Gaussian Splat vs Mesh: Which Deliverable Should You Hand Over?

A buyer walks a listing at midnight from another city, loves the space, and asks for the floor plan. An architect gets the same capture and asks where the load-bearing wall actually sits. A studio wants the room in their game engine with working collision. One capture, three different questions, and the gaussian splat vs mesh decision decides which of them you can actually answer. Picking wrong does not ruin the walkthrough. It shows up two weeks later, when someone needs to do something with the file that the format was never built to do.

Gaussian splat vs mesh: the short answer

Hand over the splat when the job is a photoreal walkthrough for a buyer, a guest or an investor, and nobody downstream needs to measure or rebuild the space. Hand over a mesh when someone has to model on top of the capture, issue it into CAD or BIM, or drop it into a physics-driven engine. Most real projects that outlast the first client call need both, captured once and split into two deliverables at export.

How we are comparing

The gaussian splat vs mesh choice is really six smaller choices, and they surface in the order a project actually runs into them.

• Measurement and dimensional accuracy. Can a viewer of the file pull a real distance off it, or only a picture of one.
• CAD and BIM handoff. Does the file open in Revit, AutoCAD or Navisworks and let someone trace or annotate it.
• Collision and physics. Can a game engine or simulation use the geometry to stop a character walking through a wall.
• Game engine import path. Is the format a native asset type, or does it need a plugin that may not exist for the engine version in use.
• Editability and retopology. Can another artist reshape, simplify or repair the geometry after capture.
• Capture-to-viewable time and fidelity. How much manual work sits between the raw capture and something a client can actually walk, and how convincing that result looks once it is there.

Gaussian splatting, in short

A gaussian splat is not a surface. It is a cloud of millions of soft, oriented blobs, each one carrying a position, a colour, an opacity and a direction-dependent shading term, trained from ordinary photographs or video until the rendered views match the source images. There is no wireframe underneath and no defined edge between one object and the next. That is exactly what lets it render glass, foliage and reflections as continuous, photoreal surfaces instead of the tessellated approximations photogrammetry produces for the same subjects, per the original technique described in Kerbl et al.'s 2023 paper on real-time radiance field rendering. A companion piece, what a 3D gaussian splat actually is, goes further into how the training works and where the representation still struggles.

Triangle mesh, in short

Set the gaussian splatting vs polygon mesh question aside for a moment: a mesh is vertices, edges and faces, an explicit, storable definition of where a surface is. It arrives one of two ways. Photogrammetry runs structure-from-motion and multi-view stereo over a photo or video set, then meshes the resulting point cloud, a pipeline covered in more detail in the difference between gaussian splatting and photogrammetry. Or a laser scanner produces a point cloud directly, in formats like E57, LAS or LAZ, which gets meshed afterward. Site teams doing the latter at scale usually run it as a managed pipeline rather than by hand, which is the case our own lidar to 3D service is built around. Either origin ends at the same thing: a format every CAD, BIM and game engine tool has understood for decades.

The comparison table

Where the mesh wins

Knowing when to use a mesh instead of a gaussian splat starts here. This is the half of the comparison a splat-first pitch tends to skip, and it is where the mesh is simply the correct answer.

Measurement is the clearest case. A splat has no defined surface, only millions of semi-transparent blobs that look right from the angles they were trained on. There is nothing in that representation for a tool to snap a dimension line to. A mesh is the opposite: an explicit set of coordinates, which is why a point cloud, sparse and hard to read as it is for anyone outside the survey team, still beats a photoreal splat the moment someone needs a real number off the file.

CAD and BIM handoff follows from the same fact. Revit 2026 and ReCap Pro 2026 support a scan-to-mesh-to-Revit workflow, including RCMR segmented mesh import, that lets a modeler trace walls, floors, ceilings and MEP runs directly from captured geometry. Nothing equivalent exists for a splat, because there is no wall edge in a splat file for a BIM tool to snap to.

Collision and physics are a mesh requirement, not a preference. Unity's Mesh Collider builds its collision shape directly from mesh geometry through a cooking step, and needs the mesh marked convex before a rigidbody can use it at all. A splat carries no equivalent shape a physics engine can query. Any project where a character needs to stop at a wall, or a ball needs to bounce off a table, needs a mesh under the visuals somewhere, whatever renders on top of it.

Game engines lean the same way for import itself. FBX, OBJ and glTF mesh import has been a native, first-class path in Unity and Unreal for over a decade. Gaussian splats have not caught up: Unreal Engine 5.7 ships with no first-party splatting module, and getting a splat into the engine in 2026 still means a third-party plugin such as NanoGS or Volinga, tied to a specific engine version and platform.

