplat, DJI Terra, etc.). The ecosystem is rapidly taking shape, but there is still a significant learning curve.

Conclusion

In less than three years, Gaussian Splatting has evolved from a research paper to a technology integrated into tools from Autodesk, Esri, and DJI, and will soon be part of the glTF standard itself. Its ability to generate photorealistic 3D environments that can be navigated in real time, using simple photos or videos, makes it a strategic asset for players in Industry 4.0, digital twins, augmented reality, and immersive training.

The standardization process underway in 2026 signals that the technology is transitioning from the experimental stage to large-scale professional deployment. For manufacturers, the question is no longer whether Gaussian Splatting will become mainstream, but how to prepare for it today, in terms of data capture, 3D pipelines, and internal expertise.

Sources

Kerbl B. et al., 3D Gaussian Splatting for Real-Time Radiance Field Rendering, INRIA / Université de Tübingen, ACM Transactions on Graphics, 2023 — repo-sam.inria.fr
Clarté / augmented-reality.fr — À la découverte du Gaussian Splatting et de ses usages avec Grégory Duvalet, septembre 2025
Widdim — 3D Gaussian Splatting : la nouvelle révolution de la visualisation 3D, octobre 2025
Artlight — Gaussian Splatting : de la photogrammétrie au rendu temps réel, avril 2026 — artlight.fr
Escadrone — Gaussian Splatting : impact et intégration dans DJI Terra, novembre 2025
Synima — The Differences between Photogrammetry, NeRF and Gaussian Splatting, 2025
Teleport / Varjo — Gaussian splatting vs. photogrammetry vs. NeRFs, 2024
arcorama.fr — Gaussian Splats : une nouvelle technologie de rendu 3D pour votre SIG, décembre 2025
AEC Magazine — Introducing Gaussian Splats for AEC, février 2026
Khronos Group — Khronos Announces glTF Gaussian Splatting Extension (KHR_gaussian_splatting), février 2026 — khronos.org
Khronos Group / OGC — Khronos, OGC, and Geospatial Leaders Add 3D Gaussian Splats to the glTF Asset Standard, août 2025
Bentley Systems — Why an Open Standard For Gaussian Splats Changes 3D Modeling, février 2026
Leaxea — 3D Gaussian Splatting arrive dans glTF, mars 2026
The Future 3D — The State of Gaussian Splatting in 2026: Standards and Tools, avril 2026
Tu L. et al. (VRSplat) — Fast and Robust Gaussian Splatting for Virtual Reality, arXiv, mai 2025 — arxiv.org/abs/2505.10144
Zhu S. et al. — 3D Gaussian Splatting in Robotics: A Survey, arXiv, 2024 — arxiv.org/abs/2410.12262
Fan S. et al. — A review of recent advances in Gaussian splatting, Applied Intelligence / Springer Nature, 2026

A Better Understanding of Gaussian Splatting