Ask an image model for an interior and it will hand back something that reads as a room. Measure it and the room falls apart. The ceiling and floor planes do not agree on a horizon, a run of downlights fails to converge toward the same point, and the table in the foreground belongs to a slightly different camera than the wall behind it.
This is not a rendering error in the ordinary sense. The model never constructed a space and then photographed it. It produced an image that resembles photographs of spaces, and perspective is one of the patterns it imitates rather than one of the rules it obeys. The same gap runs through everything a model does with light. Understanding that difference is what makes it possible to work with the tool productively, because it points at the fix: geometry has to come from outside the model.
The three failures worth recognising
The horizon drifts. In a correctly constructed perspective, parallel horizontal lines receding in the same direction converge towards a common vanishing point, with the relevant vanishing points positioned along a consistent horizon line. Generated interiors routinely carry two or three competing horizons — the ceiling reads from one eye height, the floor from another. It is subtle enough to survive a quick look and obvious once you trace a line along the ceiling junction.
Repeating elements do not converge. A row of pendants over a bar, a track of spots along a corridor, a grid of ceiling luminaires: these are the elements that expose the problem, because a regular array in perspective has to compress as it recedes at a specific rate. Models tend to keep the spacing roughly even, or compress it inconsistently, and the array reads as slightly wrong without the viewer being able to say why.
Objects carry their own cameras. A luminaire drawn from slightly below in a scene viewed from slightly above, a fitting whose ellipse implies a different viewing angle than the ceiling it sits in. Each element is individually plausible and the set is not mutually consistent.
Constraining the model with real geometry
Longer prompts do not fix any of this. A more detailed prompt may help, but it cannot reliably enforce consistent three-dimensional geometry. Supplying a photograph, model view or structural guide gives the image model a much stronger spatial reference. What works is supplying the geometry rather than requesting it.
Start from a real image. A photograph of the actual space, or a SketchUp view of the modelled space, gives the model a correct perspective structure to work within. Image-to-image work keeps the room and changes its treatment, which is almost always the real requirement — the space is fixed, the lighting is the variable. This is the single highest-value technique, and it is why concept imagery built on a project photograph holds up while imagery generated from a description does not.
Constrain with a line drawing or depth information. Where the source image needs to change substantially, a structural guide extracted from it — an edge map, a depth map, a simple line drawing of the room — can be supplied alongside the prompt so the model can reinterpret surfaces and lighting while more closely preserving the underlying geometry. The mechanics vary between tools; the principle does not.
Change one region at a time. Regenerating an entire image to adjust one wall reintroduces error everywhere else. Masking the region to be changed and leaving the rest untouched preserves the parts that were already correct. It also makes iteration reviewable: each version differs from the last in one respect, so it is possible to say which change helped.
Mark up the image rather than describe the change. When an output is nearly right, annotating the image itself — an arrow to where a beam should fall, a circle around a fitting to be removed — communicates the correction more reliably than another paragraph of prose. Spatial instructions given spatially survive the round trip; the same instruction in words tends to be interpreted somewhere else in the frame.
Where a lighting scene is harder than a general interior
Perspective errors in a furniture render are cosmetic. In a lighting image they propagate, because the light patch is a geometric consequence of the fitting.
A luminaire with a nominal 26° beam angle, mounted and aimed at a given position, produces a light distribution whose approximate size and location can be predicted from its geometry and photometric data. Its edge quality and visual appearance also depend on the optical system, surface material and surrounding conditions.
Where the generated image and that prediction disagree, the consequence is commercial. An unconvincing chair costs nothing. An image showing a wall grazed evenly from top to bottom by a fitting that could not physically do it sets a client expectation that the eventual installation will not meet.
The same applies to shadow direction, to the way a beam interacts with a textured surface, and to the visible source itself — a fitting shown without glare in a position where it would be directly in a diner's field of view.
The review checklist
Every generated lighting image gets checked against the design before it goes anywhere near a client.
- Fitting positions against the reflected ceiling plan — does each visible luminaire correspond to something in the layout?
- Patch geometry against the specified beam angle and mounting height — is the pool of light the size that fitting would produce from that height?
- Shadow direction consistency — do all shadows agree on where the sources are?
- Source visibility — is a fitting shown in a position where a seated or standing occupant would look straight into it?
- Surface behaviour — does the material respond to light the way the actual specified material would?
- Unsupported elements — has the image introduced any fitting, aperture or lighting effect that does not exist in the actual design?
An image that passes these checks can provide a fair representation of the design intent. It is still not a calculation, and photometric performance must be assessed separately in software such as AGI32 or DIALux EVO against the relevant project requirements and Australian Standards.
FAQ
Why does AI get perspective wrong so often?
Because it imitates perspective rather than constructing it. A render engine builds a three-dimensional scene and projects it through a camera; a generative model produces pixels that resemble such projections. Nothing enforces a single consistent viewpoint, so local plausibility and global consistency come apart.
How do you keep the room correct when generating lighting concepts?
Work from a real image of the space — a photograph or a view from the model — and change the lighting treatment rather than generating the room from a description. Where larger changes are needed, supply a structural guide such as an edge or depth map so the geometry is held while the surfaces are reinterpreted.
Can AI images be used in a client presentation?
As a statement of intent, yes, provided they are labelled as concept imagery and have been checked against the actual design — fitting positions, beam geometry, shadow consistency. They should not be presented as a prediction of measured performance.
Is a longer, more detailed prompt the answer?
Rarely. Prompt detail influences style and content; it does not impose geometry. Supplying the geometry as an image is more effective than describing it in words, and annotating an existing output is more effective than rewriting the description.
SL Design Consultant is an independent lighting and electrical design consultancy in Sydney, delivering lighting and electrical design across commercial, hospitality, retail, heritage and public projects. Our product-neutral recommendations are guided by design intent, technical performance and project requirements.