Editability closes the case. A mesh can be decimated, retopologised, UV-unwrapped and patched where a scan left a hole, using tooling that has existed since the 1990s. A splat has no such surface to work on. If an artist needs to reshape a captured room, they are not editing the splat. They are starting over with a mesh.

And at low detail, a mesh is simply lighter. A decimated, low-poly export can be a fraction of the size of even a reduced-quality splat, where the original Gaussian Splatting research reports full-quality scenes running from around 366MB up to 1.4GB each, well above the 10 to 50MB a comparable NeRF needed for the same view. For a mobile app or a bandwidth-constrained embed, that difference is the whole ballgame.

Where the splat wins

None of that makes the splat the lesser format. It wins the cases a mesh is structurally bad at, and they happen to be the cases most client-facing deliverables actually need.

Photoreal fidelity is the first. A splat renders continuous, view-dependent surfaces because it was never asked to define a hard edge in the first place, which is exactly what lets it hold up on the subjects that defeat mesh reconstruction: foliage, glass, water, and thin rails or wire fencing that a mesh either merges into a solid blob or drops entirely. Multi-view stereo needs a surface it can triangulate consistently across viewpoints, and reflective or translucent material breaks that assumption before the mesher even starts.

Capture-to-viewable time is the second, and it is a labour difference more than a raw compute one. A mesh pipeline runs structure-from-motion or LiDAR registration, then a separate meshing stage, then UV-unwrapping, then texture baking, each one a distinct manual or semi-manual step before the result is something a client can walk. A splat skips all of it: train on the source photos or video and the trained result is already the viewable scene, with no meshing or texture-baking stage in between. Our own pipeline processes a capture in the cloud in 10 to 120 minutes with no local GPU and no installs, and that number covers the splat end to end, not a partial stage of it.

Messy geometry is the third, and it overlaps with fidelity without being identical to it. A construction site with scaffolding, a lobby with a chandelier, a room with a floor-to-ceiling window: these have no clean, closed surface for a mesher to reconstruct, and a mesh pipeline handling them either produces holes, floating fragments or a flattened approximation that reads as fake the moment a client rotates the view. A splat has no such requirement to fail. It represents what the cameras actually saw, not an idealised watertight solid standing in for it.

Which to choose

If the deliverable is a listing walkthrough, a hospitality preview or an investor tour, and nobody downstream needs to measure or rebuild the room, take the splat. It is the better looking file for less work, and looking good is the entire job.

If an architect, engineer or facilities team needs to model, measure or issue the capture as a BIM deliverable, take a mesh, sourced from photogrammetry for a quick single-room job or from a laser scan run through a service like our lidar to 3D pipeline when the site is large or the accuracy bar is higher.

If the target is a game engine and the geometry has to do work, drive collision, block a camera, trigger an event, take a mesh. Treat a splat there as a decorative visual layer at most, and budget for a plugin tied to your specific engine version rather than assuming native support.

If the project genuinely needs both a client-facing look and downstream engineering use, do not choose. Capture once, whether from video or from a scan, and export twice: a mesh for the geometry work and a splat for the walkthrough. The capture is the expensive part. Producing two deliverables from one capture session costs far less than running the shoot twice.

FAQ

When should you use a mesh instead of a gaussian splat?

Use a mesh instead of a gaussian splat whenever the next step needs the geometry to do something rather than just look like something: pulling a dimension, tracing a wall into a BIM model, generating a physics collider, or handing the file to an artist for retopology. A splat has none of that underneath it, so any of those tasks starts with a mesh regardless of how good the splat looked in the client demo.

Is splat to mesh conversion a real option?

Techniques exist to extract a rough surface from the centres of a trained splat, but the result is a lossy approximation of the geometry the splat implies, not a substitute for a mesh built directly from photogrammetry or a laser scan. Treat splat to mesh conversion as a fallback for a quick reference shape, not as a way to skip capturing a proper mesh when one is actually needed.

What is the right 3D deliverable for client walkthroughs?

For most client walkthroughs, the splat is the right 3D deliverable: faster to produce, better looking on difficult subjects like glass and foliage, and viewable on a phone with no plugin. Switch to a mesh only once the client's own team, not the sales process, needs to measure or build on top of what was captured.

The gaussian splat vs mesh question only feels like a rivalry until you notice both formats are answering different jobs from the same capture. A splat is the format for showing someone a space. A mesh is the format for doing something to it afterward. Knowing which question your project is actually asking, before you export anything, is most of the decision.

Unity, Unreal Engine, Autodesk, Revit, ReCap and Navisworks are named here as the tools this decision runs into in practice. All trademarks belong to their respective owners. No affiliation, partnership or endorsement is implied.