Research · · 8 min read
How AlphaRender holds your geometry
The reason most AI renders cannot be used is not that they look bad. It is that they are pictures of a different building. Here is what is locked, what is free, and how to check.
Every architect who has tried a general image model has the same story. The picture is beautiful. The building in it has five floors instead of four, the entrance has moved to the other end, and the cantilever the whole scheme is about has quietly become a column. You cannot show it to a client, because it is not your project — it is a picture in the style of your project.
That is not a bug in those tools. It is what they are: they turn a description, or an image treated as a description, into a new picture. Nothing in that process is obliged to keep your dimensions, because nothing in it ever knew about them.
The model is read as geometry, not as a picture
The difference here starts at the upload. What goes in is an OBJ or an FBX — a set of surfaces in three-dimensional space, with a camera. It is not a screenshot of your viewport. The massing, the proportions, the position of every opening and the viewpoint you framed are all measured out of that file, and they are fixed before anything is generated.
Generation then happens on top of a scene that is already spatially decided. The run is not asked what the building looks like; it is asked what this building, from this camera, looks like in this light, in these materials, in this weather. That is a much narrower question, and narrow questions are the ones machines answer reliably.
What is fixed and what is free
The clearest way to think about a run is as two layers with a hard line between them.
- Fixed: massing and volumes, floor-to-floor and overall proportions, the position and size of openings, the camera position and lens. These come from your file and no preset can move them.
- Free: time of day and sky, materials and their colour, atmosphere and weather, planting, vehicles and people, and everything beyond the site boundary.
- In between: small elements that the model implies but does not describe — a handrail detail, the profile of a mullion, the exact leaf of a tree. These are generated, and they are the ones to look at twice.
This is also why a set of twenty directions is comparable. Every image in the set is the same building from the same point; the only variables are the ones you were trying to explore. Reading that set would be pointless if the massing changed between frames — you would be comparing photographs of twenty different projects.
Standard and Creative
Two modes decide how much freedom the free layer gets. Standard follows the model and any references closely, which is what you want once the design is settled and the image has to be defensible. Creative allows bolder moves in mood, palette and material — useful on the first day of a competition, and worth a second look before it goes to a client, because “bolder” includes the in-between elements above.
How to test it on your own project
Do not take a claim about accuracy on faith. The check takes five minutes and it is the same one a good visualiser does on their own work:
Render the clay
Take a screenshot of your model from the camera you are about to export — the flat grey viewport shot, nothing else.
Generate
Upload, generate a run, and pick any frame from the set. Do not pick the prettiest one; pick a random one.
Flip between them
Put the two images on top of each other in any viewer and switch back and forth. Watch four things: the roof line, the head heights of the windows, the ground line at the entrance, and the horizon. Drift shows up as movement — a fixed geometry reads as one image with the lights turned on.
Repeat on a second frame
One frame is an anecdote. If frames three and seventeen from the same run both land on the same silhouette, the lock is doing what it says.
“AlphaRender has become an essential part of my interior design workflow, significantly speeding up the revision process while preserving the original composition and proportions better than any other AI tool I have used.”
Kateryna Shumikhina, CEO & Founder, k3nstudios.com
Where the limits are
Honesty is more useful here than a slogan, so: what is not in the model is not in the render. A canopy you did not draw will not appear because the brief mentions it, and a mullion pattern you did not model is being invented plausibly rather than reproduced. If a detail is contractual, model it — the lock protects what you gave it, and only that.
The other limit is scale. Very small elements sit below the resolution the run is working at, which is one of the reasons the upscale is a separate step: detail that was implied at 1080p is resolved on the frames you keep, not guessed at across the whole set.
Why this is the whole product
Speed is what gets attention — 68 seconds, 10–20 options, a revision round in 41 minutes. But speed without accuracy is just a faster way to produce images you cannot use. The reason the rest of the pipeline works is that the thing coming back is always your building, so every minute you spend after the run is spent on the design instead of on defending the render.
Questions we get
- Will the AI change my building’s proportions?
- No. Massing, proportions, opening positions and the camera are measured from the uploaded OBJ or FBX and fixed before generation. Light, materials, atmosphere and entourage are what vary between directions.
- Does the camera I set in my 3D editor carry over?
- Yes — the viewpoint saved in the file is the viewpoint the render uses, which is why framing in your own editor is worth doing before export.
- What about details I did not model?
- They are generated plausibly rather than reproduced. Anything contractual — a specific mullion profile, a particular handrail — should be modelled; the lock protects what you supplied.
- Does Creative mode break the geometry lock?
- No. Creative widens mood, palette and material freedom, not the geometry. The massing and camera are fixed in both modes.
