If RTK and PPK are about where each photo sat in space, a gaussian splat is about how the scene looks when you spin it in a viewer. The phrase shows up next to mesh and point cloud on processing menus and client wish lists. Plenty of mapping pilots who already deliver orthos still cannot say what a splat is made of — or what it is not.
A 3D gaussian splat is a scene representation built from millions of tiny, soft blobs. Each blob has a position in 3D, a shape (covariance), an opacity, and a color that can change with viewing angle. Stack enough of them and a dedicated viewer can look photoreal as you orbit. That is the product: appearance that interpolates between the camera positions you actually flew.
Pilot Ledger is ops and CRM for commercial pilots, not a rendering engine. This is the field explanation of what a splat is doing when the client says it "looks like the site." For the positioning story that still decides whether a map belongs on someone else's control, start with What Is RTK? Drone Mapping Explained and What Is PPK? Drone Mapping After the Flight.
What a 3D gaussian splat is
Strip the marketing. A splat scene is not a continuous surface you can click and trust. It is a cloud of primitives, each one a small 3D Gaussian — a soft ellipsoid rather than a hard triangle or a single lidar return.
Each primitive typically carries:
- Position — where that blob sits in the reconstructed volume.
- Covariance (shape) — how wide and how oriented the blob is along three axes. Thin along one direction, fat along another. That is what makes edges soft instead of faceted.
- Opacity — how much the blob contributes when a ray from the viewer hits it.
- Color — often view-dependent, so the same blob can tint differently as you rotate. That is how reflections, foliage, and thin structures can look convincing from nearby viewpoints.
Training (or optimization) adjusts those parameters so that when the renderer looks from the original camera poses, the composite image matches your photos as well as it can. Between those poses, the viewer interpolates. Near the flight path, a good splat can look photo-real. That is the strength.
A photoreal viewer is not a survey instrument. The splat is optimized for what the cameras saw, not for a civil set.
What a gaussian splat is not
This section is the one that keeps clients and scopes honest.
A gaussian splat does not replace photogrammetry for survey deliverables. Photogrammetry, when flown and controlled for mapping, still produces the orthomosaic, the DEM or DSM, the mesh tied to check shots, and the report that says how the map sits on independent points. A splat can share the same photos as input. It is not the same product as a controlled map.
Looks right is not the same as measurable. A curb that looks sharp in the viewer can still sit in the wrong place relative to a monument. A stockpile that "feels" full can still be a visual guess. If the deliverable has to answer a quantity, a cut-fill, a setback, or a legal boundary, you need a controlled mesh, ortho, or point product with checks — not a pretty orbit.
You cannot survey from a splat as if it were a controlled mesh or orthomosaic. Do not extract stakeout coordinates from a splat and treat them like GNSS-tied vertices. Do not claim absolute accuracy because the foliage looks real. Do not hand a client a splat link and call the mapping job closed when the scope asked for survey-grade products.
A splat is also not a substitute for a known point, a datum, Fixed versus Float, or independent check shots. Those still decide whether any photo-based product belongs on someone else's control. The splat does not forgive a float-heavy flight or a missing base any more than a nice-looking mesh does.
Brief contrast: splat vs mesh, point cloud, and ortho
Keep the distinctions short. A deeper head-to-head belongs in a later post.
Mesh. A mesh is a surface of triangles (or similar faces). You can often measure along it, drape textures, and export something a CAD or GIS workflow expects — when that mesh was built and controlled for mapping. A splat is soft blobs optimized for appearance. It is not a triangle surface you treat as geometry by default.
Point cloud. A point cloud is discrete samples with coordinates (and often color or intensity). Dense photogrammetry clouds and lidar clouds are still the language of many survey and inspection scopes. A splat may start from similar imagery, but the stored object is not "points you can trust for coordinates." Do not confuse a photoreal orbit with a LAS you would send for classification or volume.
Orthomosaic. An ortho is a map-like image: nadir (or near-nadir) imagery corrected so it can sit on a plane and overlay other geospatial layers when georeferencing is done right. A splat is a free-orbit 3D view. It does not replace the plan-view deliverable the civil set expects.
When you need the full comparison — when to fly for each, what fails when you treat one like the other — that is the next educational piece in this series. This post only draws the boundary: splat for look and walkthrough; mesh, cloud, and ortho for the products scopes still name.
Why it looks real near the path — and invents away from it
The optimizer mostly pins surfaces that many views saw from different places. Fly a normal mapping grid at one height, or a single pass down a hall, and the splat can look excellent from those training cameras. Step off that path — drop under the canopy, peek behind a wall the cameras never saw, or pull far off the flight lines — and the renderer fills gaps with soft guesswork.
That is not a software insult. It is how appearance models behave. Photos constrain what the cameras occupied. Everything else is interpolation. Near the path: photoreal. Away from the path: invention. Treat invention as invention, not as a measurement.
What to confirm before you sell or accept a splat
- The client scope names a splat (or 3D walkthrough) as a deliverable on purpose — not as a substitute for ortho, DEM, or controlled mesh.
- You still have a plan for the survey products if the job needs quantities, overlays, or legal placement.
- Photo positions and control still follow the same GNSS discipline you use for mapping (RTK live, PPK after the flight when the link drops).
- Nobody on the team is extracting stakeout points from the splat viewer and calling them survey.
- The handoff language says "looks like the site in a viewer," not "survey-grade model."
None of that requires a graphics paper. It requires treating a gaussian splat as an appearance model built from soft 3D blobs — position, covariance, opacity, color — that can look photo-real in a viewer without becoming a measurable map.
If a viewer format, a scope line, or where a splat sits next to mesh and ortho still is not clear, ask the follow-up on Ask Mav at pilotledger.com. SkyView is on the waitlist for September 2026.