AI in archviz, an honest year end review of what actually helped and what wasted our time

A year ago our studio decided to be deliberate about which AI tools we brought into the pipeline instead of just piling them in and seeing what stuck. We tracked what got used, what got quietly abandoned, and what changed how we work. This is that review. No sponsors, no affiliate links, just what happened.

The tool that changed our workflow most was Intel Open Image Denoise 2.3. Not because it is new. We have been using OIDN for years. What changed is version 2.3, which came out in late 2025, is genuinely better at preserving fine detail than earlier versions. We used to render our hero stills to a much higher sample count than needed just to avoid denoiser artefacts. With 2.3 we can drop samples by about 30 percent and get the same or better final quality. That is a real time and cost saving on every project.

The tool we thought would change everything and mostly did not is Midjourney for final delivery. We tried, on three separate projects, to use Midjourney concepts as the actual client deliverable when the architecture was still in early sketch phase. It failed twice and worked once. The two failures were both because the client, once they saw the images, wanted them to be accurate to the eventual building, which of course Midjourney cannot do. The one that worked was a property developer pitch deck where accuracy was explicitly not the point. So Midjourney is in our pipeline for concept and pitch work only, never for final delivery.

The tool we use every single day now and did not use a year ago is ControlNet with SDXL for concept variation. This is the one AI workflow that has genuinely changed how we handle the middle phase of a project. Between rough model and final render we used to do a lot of Photoshop overpaint to test moods. Now we do it in ComfyUI in a fraction of the time. Every artist in the studio has a ControlNet setup on their workstation.

The tool that was hyped hard and disappointed is realtime AI style transfer in Twinmotion and Chaos Vantage. It looks great in a demo. In practice, on a real client presentation, the style transfer does unpredictable things to specific materials, and clients notice. We tried it live in two presentations, went back to standard PBR both times, and have not touched the style transfer feature since. Maybe next generation.

The tool that surprised us was Topaz Video AI Rhea model for animation upscaling. We were sceptical because we had bad experiences with earlier upscalers producing plasticky output. Rhea is genuinely different. On a 1080p animation upscaled to 4k it produces output that is indistinguishable from a native 4k render in most cases, and takes about a fifth of the time. This has changed how we quote animation projects. We now render most animations at 1440p and upscale, which cuts render times enough to make animation projects viable at price points where they were not before.

The tool everyone in the studio quietly stopped using is text based scene generation. There were several tools this year that promised you could describe a scene and get a full 3d file. We tried three of them across six test projects. The output was, without exception, a starting point that took longer to fix than modelling from scratch would have taken. Not because the tools are bad, but because architectural work requires precision from the first click. Generative scene tools do not do precision. They do vaguely convincing, which is exactly what we do not need.

The tool we are still arguing about is Nvidia RTX Neural Materials for real time presentation. On paper it is compelling. In practice it requires specific hardware setup and material authoring workflows that our team is not set up for yet. Half the studio wants to invest a week in learning it. The other half wants to wait for the second generation before committing. Probably we will invest that week in Q1 next year.

The workflow question that came up repeatedly is whether AI has changed how we quote and price projects. The honest answer is yes, but not in the way we expected. AI has made concept phase cheaper and faster, so we bundle more concept variations into the same quote without changing price. Final delivery is roughly the same cost because the pipeline still relies on traditional rendering. Animation has become genuinely cheaper thanks to upscaling. Net effect is that we deliver more per project, which is what clients notice, more than we deliver at lower prices.

The team question is worth addressing. There was a real worry a year ago that AI would replace junior artists. It has not. What it has done is shift what junior artists spend time on. Instead of hours of Photoshop overpaint for concept variations, they run ControlNet workflows. Instead of building simple entourage from scratch, they source or generate and integrate. The judgment calls, the client communication, the lookdev decisions, the final grading, none of those have moved. The craft is still the craft, the tools around the craft have shifted.

The one prediction I feel confident about is that this year, the difference between studios that use AI tools well and studios that ignore them will start being visible in the work. Not because AI produces better renders. It does not. But because studios using AI well ship more concept options, more animation shots, and iterate faster on client feedback. The output looks the same. The workflow is faster. Clients will feel that difference even if they cannot name it.

The last thing to say is that this is a year end review, not a permanent verdict. Every tool in this list is on a version that will not exist in six months. What matters is having a habit of testing new tools on real client work, keeping the ones that earn their keep, and quietly deleting the ones that do not. That habit is the actual competitive advantage, not any specific tool.