How AI Actually Fits Into a 3D Walkthrough in 2026

Half the studios selling an "AI 3D walkthrough" right now are selling a slideshow of AI stills with a slow zoom on each one. It looks like motion until you watch it, and then it falls apart, because an AI image generator has no idea your boardroom is the same boardroom from one shot to the next. I run AI in production at 100CGI Studio, on several RTX 5080s and RunPod serverless with custom ComfyUI workflows, mostly with Nano Banana Pro, and I have fine-tuned our own models. So this is where AI genuinely earns its place in a walkthrough, and where it still cannot do the job.

Start with the honest split, because it is the thing the hype skips. AI is excellent for concept and mood, and useless for an accurate as-designed render from drawings. Those are two different jobs. When a design-and-build firm needs to show a client exactly what was specified, the exact partition layout, the real materials, the lighting as it will actually fall, that comes off our 3ds Max, V-Ray and Chaos Vantage pipeline, not off a prompt. Get that wrong on a £4 million fit-out and you have shown the client a building they are not getting.

Where AI does real work in the pipeline

Inside that pipeline, AI does plenty, and we use it daily.

Voiceover is the clearest win. Every office walkthrough we produce ships with AI voiceover, and it carries the design story well enough that the film works in an email or on a reception screen with nobody from your team narrating live. The practical benefit is speed: when a brief changes at 6pm before a 9am pitch, I can redo the script and the voice track in minutes instead of booking a studio and a human voice artist for a re-record that will not happen in time.

People are the second. A walkthrough of an empty office feels like a morgue. Populating it with believable people used to mean expensive 3D crowd assets or stiff library figures. AI people integration, done carefully, puts plausible humans into the shot at a fraction of the cost and time, which is why it sits in our standard walkthrough package rather than as a luxury add-on.

AI also speeds the heavy technical steps. Upscaling, denoising, generating quick concept frames to agree a look before we build, spinning variations on a mood. A pass that used to tie up an artist for most of a day now takes an hour or two, so a 60-second sequence reaches first draft faster without cutting the quality that matters. The craft still comes from a real 3D artist. AI just clears the technical overhead so that artist spends their time on the decisions that sell the room, not on waiting for a machine. On a Publicis project, an office across several floors for over 1,000 staff, that speed is what let us invent a new presentation method, Dynamic Flow Video, instead of grinding through manual busywork.

The as-designed line nobody should cross

Here is the call I will not change, because clients get burned on it. For the shots that have to be accurate, you cannot let a generative model improvise.

I have used ControlNet to AI-render from drawings in production, accurate to the as-designed layout, and it works because the geometry is locked first. The AI is colouring inside lines we drew, not inventing the room. The moment you hand the whole job to a text prompt, it hallucinates a beautiful office that is not your office: a window that does not exist, a ceiling height that is wrong, a material the client never chose. For a marketing teaser where vibe is the point, fine. For a tender where the client will hold you to what they saw, that is how you lose trust and the next project with it.

So the rule we run is simple. Real geometry and real spec drive the accurate shots. AI accelerates everything around them.

Why AI did not kill arch-viz

The fear I hear most is that AI has made studios like mine obsolete. It has not, and the evidence is in who is still standing. AI has not replaced manual arch-viz for premium pitches. What it did do was wipe out the studios that competed on price instead of service, the ones already shipping cheap, forgettable renders. Their work was the easiest to replace with a free image generator, so it got replaced.

Clients making their own "good enough" AI images is not the threat people think either. We cook at home and still book a table on Friday. A director assembling a few AI frames for an internal review is not commissioning the film that wins a 1,000-person office. Different stakes, different tool.

What I have not cracked

To keep myself honest: I have not run Gaussian Splatting on a live client project. I wanted it for a real window view on a scheme, the actual city outside the glass, but you cannot fly a drone in most central locations now to capture the plate, so it stayed an experiment. And I have tried every major AI video tool on real work, Runway, Sora, Kling, but none of them yet holds a consistent, accurate office across a 90-second continuous move the way our render pipeline does. For a controlled architectural walkthrough, the hybrid approach still wins: real 3D for accuracy, AI for everything that makes it faster, cheaper and more human.

For how this comes together on a real office pitch, see our 3D walkthrough service.