Introduction

Somewhere between a photograph and a hologram, a new medium of spatial representation is taking shape. It does not look like the wireframe models of early Computer-Generated Imagery (CGI), or the painstaking polygon meshes of modern game engines, or the slow-burning neural reconstructions that briefly captured the industry's imagination in the early 2020s. It looks, with startling fidelity, like the world itself.

That is the promise of Gaussian Splatting: the ability to take a set of ordinary photos or video frames of any real environment and reconstruct it as an explorable, photorealistic 3D scene, in real time, on consumer hardware. Two years after the technique's publication at SIGGRAPH 2023, it has cleared the proof-of-concept stage and entered commercial deployment across industries ranging from surgery to cinema. More than 665 job listings requiring Gaussian splatting skills have appeared since 2024, and dedicated startups in the space have raised over $480 million in disclosed capital. The question now is not whether this technology will reshape 3D content creation, but how fast and how broadly.

What Is Gaussian Splatting?

To understand Gaussian Splatting, it helps to understand what came before it and why it fell short.

For decades, 3D scenes in games, films, and simulations were built by hand: artists modelling every surface as a mesh of triangles, painting textures onto them, and setting up lighting rigs. This process produces excellent results but is slow, expensive, and limited to synthetic content. Capturing the real world in 3D required expensive laser scanners or hours of processing time, and the outputs rarely matched the visual quality of a photograph.

In 2020, Neural Radiance Fields (NeRF) appeared to change this. NeRF could take a set of photographs and learn an implicit neural representation of the scene that enabled photorealistic novel view synthesis. It was remarkable, but it had a critical flaw: rendering was slow, often taking seconds or minutes per frame, making real-time use impossible. It was also computationally expensive to train and difficult to edit after the fact.

3D Gaussian Splatting solved the rendering bottleneck. Instead of a neural network queried per ray, it represents a scene as a cloud of millions of small 3D ellipsoids, each called a Gaussian, each defined by its position, size, orientation, colour, and opacity. When rendered, these ellipsoids are projected onto the screen as soft, overlapping 2D splats that combine to produce a photorealistic image. Because this projection is a rasterisation operation rather than a neural computation, it is orders of magnitude faster. The original 2023 paper demonstrated real-time rendering at 30 frames per second or above, a threshold no prior photorealistic reconstruction method had reached.

Key Definitions

1. 3D Gaussian Splatting (3DGS)

A technique for representing real-world scenes as millions of 3D ellipsoidal primitives (Gaussians). Each primitive encodes position, shape, colour, and opacity. Together, they can be rapidly projected onto a 2D image plane to produce photorealistic views from any angle in real time.

2. Novel View Synthesis

The task of generating a realistic image of a scene from a camera angle that was not present in the original set of training photographs. It is the core capability enabled by both NeRF and 3DGS.

3. Neural Radiance Field (NeRF)

The predecessor technology to 3DGS. NeRF represents a scene implicitly inside a neural network and produces high-quality novel views, but at rendering speeds too slow for real-time use.

4. 4D Gaussian Splatting (4DGS) 

An extension of 3DGS that adds a temporal dimension, enabling the representation of dynamic, moving scenes rather than static environments. Critical for video reconstruction and real-world simulation.

5. Splat File (PLY / SPZ)

The file format in which a Gaussian splatting scene is stored. A standard PLY file for a scene can reach 250 MB or more; the SPZ format developed by Niantic compresses this to roughly 25 MB, making web and mobile delivery practical.

2025: The Breakout Year

The original 3DGS paper was a proof of concept. What has happened since is a co-evolution of the research base, the tooling ecosystem, and the commercial pipeline that transforms a proof of concept into a platform.

1. The Research Explosion

The pace of follow-on research has been extraordinary. RadianceFields.com indexes over 3,333 radiance field research papers, with at least 749 carrying 2026 publication dates through July of that year, a rate of more than four new papers per day. Key advances include Mip-Splatting, which resolved aliasing artefacts in rendered outputs; 2D Gaussian Splatting (SIGGRAPH 2024), which improved geometric accuracy; and 4D Gaussian Splatting, which extended the technique to dynamic scenes. Compression research has produced tools like HAC++, which targets up to 100 times compression of scene files, and SPZ, which delivers roughly 90% file size reduction in practice, making streaming and mobile delivery viable.

