Every visualization project hits the same wall. The 3D model is accurate, the cameras are approved, the layout is final, and then the polishing begins. Materials, lighting, vegetation, people, atmosphere. That last stretch from “correct render” to “image that sells the project” is where most of the time went.
Over the past year and a half I built a system, ArchUpgrade, that changes the economics of that stretch. It takes working 3D renders and turns them into photorealistic images, in controlled variations, with automated quality scoring on every output. Daily production use since early 2026. It runs on my own hardware, on my own farm, and no client image leaves the building.
Drag the slider. One side is the working render out of Archicad, the other is the same image after the pipeline. Same geometry, same camera, same layout; the labels say which is which.
Nothing about the design changed. The furniture is where the designer put it. Signage, screens, partition glass, the little green poufs in the kids’ corner, all preserved. What changed is what makes an image read as a photograph: material response, light falloff, soft shadows, reflections, the sense that a camera stood in the room.
Variations. Lighting moods, seasons, furnishing styles, budget tiers, from the same base render. Where I used to show two or three polished images, I now show a curated selection from dozens. These four are one source render, one batch:
Four variations of one bank branch render, one batch.
Relighting without a re-render. New since I first wrote about the system. A finished view can go to golden hour, blue hour or night — interior lights on, street lamps on, sky replaced — as a pass on the image, scene locked. The three pairs below are archive renders put through the current pipeline while writing this. Nothing was re-rendered.
These are look passes, not lighting simulations. The pipeline re-tints materials under the warm light it adds, so a daylight study still comes from the 3D model. For a brochure that needs the building at dusk, it is a fraction of the price of a night render, and the price list says so.
Quality check before a human picks. Each generated image is scored by a local AI reviewer on five aspects: aesthetics, fidelity to the source design, prompt adherence, technical quality, style match. Images below threshold are retried automatically. The review screen shows every candidate with its scores, so the final pick is made from a field that already passed inspection.
The design stays under control. This is the part I care most about, and the part most “AI rendering” services get wrong. The pipeline works from the accurate 3D model. It enhances, it does not invent, unless I instruct it to. Walls do not move. Proportions do not drift. Brand colours stay brand colours. When a generated image deviates from the source design, the fidelity score catches it and the image is rejected.
A larger space, an office atrium, through the same pipeline:
Your material never leaves the network. Every model is open-weight and runs on my own GPU farm. Nothing is uploaded to a third party, and there is no per-image cloud cost quietly deciding how many variations I am willing to try. Every model is classified by licence; anything non-commercial is locked out of paid deliverables at the configuration level. Each delivered file carries a record of the tools that touched it.
- I build the 3D scene as always. Accurate model from your plans, approved cameras, correct proportions. This part does not change. Everything else depends on it.
- The pipeline generates candidates. For each approved view, a set of photorealistic versions across the moods and styles we agreed. A version renders in about a minute on one GPU; the farm runs many in parallel.
- Machines filter, humans choose. Automated scoring removes the failures; I review what is left and pick. You can sit in on that round if you want to.
- Finishing on the picked images. Relighting, sky, generative expand to a new aspect ratio, upscaling to print: passes on the chosen image rather than a return to the 3D scene.
- Delivery. A clean, numbered package, with the full generation history archived on my side. If you come back in three months and want “that version, but in evening light”, I branch from exactly that version.
The same system carries the rest of a presentation package: AI motion from stills for short films, AI narration and AI-composed music for them, upscalers for print, and experimental 3D tools, mesh generation for props, Gaussian splats for captured context. Those are line items on the price list where they are production-ready and internal experiments where they are not. I say which is which.
It does not replace 3D work. It is built on top of it. The accuracy comes from the model. The judgement about which image tells the right story is still mine, or yours. What the pipeline changes is the iteration count: more variations, tighter quality control, and a shorter polishing stretch at the end of every schedule.
I built it because I render every day and know exactly where the time goes. Since this year it also runs on other studios’ renders: ai.visiongraphics.eu. The engineering is in the technical companion. If you want to see what it does on your project, get in touch.