Two representations, each missing exactly what the other has.
Most organisations trying to build a digital twin end up choosing between two things that are each half of the answer.
CAD gives you structure. Exact geometry, parts that are named and separable, assemblies, materials, tolerances. Everything downstream wants this: simulation, clash detection, parts ordering, maintenance planning. What it does not give you is any relationship to the object that actually exists. A CAD model is the design intent, rendered in a way that looks like an engineering drawing because that is what it is.
Gaussian capture gives you reality. Photoreal, immediate, cheap to produce, and correct about the things CAD is silent on — wear, modification, the pipe somebody added in 2019, what the equipment genuinely looks like under the lighting it lives in. What it does not give you is structure. A splat is millions of coloured blobs. It has no idea which of them are a pump.
CAD knows what the thing should be. A capture knows what the thing is. Almost every interesting question about an asset lives in the gap between those two.
The standard workaround is to keep both and align them by hand, which is slow, and which nobody redoes when the asset changes. The interesting research direction is to stop treating them as two deliverables and start using each to compensate for the other’s weakness during reconstruction itself.
What the research actually demonstrated.
CADSplat, published in September 2026 by Overdulve, Jorissen and Michiels at the Digital Future Lab, Flanders Make and Hasselt University, is the clearest statement of the idea so far. The mechanism is worth understanding because it explains both the promise and the limits.
How the two are joined
1. photographs of the real object
sparse, wide baseline, under 15
2. segment the object silhouette
and match it against silhouettes
rendered from a CAD library
-> recovers camera-to-object pose
WITHOUT a dense photo set
3. anchor Gaussian primitives to the
surface of the retrieved CAD model
-> the splat inherits structure
instead of inventing it
4. jointly optimise
- the Gaussian parameters
- the camera registration
- a non-rigid deformation field
the deformation field absorbs the
difference between the CAD model
and the physical object
Result: photoreal, geometrically
anchored reconstruction from few views,
degrading gracefully to as few as threeStep two is the one that does the commercial work. Ordinary splat reconstruction needs many overlapping photographs because it has to infer both where the cameras were and what the shape is, simultaneously, from nothing. Give it a CAD model and half that problem disappears: the shape is already known, so a handful of views is enough to work out where you were standing.
Step four is the one that makes it honest. The physical object is never exactly the CAD model, and a method that pretended otherwise would be useless. The deformation field exists precisely to represent that divergence.
Why "fewer than fifteen photographs" is the number that matters.
It is easy to read the accuracy claims and miss the commercial point. The interesting figure in that research is not fidelity. It is view count.
Capture effort is what actually kills digital twin programmes. Not software, not licensing — the fact that getting good coverage of an asset means a planned visit, a trained operator, a shutdown window, hundreds of images, and a repeat of all of it when something changes. Multiply that across a few hundred assets and the programme quietly dies at the pilot.
A method that works from a dozen wide-baseline photographs is a different activity. It is a walk round an asset with a phone or a compact camera, done by someone who already works there, during a normal shift. That changes three things at once.
Coverage becomes achievable. The question shifts from “which assets can we justify scanning” to “which assets do we have CAD for”, and for most manufacturers the second list is much longer.
Refresh becomes plausible. A twin that is captured once is a photograph. A twin that can be cheaply recaptured is a record. Only the second one tells you anything about change over time.
The cost centre moves. When capture is cheap, the budget goes where it should have been all along: into what the twin is connected to and what decisions it supports.
The as-built gap is the product, not a footnote.
There is a subtler idea buried in the deformation field, and it may turn out to be worth more than the photorealism.
That field is a measurement of how the physical object differs from its design. Normally that divergence is invisible — it exists as tacit knowledge in the heads of people who maintain the equipment, or as a note on a drawing nobody has opened since commissioning. A reconstruction method that has to model the difference explicitly, in order to work at all, produces that divergence as data.
For anyone responsible for an ageing estate, that is the interesting output. Where has the asset drifted from the design? Which units differ most from their drawing? What changed between this capture and the one eighteen months ago? Those are maintenance and compliance questions, and they are considerably more valuable than a nice-looking model.