2. Standardisation

Perhaps the clearest signal that a technology has graduated from research to infrastructure is the formation of standards bodies around it. In February 2026, the Khronos Group announced the KHR_gaussian_splatting extension for glTF 2.0. Khronos is the consortium behind OpenGL, Vulkan, and WebGL; its endorsement means that Gaussian splats will be a first-class citizen in the same 3D data pipeline used by every major web browser, game engine, and CAD application. MPEG has opened a parallel Explorations track on Gaussian Splat Coding, signalling that formal broadcast standards may follow.

3. Tooling and Accessibility

In 2023, creating a Gaussian splat required a working knowledge of Python, access to a high-end GPU, and significant patience. By 2025, over 212 platforms and tools are on the market, from smartphone apps (Polycam, Luma AI) that process a scene from a phone scan, to enterprise renderers integrated into Esri's ArcGIS Pro and DJI's Terra drone-mapping software. The democratisation of the capture-to-splat pipeline is accelerating adoption across organisations that have no computer graphics expertise.

Deployment Across Industries

1. Media, Entertainment, and Advertising

The earliest and most visible commercial deployment has been in film and advertising. In 2025, the production of the film Superman became the first major motion picture to use dynamic Gaussian splatting in its visual effects pipeline. At the 2025 BAFTAs, photographer Harry Nelder captured over 40 award winners and presenters as Gaussian splats, using the format as a new medium for celebrity portraiture. A broadcast-quality advertising campaign produced in Australia demonstrated that splat-based production is now viable as a commercial workflow. For brands, the technology offers a structural change in cost: a single real-world scan can yield infinite camera angles, product configurations, and lighting conditions, replacing multi-day studio shoots with a single capture session.

The market data reflects this momentum. Media and entertainment commanded $1.16 billion of the $2.8 billion 3DGS market in 2025, representing 41.3% of total revenue. Film studios, streaming platforms, and game developers are making strategic investments to reduce post-production costs and accelerate content timelines.

2. Healthcare and Surgery

In clinical settings, Gaussian Splatting is emerging as a visualisation layer for complex anatomy. The technology's ability to blend multiple imaging modalities, including CT scans, MRI, and ultrasound, into a single real-time explorable 3D representation is allowing surgeons to plan procedures with a spatial understanding that 2D medical images do not provide. Medical institutions have reported reduced surgical complications through immersive 3D visualisation during complex procedures.

The surgical robotics research community is further ahead. The Instrument-Splatting framework uses Gaussian Splatting to build real-time digital twins of surgical instruments and tissue, enabling robotic systems to reason about the physical state of a surgical field. This real-to-simulation-to-real pipeline, where a robot learns in a Gaussian splatting environment and transfers that learning to the operating theatre, is one of the most promising directions in embodied AI.

3. Robotics and Autonomous Systems

Autonomous navigation systems need rich, updatable world models. Gaussian Splatting is attractive for this application because it combines dense photorealistic reconstruction with fast rendering, allowing a robot to reason about the appearance of its environment at interactive frame rates. Frameworks like SplaTAM and GS-SLAM adapt splats to online mapping and tracking, enabling robots to build and navigate persistent, photorealistic maps of their environments in real time.

In autonomous vehicle development, Gaussian Splatting enables photorealistic simulation of real-world driving scenarios captured from sensor data, accelerating training and validation without requiring exhaustive real-world mileage. Wayve, the autonomous driving company, has integrated splatting-based simulation into its development stack and has raised over $1 billion in funding for its approach.

4. Geospatial, Infrastructure, and Digital Twins

The geospatial industry is integrating Gaussian Splatting rapidly. Esri has announced that ArcGIS Pro will support Gaussian splatting as a native layer type, expanding applications from infrastructure inspection to smart city digital twins. DJI Terra already offers Gaussian splatting output from drone data, enabling faster and higher-quality 3D reconstructions of construction sites, utility corridors, and urban areas. The 3D scanning market is projected to reach $19 to $22 billion by 2033 to 2035, and Gaussian Splatting is positioned as the primary driver of that growth.

