Gaussian splatting

Gaussian splatting represents a scene as millions of small, soft, coloured ellipsoids called splats instead of triangles, usually created automatically from photos of a real place or object. The technique became widely used after a 2023 SIGGRAPH paper by Kerbl and colleagues, which represented scenes "with 3D Gaussians" and achieved "high-quality real-time (≥ 100 fps) novel-view synthesis at 1080p resolution" (Kerbl et al. 2023).

For game artists it sits somewhere between photogrammetry and a photograph: captures look realistic from any angle, but they are hard to light, animate, edit or collide with. This page explains how splats work, what they cost, how to get them into engines, and where they fit today.

#How it works

1. Capture. You take many photos or a video around a subject. As with photogrammetry, coverage and sharp images matter most. See photogrammetry and scanning tools.
2. Camera solve. Software works out where each photo was taken. The method starts "from sparse points produced during camera calibration" (Kerbl et al. 2023).
3. Training. An optimiser turns those points into Gaussians and refines them until renders match the photos. It adds and removes Gaussians as it goes ("interleaved optimization/density control") and optimises each one's anisotropic shape (Kerbl et al. 2023).
4. Rendering. Each frame, the splats are sorted by depth and blended on the GPU, using "a fast visibility-aware rendering algorithm that supports anisotropic splatting" (Kerbl et al. 2023).

PlayCanvas lists what each splat stores: a position, a covariance matrix that "defines the size, orientation, and shape of the ellipsoid", a colour and an opacity (PlayCanvas Developer Site: What is Gaussian Splatting?). Khronos's draft glTF extension stores the same idea as position, orientation, scale, colour and opacity (khronos.org: Khronos announces glTF Gaussian Splatting…).

#Strengths and weaknesses of splats

Good at: capturing real places quickly, from photos, with little manual modeling. PlayCanvas names "real-world scene capture", product visualisation and virtual tourism as ideal uses (PlayCanvas Developer Site: What is Gaussian Splatting?).

Bad at (according to PlayCanvas) (PlayCanvas Developer Site: What is Gaussian Splatting?):

• "Animated content - Each frame would require separate training"
• "Procedural geometry - Traditional meshes are more flexible for generated content"
• "Interactive objects - Editing and deformation are challenging"

Lighting is another major limit. A capture records the light that was there on the day. A 2026 browser shooter built from a scanned building showed the problems: splats "don't capture anything about the world except how it looks", the lighting "doesn't apply very well to the characters, and there aren't any shadows", and low-quality up-close detail "will continue to be a challenge" (Tom's Hardware: Developer creates a basic first person…).

#Performance and memory

Splat scenes are large. Aras Pranckevičius's open-source Unity viewer gives detailed numbers for one standard test scene (GitHub: Gaussian Splatting playground in Unity):

The same project also shows the effect of compression: a "Very Low" preset capture came in under 8 MB (GitHub: Gaussian Splatting playground in Unity).

#Getting splats into engines

Capture tools are also being developed. The Unreal plug-in Atlux, for example, generates camera volumes and COLMAP data inside Unreal for training in tools such as Postshot and LichtFeld Studio (80 Level: All-in-one plug-in for Unreal Engine 5 to…).

#File formats and standards

The original paper's output is PLY files; the compressed SPZ format from Scaniverse is also supported by tools such as the Unity viewer (GitHub: Gaussian Splatting playground in Unity). In February 2026 Khronos released KHR_gaussian_splatting as a release candidate glTF extension, with ratification expected in Q2 2026 (khronos.org: Khronos announces glTF Gaussian Splatting…). It stores splat attributes in glTF 2.0, falls back to a point cloud in viewers without splat support, and was developed with contributors including Autodesk, Cesium/Bentley Systems, Esri, Huawei, Niantic Spatial, NVIDIA and XGRIDS (khronos.org: Khronos announces glTF Gaussian Splatting…). Compression extensions (Niantic's SPZ and Qualcomm's L-GSC) were proposed alongside it (khronos.org: Khronos announces glTF Gaussian Splatting…). Check its current status before relying on it. See 3D file formats.

#Common mistakes

• Expecting to relight a capture. The lighting is baked into the colours, as in the browser shooter (Tom's Hardware: Developer creates a basic first person…).
• Ignoring splat count. Memory grows with every splat (GitHub: Gaussian Splatting playground in Unity).
• Capturing for close-ups. Up-close detail is still a weak point (Tom's Hardware: Developer creates a basic first person…).
• Forgetting collision and navigation. Splats have no surfaces; you need proxy meshes (Tom's Hardware: Developer creates a basic first person…).
• Shipping with a non-commercial trainer. Check licences (GitHub: Gaussian Splatting playground in Unity).

#Further reading and watching