In early 2025 I published a version of this article that said tools “cannot produce images that meet the accuracy standards required for architectural visualization”. At the time that was correct. I am rewriting it because it stopped being correct. Not because AI became magic, but because I changed where it sits in the pipeline, and I expect to rewrite it again.
The short version: AI does not build the picture. It finishes it. Everything that makes a visualization right — geometry, camera, proportions, brand colours, where the designer put the furniture — still comes from a 3D model built from your drawings. Everything that makes it read as a photograph — material response, light falloff, reflections, atmosphere — is now largely an AI pass on top of that model, on my own machines, checked by an automated reviewer before a human picks.
Compare:
Drag the slider. One side is the working render out of Archicad, the other is the same image after the pipeline; the labels say which. Nothing about the design changed. The partition glass, the screens, the green poufs in the kids’ corner are where the designer put them.
Finishing. The last stretch from “correct render” to “image that sells the project” used to eat most of the schedule: materials, vegetation, people, atmosphere. ArchUpgrade, the system I built, takes the working render and produces photorealistic versions in about a minute per image per GPU, in controlled variations, each scored by a local vision model on five criteria before anyone looks. On a recent 18-view interior project it generated over 300 candidates, the reviewer filtered them, I picked 18.
Variants without re-rendering. Season, time of day, weather, furnishing style, budget tier, from the same base render with geometry locked. A night or dusk variant of an existing view is a fraction of a new view on the price list for exactly this reason.
Compare:
Relighting and sky work on finished images. Time-of-day relighting, light migration from a reference photo, sky replacement, material swaps: targeted tools on a finished render, not a re-render. Look passes, not physically measured ones. Good for presentation, useless for a daylight study.
Interiors from an empty model. For sales material, an empty 3D apartment can be furnished and finished by the pipeline from a viewpoint I set. Consistent across a building, a fraction of hand-dressing every flat type. The furniture is plausible, not specified; where a client sells a specific fit-out, I model it.
Renders from sketches and flat exports. A pencil sketch, a SketchUp or Archicad export, an old render from a previous phase: all usable as a starting point for a photorealistic image before any 3D work begins. Fast for competitions and early conversations; see AI Render Upgrade.
Compare:
Motion from stills. Camera moves, time-of-day and weather transitions generated from a still, cut into 30-second to 5-minute films: AI Animation. Right for social and sales-office loops. A flythrough through a building is still a 3D animation.
Upscaling to print. A 6-megapixel delivery upscaled to 4K or 8K with detail added rather than smeared, so an older render survives a hoarding.
Everything around the picture. AI narration and AI-composed music for films are line items now instead of a hunt through a stock library. Batch quality checks catch a wrong-resolution frame before a client does. And the scripting that runs my 3ds Max pipeline is written with AI assistants: Automation, MaxScript and AI in 3ds Max.
New viewpoints. An AI pass works from one image. It cannot walk around the building, look through a door it cannot see, or add the elevation you forgot to render. That is what the model is for.
Material-exact finals. The pipeline generates surface quality probabilistically. It will produce a convincing oak floor; it will not reproduce a specific oak floor from a specific supplier’s catalogue unless I model it. For a manufacturer’s own product or a listed facade, a full V-Ray render from the model is still the answer.
Geometry generation. Text-to-3D and image-to-3D are in my toolkit for props and context, and they have improved a lot. They do not produce a building from a plan. The model is still built from the drawings, by someone who knows what a stair needs.
Long, consistent video. AI clips are 3 to 8 seconds each and hold a camera move about that long. A ninety-second film is an edit of many clips. A continuous walkthrough with exact geometry in every frame is a 3D animation.
Judgement. Which camera tells the story, which hour the building looks like itself, whether an image survives a planning committee. Nobody has automated that, including me.
This list is the least reliable part of the article. Something on it will be wrong within weeks. I check.
- ArchUpgrade, my own render-to-photograph system: multi-GPU ComfyUI farm, structured prompts, version tree, automated five-score review, delivery packaging. Daily use since early 2026. Technical write-up.
- Relighting, sky and material tools, generative expand, upscalers, camera-angle passes: the same system, on finished images.
- Local models only. Every model is open-weight and runs on my hardware. No client image is uploaded anywhere. Each model is classified by licence, anything non-commercial is locked out of paid deliverables at the configuration level, and every delivered file carries a record of the tools that touched it.
- A vetting database of AI tools, re-checked weekly for release state, licence and hardware needs: how I vet the flood.
- Scheduled agents that run generation and review overnight, with me reading the results in the morning: the boring version of AI automation.
- Unreal Engine 5 for VR and interactive deliverables, with AI denoising and upscaling in the real-time path.
- AI coding assistants for every MAXScript and Python tool I build, plus an MCP server that lets an assistant drive 3ds Max directly.
The job moved. The picture used to be made once, by hand, over days. Now it is made many times, by machine, in minutes, and the craft is in the model that anchors it, the constraints I put on the generator, and the choice of which result to show you.
It does not replace 3D work. It is built on top of it. The accuracy comes from the model. The judgement is human. What changed is the iteration count and where the hours go.
When AI touches your project, I say so. An enhanced render from your approved model, an AI variant of an existing view, a concept image from a sketch: each is labelled as what it is, and each delivered file carries its provenance.
If you want to see what this does on your project, get in touch.