Current Limitations and Active Research Frontiers

Gaussian Splatting is not without significant unsolved problems. Understanding these limitations is essential for any organisation planning to deploy the technology in production.

1. Memory and Storage Footprint

A typical 3DGS scene requires millions of Gaussian primitives, each carrying dozens of parameters. The resulting storage and GPU memory requirements are substantial. A 2025 survey of 3DGS compression methods identified this as the primary barrier to deployment on mobile devices and edge hardware. Most compression techniques reduce file size but do not reduce the rendering memory footprint, because the scene must be decoded before it can be rendered. Active research areas include spherical Gaussian approximations, codebook-based quantisation, and learnable factorisation, but a robust solution for real-time streaming on consumer devices remains an open problem.

2. Dynamic Scenes and Scalability

Static scene reconstruction is largely solved. Dynamic scenes, those involving moving people, changing lighting, or evolving environments, remain significantly harder. 4D Gaussian representations can consume over a gigabyte of memory for a short sequence, making them impractical for long video or real-time streaming without aggressive compression. City-scale scenes introduce additional challenges: block-based training approaches produce visual artefacts at block boundaries, and the sheer number of Gaussian primitives required strains even multi-GPU setups.

3. Geometric Fidelity

Gaussian Splatting produces excellent images but is not primarily a geometry reconstruction technique. The Gaussian ellipsoids optimise for visual appearance from the training viewpoints, not for physically accurate surface representation. This limits its utility in applications that require precise measurements, such as surveying, structural engineering, and some medical use cases. 2D Gaussian Splatting (2DGS, SIGGRAPH 2024) addresses this by using flat disc-like primitives that better represent surfaces, and hybrid approaches combining 3DGS with signed distance fields are showing promise.

The Road Ahead

The convergence of standardisation, compression research, and multi-industry deployment suggests that Gaussian Splatting is transitioning from an advanced capability to foundational infrastructure. Several specific developments will shape that transition.

1. The Splat as a New Content Format

With the Khronos Group's glTF extension and MPEG's Gaussian Splatting Coding work, the splat file is moving toward the same status as JPEG for images or MP4 for video: a universal, browser-native format for 3D spatial content. When that standardisation completes, any website, application, or platform that currently embeds images or video will be capable of embedding photorealistic 3D scenes. The implications for e-commerce, real estate, education, news media, and social platforms are substantial.

2. Generative Gaussian Splatting

The intersection of Gaussian Splatting with generative AI is an active research frontier. Tools like GaussianDreamer already demonstrate the ability to generate 3D Gaussian scenes from text descriptions using diffusion models. Where generative AI has historically struggled with spatial consistency, a Gaussian splat provides a consistent, explorable 3D anchor that downstream generation can operate against, enabling spatially coherent variant generation, relighting, and scene editing. This combination is likely to produce new creative tools for advertising, game development, and architectural visualisation over the next two to three years.

3. Persistent World Models

The most ambitious near-term application is the persistent, updateable world model: a continuously maintained Gaussian splat of a physical space that is updated as the environment changes. For autonomous systems, this means a shared map that robots and vehicles can read and write in real time. For consumers, it means AR applications that anchor digital content to physical locations with photorealistic fidelity. Niantic, which spun out its spatial computing division in 2025 with $250 million in dedicated funding, is explicitly building toward this vision.

Conclusion

Since its 2023 breakthrough, Gaussian Splatting has rapidly evolved from a technical novelty into a scalable, commercially viable medium deployed across cinema, healthcare, autonomous navigation, and geospatial digital twins. While challenges in memory overhead, dynamic scenes, and geometric precision remain, active research and formal standardization by bodies like the Khronos Group are embedding splats directly into the core web and media infrastructure alongside traditional images and video. For forward-looking organizations, the imperative is clear: building early internal capabilities now will secure compound workflow advantages as photorealistic 3D spatial content becomes standard across industries.