Oct 2026 ยท Technique

The render is not the building

AI image models are very good at making a scheme look better, and often they do it by quietly changing the scheme. We laid the Soldagen model over its own final images to see what held and what moved.

The image that wins the room

An AI image model does not know which parts of a picture are the design. It sees a building, a coastline, a sky and some steam, and it is equally willing to improve all of them. Ask for something more atmospheric and it will lower a parapet, add a bay, deepen an overhang or tidy a neighbour's roof, because those are the changes that make an image more convincing. It has no way of knowing they are also the changes you are not allowed to make.

This is the default behaviour of the tools, and it matters because the image that comes out can be more persuasive than the scheme that went in. It may win the room, but once someone who knows the drawings looks closely, the lowered parapet and the extra bay are the first things they see.

The more useful question to ask of an AI image is which parts of it were decided by the design and which by the image model. Below, we answer it for Soldagen: In the Interval, the bathhouse that took first place in GENDO 48.

Starting from the model

Every image in the Soldagen set started as a view of the detailed 3D model, developed to the level of the hoods, pools, walkways, floor levels and glazing. That model sat in a site model of Vindstad, with the mountain and the fjord in it, so the camera, horizon and skyline were fixed before any image was generated.

All of the image generation was done in Gendo, using the model view as the input the generation had to hold to. The model settled the building's geometry. Gendo was left to supply the conditions around it: weather, light, water, steam, the wear on the materials, and the people.

In the slider below, the model view sits underneath the final image. Drag across to compare them.

What each side was allowed to decide

The distinction is easier to state than to apply, because much of what makes an image persuasive sits close to the line. The table below sets out where we drew it for Soldagen.

Fixed by the modelLeft to the image
The buildingRoofline, the curved hoods and their spacing, pool forms, walkways, floor levels, glazing linesSurface finish, weathering, light inside the rooms
The siteCamera, horizon, mountain skyline, the line of the shoreSnow, sky, the surface of the sea, the texture of the rock
The atmosphereNothingTime of day, steam, water spilling from the pools
The peopleNothingWho is there, what they are wearing and, in the section, the furniture
Who answers for itThe designThe image

The right-hand column still matters. Light and weather shape much of how a building is experienced, and on a site where the whole brief is about the absence of the sun, they come close to being the design. Even so, they describe conditions around the building, and a jury expects that column to be interpreted.

That was the intention. The next step was to check whether it held.

Laying one over the other

To test whether the model really fixed the geometry, we traced the edges of the model view and laid them directly over the finished image. Where a red line sits on an edge in the picture, the image held. Where a red line runs across open water or through a rock, something moved.

The building matches. Every hood along the row lands on its line and the round bathing towers sit inside their outlines. The walkway rail follows the model so closely that the two are hard to tell apart.

The ground moved. In the model the walkway loops out over open water, with its posts standing in the fjord. In the image a rock shelf has grown underneath it, and the flat ground in the foreground has become boulders and rock pools, with the model's shoreline running straight through them. We did not ask for this. The image model inferred a rocky shore, which suits Lofoten but departs from the drawing.

For a competition image we let it stand. The rocks are true to the place, and nothing in the scheme depends on exactly where the water meets the stone. In a planning view we would have rejected it, for the reasons set out below.

The section nobody photographed

The second pair raises a different issue. The section published with the project looks like a photograph of a card and timber model on a table, with the mountains behind built up in stacked contours. In fact it is a section render from the same 3D model, re-rendered in Gendo to resemble a physical model.

This pair held more closely than the terraces. The roofline steps where the model steps, the curved end of the first hood is where the model put it, and the floor plates drop down the slope at the same levels. Even the mountain skyline behind is the model's, translated into card.

What the image added is the interior. In the model the rooms are empty glass; the image furnished them with chairs, tables, a stair, figures in robes at the windows and a bather sitting on the edge at the far end. None of this was designed; the image model supplied a plausible reading of how a bathhouse is used. A section full of people makes claims about how a building works, and here those claims came from the image model.

Where the line sits

We continue to use AI for imagery, but the line between what the design decides and what the image decides has to move with the purpose of the image.

In a competition, the jury is judging an idea and knows the images are interpretations. Atmosphere and context can take some licence, provided the building matches the model.

In planning and consultation the context matters as much as the building, because the question being asked is what the building does to the place. That is why verified views exist: images built from surveyed camera positions, so that what a committee is shown can be checked against the site. That process exists to catch a coastline that has quietly moved, and the terraces image shows how readily an image model will move one.

In our planning and consultation work, the context therefore comes from survey and photography, and the overlay check covers the ground as well as the building. Our AI policy limits generative work to elements that cannot affect what is being represented.

The limits of the overlay

The overlay has two limitations. First, it checks the image against the model; it cannot judge the design itself.

Second, an accurate image can still mislead. The terraces image shows a low sun on the horizon and clear air, at a site where for three weeks each December the sun does not appear at all. The return of the light is the subject of the scheme, but every image depicts a chosen moment, and that choice shapes how the scheme is read. Choosing the moment remains an editorial judgement, with or without AI.

The full project, with the film and the research, is published as Soldagen: In the Interval. How we hold geometry under AI work is set out under AI architectural visualisation. To discuss a scheme that needs images people can check, contact the studio.

←All of AF_LAB Start a project→