Related work is pushing in the same direction from other angles: material-informed Gaussian splatting attaches semantic material properties to reconstructed scenes so they can drive sensor simulation, and ArtiTwinSplat reconstructs articulated, interactable twins from RGB-D video. The common thread across all of it is the same shift:
The first generation of Gaussian splatting reconstructed how the world looks. The second generation is trying to reconstruct how the world behaves.
Where this sits against LiDAR, honestly.
This will be read by some people as an argument that scanning is finished. It is not, and pretending otherwise would waste your money.
| If the twin is for | Use | Why |
|---|---|---|
| Fabrication, clash detection, retrofit against tight tolerances | Survey-grade scanning | You need certified accuracy with a stated tolerance. A reconstruction is an estimate, however good it looks. |
| Operator familiarisation, training, remote inspection, stakeholder review | Gaussian capture, CAD-anchored where CAD exists | Recognition and context matter more than millimetres, and cost per asset is the binding constraint. |
| Condition and change tracking on assets you have drawings for | CAD plus capture | The divergence between design and reality is the actual output. |
| Anything legally or contractually dependent on a measurement | Survey-grade scanning | If a number will be argued about, it needs provenance a reconstruction cannot supply. |
| Assets with no CAD and no drawings | Photogrammetry or scanning | The method in this article has nothing to anchor to. |
The realistic picture is not replacement. It is that a scanning budget which previously covered ten assets can now cover the ten that genuinely need survey accuracy, while a camera covers the two hundred that only needed to be recognisable and current.
What you need in place before this is worth trying.
Four preconditions, and the first one disqualifies more organisations than people expect.
- CAD you can actually retrieve. Not “we have CAD somewhere”. Findable, openable, matched to the physical unit, in a format something can render. If your models are locked in a retired seat of a package nobody has licensed since 2018, that is the project, and it is a data project rather than a 3D one.
- A decision the twin serves. The failure mode of every twin programme is building the model first and looking for the question afterwards. Name the decision, name who makes it, name what they use today.
- An honest tolerance. Write down the accuracy the use case genuinely requires before anyone demonstrates anything, because everything looks accurate in a nice render.
- Somewhere for it to live. A reconstruction that sits in a folder is a picture. A twin is a thing people open during their work, which means a viewer, access control and a refresh path.
If those four are in place, this is worth a small, bounded experiment on a handful of assets rather than a programme. Our Idea Validation Sprint at £495 is built for settling whether the question is the right one, and prototype work starts from £3,250.
Where we actually stand on this.
Being precise about the line between what we operate and what we are reading.
What we run in production: static Gaussian splat capture, cleanup, editing and browser delivery, through our own browser splat editor and Simam 3D Studio, with published scenes in our Gaussian gallery. We also run Simam BIM, a browser workbench that imports, views, edits and exports IFC, glTF, STEP and DXF — so the CAD half of this pairing is ordinary work for us.
What we have delivered against real assets: photoreal capture used operationally, including a site where the drawings had fallen behind reality and the capture became the navigational source of truth.
What is research and not a product: CAD-anchored sparse-view reconstruction itself. CADSplat is a paper from September 2026, not something you can buy, and we have not put it into a client pipeline. We are describing it because the direction is commercially important and because the preconditions above are worth getting in order now, not because anyone should procure against it this quarter.
If a supplier tells you they are already delivering CAD-anchored photoreal twins as a routine service, ask which method, ask for the view count, and ask what the deformation field told them about the asset. The answers will be informative either way.
What this means for a buyer.
Start with the business decision, audience, and evidence the project must produce. Simam Digital can turn that into a focused discovery, prototype, MVP, or production roadmap across AI applications, SaaS platforms, digital twins, real-time 3D, XR, and interactive systems.
Sources and further reading
- CADSplat: Sparse-View 3D Gaussian Splatting Aided by CAD Models for Robust, Photorealistic Digital-Twin Reconstruction
- Material-informed Gaussian Splatting for 3D World Reconstruction in a Digital Twin
- ArtiTwinSplat: Interactable Digital Twin Reconstruction via Gaussian Splatting from RGB-D videos
- 3D Gaussian Splatting for Real-Time Radiance Field Rendering (Kerbl et al.)
