Photogrammetry vs neural reconstruction in 2026 — when each wins

A customer asks us to "make a 3D model of this site." It is the most common opening line in our intake. The follow-up question — the one that decides which pipeline we actually run — is rarely the one the customer has prepared an answer for. So we ask it for them.

The question is: what are you going to do with the model?

The two families, briefly

Photogrammetry is the classical pipeline. Take many overlapping photographs. Detect features, match across frames, solve a bundle adjustment for camera poses, dense-match for depth, fuse depth maps into a textured mesh. Every step is a documented algorithm with known failure modes and known runtime. The output is a triangle mesh with a UV-unwrapped texture, and you can hand it to any CAD package on the planet.

Neural reconstruction is the umbrella for Gaussian splatting, NeRF and its descendants, and the new feed-forward 3D foundation models that produce a scene representation in a single forward pass. The output is not a mesh; it is a learned representation that is rendered by a neural network or a differentiable rasteriser. You see a photoreal scene; you do not see a parts list of triangles you can measure.

The decision tree we run

                ┌──────────────────────────────────┐
                │ What is the deliverable used for?│
                └────────────────┬─────────────────┘
                                 │
              ┌──────────────────┼─────────────────────┐
              ▼                  ▼                     ▼
   ┌─────────────────┐  ┌─────────────────┐   ┌─────────────────┐
   │ Engineering /   │  │ Visualisation / │   │ Training data / │
   │ measurement /   │  │ walkthrough /   │   │ simulation /    │
   │ regulated audit │  │ marketing       │   │ ML downstream   │
   └────────┬────────┘  └────────┬────────┘   └────────┬────────┘
            │                    │                     │
            ▼                    ▼                     ▼
     PHOTOGRAMMETRY       NEURAL RECON           EITHER, fuse
     (deterministic       (Gaussian splat,        (mesh backbone +
      mesh + texture,     feed-forward 3D,         splat overlay,
      auditable)          NeRF where stable)       depending on use)
        

The boxes hide nuance. A few notes from the road.

When photogrammetry wins outright

• Measurement is the deliverable. A surveyor, a structural engineer, an environmental compliance group — they need a mesh whose vertices are at known metric positions. They are going to import the result into their existing tools, query distances and volumes, and stand behind the numbers. Neural representations cannot be queried that way without a separate mesh extraction step that re-introduces all the photogrammetric concerns.
• Regulated reproducibility. If the customer's standard requires a documented pipeline, photogrammetry's per-step audit trail is an asset. Neural reconstruction is a single opaque step; the audit story is harder.
• Capture-rich situations. When the rig and schedule allow dense overlapping coverage, photogrammetry uses every frame productively. Neural methods do too, but the marginal value of frame number 800 is much higher in photogrammetry.

When neural reconstruction wins outright

• Sparse or non-uniform capture. A single drone sortie, a handful of ground angles, mixed sensor types — modern feed-forward neural methods handle this gracefully. Classical photogrammetry will produce holes or refuse to converge.
• The deliverable is "see it from any angle." Walkthroughs, marketing visualisation, training-data generation for downstream ML. The customer never measures off the result; they look at it. Splats render beautifully and stream cheaply.
• Time-to-first-result matters more than precision. Feed-forward 3D foundation models give you a viewable scene minutes after capture. The customer is on-site, the value is in showing them something now, and the measurement-grade refinement happens later if at all.

When you fuse them

The honest answer for a lot of mid-size engagements is "both."

We will sometimes ship a photogrammetric mesh as the metric backbone of the deliverable — the customer measures off it, the surveyor signs off on it — and a Gaussian-splat overlay as the visualisation surface. The splat lives in a web viewer the customer's stakeholders can spin around. The mesh lives in their CAD package. Both are derived from the same capture session; the extra cost is one additional reconstruction pass.

Customers like the result because they get the deliverable for the engineering team and the deliverable for the executive walkthrough off a single capture. We like it because the failure modes of the two pipelines are largely uncorrelated — when one struggles, the other usually has the answer.

What I want the next buyer to take away

• Pick by deliverable, not by buzzword.

  A buyer who needs a measurable mesh for engineering work is a photogrammetry customer. A buyer who needs a viewable scene for visualisation, walkthroughs, or training data is a neural-reconstruction customer. The split is almost that clean.

• Capture density is the hidden constraint.

  Neural methods are forgiving of sparse capture; classical photogrammetry punishes it. If the engagement only allows a single drone pass or a handful of ground images, neural methods will produce a usable result where photogrammetry will not.

• Determinism still matters in regulated domains.

  A photogrammetry pipeline is a sequence of well-understood steps you can audit. A trained neural reconstruction is, by definition, a function of weights and the order it saw frames. Pick the deterministic pipeline when audit and reproducibility are part of the deliverable.

• Compute is no longer a tiebreaker.

  Both pipelines now run on workstation-class GPUs in reasonable time. The decision is not "which is cheaper to compute" — it is "which produces the deliverable the customer can actually use." Plan around the deliverable.

• You can fuse them.

  A useful pattern: photogrammetric mesh as the metric backbone, neural splat overlay for the visualisation surface. Customer measures off the mesh, walks the scene off the splat. Costs more to build; ships a deliverable neither family can produce alone.

— AK-mee Engineering