Gaussian Splatting (formally 3D Gaussian Splatting) is the 3D scene representation introduced in a 2023 SIGGRAPH paper that has rapidly displaced NeRF for many real-time photoreal use cases. Instead of encoding a scene as a neural network you query, Gaussian Splatting represents the scene as millions of coloured 3D Gaussians (small fuzzy ellipsoids) that are blended together when rendered. The result trains in minutes, renders in milliseconds and looks photorealistic.
The basic recipe:
- Take 50–200 photos of a scene from different angles.
- Use a structure-from-motion pipeline (COLMAP) to estimate camera positions.
- Initialise a few hundred thousand Gaussians from the resulting sparse point cloud.
- Optimise position, colour, opacity, scale and orientation of every Gaussian against the input photos.
- Periodically split, clone and prune Gaussians as needed.
- Render by projecting Gaussians to the camera and alpha-blending.
Why Gaussian Splatting won so quickly:
- Real-time rendering — 30–120 FPS on consumer GPUs; NeRF was seconds per frame.
- Fast training — minutes to hours, not days.
- High quality — visually competitive with the best NeRF variants.
- Editable — you can move, recolour or delete individual Gaussians, which is awkward with NeRF.
- Easier integration — Gaussians are a more familiar data structure than implicit neural fields; tooling has caught up fast.
The 2026 ecosystem:
- Polycam, Luma AI, Postshot — consumer apps that turn phone video into Gaussian splats.
- Niantic Scaniverse — phone-based 3D capture with splat output.
- Spline, Threestudio, Postshot — desktop tools for splat editing and rendering.
- WebGPU / Three.js renderers — splats viewable directly in the browser; the foundation of the new wave of interactive 3D web experiences.
- Game engines — Unreal and Unity now have native Gaussian splat support for real-time rendering.
Where Gaussian Splatting fits in 2026 production:
- Real estate virtual tours — capture a property in 30 minutes, ship a walkable photoreal experience that beats traditional 360 photos.
- E-commerce 3D viewers — products you can rotate and zoom with photo-quality fidelity.
- AR / VR experiences — high-quality scene capture for immersive content.
- Cultural and architectural preservation — digital twins of buildings, monuments, archaeological sites.
- Visual effects and previs — environment capture for film with quick turnaround.
- Sports and live event capture — multi-camera splats for broadcast augmentation.
The trade-offs vs traditional 3D pipelines:
- Splats are not meshes — you cannot easily UV-map them, animate them with skeletons, or edit them with traditional 3D tools. Conversion to meshes is possible but lossy.
- Storage is heavier than meshes — millions of Gaussians take tens to hundreds of MB.
- Hard surfaces, glass and reflections are still challenging; specular materials look better in some implementations than others.
- Tooling is younger — the ecosystem is moving fast but still less mature than the decades-old mesh-based pipeline.
For a US studio or developer in 2026, Gaussian Splatting is the right answer when you need photoreal scene capture for real-time interactive use. For traditional VFX and game asset pipelines, meshes still rule. The two will increasingly coexist, with splats handling environments and meshes handling characters and props.
