Why a floor plan is not a 3D model
A floor plan looks like it already contains everything you would need. Walls, room boundaries, door swings, a scale bar. It reads as geometry. It is not geometry, it is a drawing of geometry, and the gap between those two things is where every 2D-to-3D pipeline actually spends its effort.
What a plan does not tell you, and what a 3D view cannot be built without:
- Height. Not marked on almost any residential plan. Ceiling height, window sill height, door head height, and whether a wall is full height or a half-height divider all have to be assumed.
- Wall thickness. Drawn as two parallel lines whose separation may or may not be to scale, and interior partitions differ from exterior walls in ways the drawing often does not distinguish.
- Which side is which. A door swing tells you the hinge side. It does not reliably tell you which room is inside.
- What the symbols mean. Symbol conventions vary by country, by practice, and by whoever drew it. A rectangle with a cross is a window in one drawing set and a service riser in another.
None of this makes the job impossible. It means the output is a plausible reconstruction constrained by a drawing, and being honest about which parts are measured and which are inferred is the difference between a useful tool and a confident liar.
The scale trap, and why it bites hardest
The single most consequential number in the whole pipeline is pixels-per-metre. Get it wrong and every downstream dimension is wrong by the same factor, silently and consistently, which is far worse than an obvious error.
The tempting approach is to take several known measurements from the drawing and average them into one scale factor. This is a mistake, and it fails in a specific, instructive way. If two calibration samples disagree, averaging them produces a scale that matches neither. A plan calibrated from one wall marked 9.27 metres and another marked 16.94 will render rooms at dimensions that appear in neither the drawing nor the calibration, and because the result is internally consistent it looks correct until someone measures it.
The better rule: if calibration samples disagree beyond a few percent, do not blend them. Prefer the longest measured span, because proportionally it carries the least error from imprecise endpoint placement, and surface the conflict rather than smoothing it away. A tool that says "these two measurements disagree" is more useful than one that quietly splits the difference.
Furniture is how people judge room size
This is the finding that surprises people, and it changes what you optimise for.
Nobody walking through a 3D tour is estimating metres. They are reading the room against objects whose size they already know. A sofa, a door, a kitchen counter, a bed. Those objects are the measuring stick, and if the furniture is the wrong size relative to the room, the room feels wrong even when its dimensions are exactly right.
The practical consequence is that furniture scaling deserves as much care as wall placement. A dimensionally perfect room furnished with slightly oversized pieces reads as cramped. The same room with undersized pieces reads as a warehouse. Both are geometrically correct and both fail the only test that matters, which is whether a buyer believes the space.
It also means a furniture item that cannot be scaled plausibly into a room is better flagged than silently squeezed in. Moving a misfit item to make it fit produces a room that is subtly, unaccountably odd.
The four approaches, and what each costs
Anyone solving this picks one of these, and they are not interchangeable.
| Approach | Needs | Strength | Limitation |
|---|---|---|---|
| Physical capture | A finished building and a site visit | Photographic truth | Useless before completion, and every revisit is another visit |
| Generated imagery | One reference image per view | Fast and cheap per frame | Views are independent, so the same room can disagree with itself between angles |
| Manual CGI | An artist and weeks | Complete control | Cost and turnaround scale linearly with the number of views |
| Reconstructed geometry | A parsed plan and a render pipeline | Consistent from every viewpoint, and walkable | Heights and materials are inferred, and rendering is compute-bound |
The distinction that matters for a tour specifically is consistency. A tour is not a set of pictures, it is one space seen from many positions. Generated stills can look better frame by frame and still fail as a tour, because a sofa that moves between viewpoints breaks the illusion of a room in a way no amount of per-frame polish repairs.
The licensing problem nobody mentions
If you are evaluating this space, check the licence on the training data and the assets before anything else. It is the constraint most likely to stop a project after the technical work is done.
The best-known public floor plan dataset for this task, CubiCasa5k, is released under a non-commercial licence. That is fine for research and disqualifying for a commercial product, and it is easy to build something on it before reading the terms. The same applies to furniture assets: every model placed in a rendered room needs a licence that permits commercial use, and needs its author and source recorded, because "we found it online" is not a provenance record.
Budget real time for this. It is unglamorous and it is the difference between a demo and something you can sell.
What to ask a vendor
If you are buying rather than building, these questions separate the serious from the demo-ware:
- What happens when my plan has no scale bar? Every plan eventually does not. The answer should involve you supplying a known dimension, not the tool guessing.
- Where do ceiling heights come from? If the answer is a default, ask what it is and whether you can change it.
- Can I correct the parse? Automated room detection will misread something. If you cannot fix a wall or a room boundary, you cannot ship the output.
- Is it consistent between viewpoints? Ask to see the same room from two angles in the same tour, not two hero stills.
- What is the licence on the furniture? You are publishing these images commercially.
- What does a revision cost? Per-view pricing and per-project pricing behave very differently when a client asks for changes.
Where this fits against staging
Worth separating two things that get conflated. Virtual staging furnishes an existing image of a room, whether that is a photograph or an architectural render, and it is a mature, cheap, fast solution to a specific problem: an empty room that a buyer cannot picture furnished.
Reconstructing walkable geometry from a plan is a different problem with a different economic profile, aimed at the case where no image of the room exists at all. Off-plan and pre-construction is where that gap is real, because the capture-based tools cannot help and there is nothing to photograph.
For most listings today, staging the photographs you have is the answer. The plan-to-geometry problem is worth understanding because of where it is heading, not because it replaces the thing that already works.
Frequently asked questions
Can you generate a 3D model from a 2D floor plan?
Partially, and it is important to be precise about which parts. Wall positions and room boundaries can be derived from the drawing. Ceiling height, window sill height, wall thickness and materials are almost never marked on a residential plan, so they are inferred or assumed. The output is a plausible reconstruction constrained by the drawing rather than a measured model of the building.
Why do dimensions come out wrong even when the plan has a scale bar?
Usually because several calibration measurements were averaged. If two samples disagree, the average matches neither, and because the result stays internally consistent it looks correct until somebody measures it. Calibrate from the longest measured span, since it carries the least proportional error from imprecise endpoint placement, and surface disagreement between samples instead of smoothing it away.
Why does a room with correct dimensions still feel the wrong size?
Because nobody estimates metres. People read a room against objects whose size they already know: a sofa, a door, a kitchen counter, a bed. If the furniture is scaled wrongly relative to the room, the room feels wrong even when its dimensions are exact. Oversized pieces read as cramped, undersized pieces read as cavernous, and both are geometrically correct.
Why can Matterport not be used for off-plan properties?
Because capture-based tools need a finished room to photograph. Off-plan and pre-construction sales happen before one exists, which rules out the entire capture category regardless of quality. That gap is the main reason anyone attempts reconstruction from a drawing at all.
What licensing issues affect floor plan to 3D tools?
Two. Public floor plan datasets are often non-commercial: CubiCasa5k, the best-known one for this task, is released under a non-commercial licence, which is fine for research and disqualifying for a product. Separately, every furniture asset placed in a rendered room needs a licence permitting commercial use, with author and source recorded. Check both before building, because it is the constraint most likely to stop a project after the technical work is finished.