Guides
Gaussian splatting vs meshes: which 3D output do you need?
3D Gaussian splatting and polygon meshes compared: how each works, what each is good for, which tools open them, and when to choose one over the other.
3 min read
Quick answer
What is the difference between Gaussian splatting and a mesh?
A mesh is a surface made of polygons with textures, which every game engine and 3D editor understands. A Gaussian splat is a cloud of millions of small, soft, coloured ellipsoids that renders photorealistic views of a captured scene. Splats look better for real places. Meshes are what you can edit, animate and put in a game.
Two ways to store a 3D thing
A mesh is the classic 3D model: points (vertices) joined into triangles or quads that form a surface, with images (textures) wrapped onto it through UV coordinates. It is what Blender edits, what game engines render and what GLB, FBX and OBJ files contain. See 3D file formats.
A Gaussian splat has no surface at all. It is a large collection of 3D Gaussians: fuzzy, semi-transparent ellipsoids, each with a position, size, rotation, opacity and view-dependent colour. Rendered together, they blend into an image. Up close it looks like soft blobs. From a normal distance it can look like a photograph.
Where splatting came from
3D Gaussian Splatting for Real-Time Radiance Field Rendering was published in 2023. It solved a problem with NeRF: NeRFs looked great but were slow to render, because each pixel meant querying a neural network many times. Splats are explicit data that a GPU can sort and draw directly, so they render in real time. Like NeRF, it is mainly a reconstruction method. You capture photos or video of a real scene and the splat is optimised to match them.
What each is good at
Splats are better for:
- photorealistic captures of real places, objects and people;
- reflections, fine hair, foliage and fuzzy materials that are hard to model as polygons;
- virtual tours, film reference and visual-effects backgrounds.
Meshes are better for:
- editing: moving a vertex, cutting a hole, adding a handle;
- animation, rigging and physics;
- game engines, where meshes are the native format with mature lighting, shadows, collision and level of detail;
- 3D printing, which needs a closed surface;
- small file sizes for simple objects.
Can you convert between them?
From mesh to splat, yes, by rendering the mesh from many angles and training a splat on the images. From splat to mesh, only approximately. You extract a surface from the splat and lose much of what made it look good, especially transparency and fine detail. If you need a mesh at the end, it is usually better to generate or capture a mesh in the first place.
Which tools open splats
Support is growing but still uneven. Browser viewers are ahead. Babylon.js Sandbox, for example, opens .splat, .spz and .sog files alongside GLB and OBJ. Game engines and DCC tools often need plugins. Check before you commit a pipeline to splats. For regular files, see how to view 3D models.
Which one does AI 3D generation produce?
Both exist. Some research models, such as TRELLIS, can decode to Gaussians, radiance fields or meshes. Tools built for game and product work output meshes, because that is what the rest of the pipeline uses.
Vi3W generates meshes: textured, automatically retopologised, and exported as GLB, FBX or OBJ, so they go straight into Blender, Unity, Unreal, Godot or a web page. If you need a photoreal capture of a real room, a splat capture app is the better tool. If you need an object you can edit and ship in a game, you need a mesh.
Last reviewed