AI denoisers in archviz: OptiX, Intel Open Image Denoise, and where each one earns its keep

The AI denoiser conversation is usually framed as OptiX versus OIDN, faster versus cleaner, and that framing is close enough to be useful and wrong enough to lose you jobs. I have watched both denoisers save frames and destroy them in the same week, so this is the honest working breakdown I give to junior artists on our team.

OptiX is Nvidia's temporal AI denoiser, trained on ray traced beauty passes and running on the tensor cores of any RTX GPU. It is fast enough to feel free inside an interactive preview. Corona, V-Ray, Arnold, Redshift, Cycles, and Octane all ship an OptiX path. Its strength is speed at low sample counts, which makes it the default for lookdev, IPR, and animation previews.

Intel Open Image Denoise runs on the CPU or on an Arc GPU, is open source, and updates on a roughly quarterly cadence. Version 2.3 introduced a substantially better model for high frequency detail and is the version I am talking about below. It is slower per frame than OptiX but preserves fine geometric edges, mullion patterns, and fabric weaves in ways that OptiX will happily eat.

The single most useful rule I know is this. Use OptiX for anything moving and OIDN for the final still. On animation frames the temporal component of OptiX is genuinely helpful, and small edge losses average out over 24 frames a second. On a hero still where a client is going to zoom to 200 percent on the balustrade, OIDN gives you cleaner detail with less of the plasticky, over smoothed look that dates a portfolio piece within a year.

Watch for these specific OptiX failure modes when reviewing a frame. Firstly, thin mullions and window frames turn into a soft blur, especially against a bright sky. Secondly, fabric weaves and rug pile lose their high frequency texture and go flat. Thirdly, small specular highlights on metal hardware get smeared into halos. Fourthly, foliage in shadow tends to bloom slightly, giving you a green haze around the canopy. None of these are fatal in an animation, but on a still they read as amateurish.

OIDN has its own failure modes, mostly at very low sample counts. Below about 40 passes on a complex interior it can produce blotchy patches in indirect light, particularly on white walls with soft light bouncing from a coloured floor. Push samples to 80 or 120 and the blotches vanish. Also, OIDN does not have a temporal model in the standard implementation, so on animation it produces a slight flickering that requires denoising the whole sequence together in a compositor.

The workflow I run in production is a two pass approach. First, iterate lookdev in Corona or V-Ray IPR with OptiX enabled so the preview is responsive. Second, for the final render, switch to OIDN 2.3 with adaptive sampling at 5 percent noise threshold and let the CPU denoiser take an extra 40 seconds a frame. On a 4k hero still that is a tiny cost for a visibly cleaner result. On animation, stay with OptiX temporal and denoise the final sequence in Nuke or DaVinci Resolve if flicker becomes an issue.

Blender users have the same choice built into Cycles. OptiX denoiser sits in the render properties, OIDN under compositing or the viewport denoise setting. My recommendation for Blender archviz is to use viewport OIDN during lookdev because Blender's OptiX viewport can be unreliable on some driver versions, then final render with OIDN in the compositor at prefilter accurate. The prefilter accurate setting reduces the blotchy artefacts on low sample interiors substantially.

One AI denoiser that deserves a mention outside the OptiX and OIDN debate is Nvidia's newer neural denoiser bundled with Chaos Vantage and some other real time engines. It is a different model, trained specifically for interactive real time output, and it does a genuinely better job than OptiX on the specific problem of stable low sample real time rendering. If you are presenting to clients in Vantage or a similar tool, that denoiser is what makes the live experience feel finished rather than noisy.

The pricing question is worth naming. All three of these AI denoisers ship free with their host engines. There is no licensing decision to make. The decision is purely a quality and speed tradeoff on a per shot basis, which is why I have argued for years that render engine choice should not be driven by denoiser marketing. Both OptiX and OIDN are excellent tools and both belong in a working archviz pipeline.

The last piece of practical advice is to always keep a beauty pass render output without denoising when you deliver a final still to a client for print use. Print reproduction can reveal denoiser artefacts that are invisible on screen, and having the raw beauty available means you can redo the denoise in post at higher quality if a printer flags an issue. Set your output to save both the denoised and raw versions in the same EXR, which every major renderer supports through the render elements or AOV panel.