Gaussian Splatting Viewer Online — View 3D Gaussian Splats in Browser

What Is 3D Gaussian Splatting?

3D Gaussian Splatting (3DGS) is a technique for representing and rendering 3D scenes using millions of small, colored, semi-transparent ellipsoids — called Gaussians or "splats." Instead of building a traditional triangle mesh with textures, Gaussian splatting represents a scene as a cloud of these overlapping ellipsoids. When rendered from any viewpoint, the splats blend together to produce a photo-realistic image that is virtually indistinguishable from a photograph.

The technique was introduced in the landmark 2023 paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering" by Kerbl et al. from INRIA. It arrived as a breakthrough solution to a long-standing problem: how do you capture a real-world scene in 3D and render it in real time with photographic quality? Previous approaches — notably Neural Radiance Fields (NeRF) — could produce stunning results but required minutes per frame to render using expensive ray marching through neural networks. Gaussian splatting achieves comparable visual quality at 30-100+ frames per second using a fundamentally different rendering approach.

How It Works

The process begins with a set of photographs of a scene taken from multiple viewpoints — typically 50 to 300 images. These images are first processed through Structure from Motion (SfM), usually using COLMAP, to estimate camera positions and produce a sparse point cloud. This sparse cloud serves as the initialization for the Gaussian splats.

Each point in the initial cloud is replaced with a 3D Gaussian — a soft, fuzzy ellipsoid defined by its position (mean), covariance matrix (shape and orientation), opacity, and color (represented as spherical harmonics coefficients for view-dependent effects). The training process then optimizes these parameters by comparing rendered images to the original photographs. Gaussians that are too large get split into smaller ones. Areas that are under-represented get new Gaussians added. Areas that are over-represented have Gaussians pruned. After 30,000 or so training iterations — typically 15 to 30 minutes on a modern GPU — the scene is fully captured.

The rendering process is what makes Gaussian splatting special. Instead of casting rays through a volume (as NeRF does), the renderer projects each 3D Gaussian onto the 2D screen as a 2D Gaussian splat. These 2D splats are sorted by depth and composited front-to-back using alpha blending. The entire process is differentiable and massively parallelizable on GPUs. The result is real-time rendering at high resolution — something NeRF fundamentally cannot achieve without distillation or baking.

File Formats for Gaussian Splats

Gaussian splat data is most commonly stored in PLY (Polygon File Format) files. Unlike traditional PLY files that store mesh vertices and faces, Gaussian splat PLY files store per-splat properties: position (x, y, z), scale (sx, sy, sz), rotation (quaternion), opacity, and spherical harmonics coefficients for color. A typical scene contains 500,000 to 5 million Gaussians, with each Gaussian requiring approximately 200-300 bytes of data. This puts file sizes in the 100 MB to 1.5 GB range for raw PLY files.

The .splat format is a more compact binary format that has emerged as a de facto standard for web delivery. It stores the same essential data — position, scale, rotation, color, opacity — in a tightly packed binary layout, typically at around 32 bytes per splat. This reduces file sizes by 80-90% compared to full PLY files, making web streaming practical. Many web viewers support .splat files natively.

More recently, compressed formats have appeared. KTX-based and codec-based compression schemes can reduce Gaussian splat data to a few megabytes for typical scenes, enabling streaming on mobile connections. The standardization of these compressed formats is still in progress, but the direction is clear: Gaussian splats will become as easy to deliver over the web as traditional 3D models.

Viewing Gaussian Splats in the Browser

Web-based Gaussian splatting viewers have advanced rapidly. The first web implementations appeared in late 2023, just months after the original paper. By 2024, multiple open-source and commercial viewers could render Gaussian splats in real time using WebGL 2.0 and compute shaders via WebGPU. Today, the ecosystem is maturing fast.

The rendering pipeline for web-based splat viewers is conceptually straightforward but technically demanding. The viewer must download and parse the splat data, upload it to GPU buffers, sort splats by depth for the current viewpoint (this must happen every frame as the camera moves), project each splat to screen space, and composite them with alpha blending. The sorting step is the performance bottleneck — sorting millions of elements every frame at 60 FPS is a hard computational problem. Efficient implementations use GPU-based radix sort algorithms, which is why WebGPU support is a significant performance win over WebGL.

Most desktop browsers with discrete GPUs can handle scenes of 1-3 million Gaussians at 30+ FPS. Mobile browsers are more constrained, typically handling 500,000 to 1 million Gaussians at acceptable frame rates. As WebGPU adoption increases and GPU hardware improves, these numbers will continue to rise.

Gaussian Splatting vs Traditional 3D

Gaussian splatting and traditional mesh-based 3D are not competing approaches — they serve different purposes and excel in different scenarios.

Gaussian splatting excels at photorealistic capture of real-world scenes: rooms, buildings, landscapes, objects with complex materials (glass, water, foliage). The capture process is as simple as taking photographs. The output preserves lighting, reflections, and fine detail that would be extremely difficult to reproduce with manual 3D modeling or even photogrammetry-to-mesh pipelines.

Traditional meshes excel at authored content: CAD models, game assets, 3D-printable objects, animations, and anything that needs to be edited, measured, or simulated. Meshes have well-defined surfaces, precise geometry, and decades of tooling support. You cannot 3D print a Gaussian splat. You cannot measure the dimensions of a splat-based object with precision. You cannot rig and animate a Gaussian splat character (though research is making progress here).

For viewing and sharing captured scenes, Gaussian splatting is increasingly the best approach. A real estate walkthrough, a museum artifact, a construction site — these are all cases where photorealistic capture matters more than geometric precision, and where the effort to create a high-quality mesh would be prohibitive.

The Current State of the Technology

As of mid-2026, Gaussian splatting has moved from research novelty to practical tool. Several commercial platforms now offer end-to-end capture-to-viewer pipelines. Polycam, Luma AI, and others let users capture Gaussian splat scenes directly from phone cameras and share them via web links. NVIDIA, Google, and Meta have all published research advancing the technology. Game engines are integrating splat rendering — Unreal Engine has experimental support, and Unity plugins are available.

Key advances since the original paper include: dynamic scene support (capturing and replaying moving scenes), relightable Gaussians (separating lighting from geometry for re-lighting), compressed representations for efficient streaming, mesh extraction from splats (bridging the gap to traditional 3D), and text/image-to-3D generation using Gaussian splatting as the output representation. The technology is evolving extremely rapidly.

The Future of 3D Capture

Gaussian splatting represents a fundamental shift in how we think about 3D content. The traditional pipeline — modeling, texturing, lighting, rendering — requires skilled artists and expensive tools. Gaussian splatting replaces this with a much simpler pipeline: take photos, train, view. The barrier to creating photorealistic 3D content has dropped from months of expert work to minutes of phone photography.

For the web specifically, the implications are significant. Imagine product pages where you can orbit around a photorealistic capture of the actual product, not a 3D artist's approximation. Real estate listings where you can walk through the actual space, not a simplified 3D model. Documentation where you can inspect a machine from any angle with perfect visual fidelity. This is the promise of Gaussian splatting, and it is rapidly becoming reality.

The technology is still young. File sizes are large. Editing captured scenes is limited. Standardized formats and tooling are still emerging. But the trajectory is clear. Just as digital photography replaced illustration for product images, Gaussian splatting will increasingly replace manual 3D modeling for representing real-world objects and spaces. The shift has already begun.