Image generated with AI
Denoisers in Maya and Blender: Clean Noise Fast
Diego Cortés
Author
A denoiser can make a 200-sample render look like 2,000 samples, but only if you find it first. In Maya it is not in Render Settings at all, and in Blender it is one checkbox plus the right passes. Here is the map for both engines.
What a Denoiser Does (and What It Does Not)
A denoiser returns a cleaner image without asking for more samples: it works on the render you already computed, as a post-process step. It scrubs the grain sampling left behind, it does not change lights or bounces. And it has a limit: the noisier the input, the more detail it throws away. With very low sample counts the result looks clean but plastic, with textures fading out, as the Blender manual shows by comparing different inputs.
If noise survives the denoise pass, the problem is not the denoiser but the samples, the bounces or the fireflies, and those get fixed earlier with adaptive sampling and light paths (optimising a Cycles render). Denoising is the last step, never the first one. And denoise images or sequences, never an already encoded video: on compressed pixels it can only redistribute the error.
Denoising in Maya: Arnold's Denoisers Are Imagers
In Arnold the denoisers are not a dropdown in Render Settings. They are imagers: post-process nodes that operate on pixels before the output driver and that can be chained, per the Arnold documentation.
Where They Live (and Why You Cannot Find Them)
You configure them in Render Settings under Imagers, and on the Post tab of the Arnold Render View. There are four: imager_denoiser_oidn, imager_denoiser_optix, imager_denoiser_noice and imager_arnold_denoiser.
Why the OptiX Option Is Missing
It is the most repeated question on the topic and it rarely has a single cause. The four real reasons:
- You are looking in the wrong place. On the AOVs tab the only related control is the Output Denoising AOVs checkbox, which generates the diffuse_albedo, Z and N passes. The denoiser lives under Imagers.
- Hardware. OptiX needs an NVIDIA GPU with a recent driver and does not exist on macOS.
- GPU memory. Arnold will not use the GPU if less than 512 MB is free at startup; Min. Memory (MB) changes that, but a saturated GPU does not render better.
- A Maya preference. In new scenes MtoA adds OIDN by default, and it is turned off under Preferences, Arnold, Create Denoiser.
On an NVIDIA card, the full walkthrough is Denoise with OptiX in Maya.
OIDN: The One That Runs Anywhere
The OIDN imager picks the fastest device and runs on CPU or GPU: it covers Intel Xe, Xe2 and Xe3, NVIDIA Turing through Blackwell, AMD RDNA 2 to 4 and Apple silicon from M1, per the official project site. Watch the minimum drivers (NVIDIA 528.33 on Windows or 525.60.13 on Linux): that is the first place to look when OIDN silently falls back to CPU.
Noice: The Sequence Denoiser
Noice is not an imager, it is a standalone post-process program, and it is the piece that solves animation work in Maya:
- It works on multi-layer EXR with Merge AOVs enabled and returns an EXR with the denoised layers under a
_denoisedsuffix. - It reads several frames and several light AOVs at once, so it gives better temporal stability than denoising frame by frame: that is the key to flicker.
- It requires Output Denoising AOVs. The flow: render the sequence, open Arnold, Utilities, Arnold Denoiser (noice), choose Input and Output and pick Single frame, Frame range or Complete sequence. Saving from the RenderView will not work (rendering an image sequence in Maya).
- Pixel Search Radius sets how far it looks for similar neighbourhoods, and the output keeps only the main RGBA layers, so keep the original EXR.
For a single frame use the imager_denoiser_noice imager, as in Denoise with Arnold (Noice) in Maya.
Imager Order Matters
Denoisers go before any imager that introduces new information, such as a bloom; the imager_light_mixer goes before the denoisers. With a denoiser active Arnold forces a Box Filter and works on complete frames, never per bucket. And since adding an imager refreshes the IPR, set them up before the session.
Denoising in Cycles: The Checkbox and the Passes
Where It Lives
In the render properties, inside the sampling section, with two independent checkboxes: Denoise (Viewport) and Denoise (Render). The viewport one shows the clean result without relaunching the final render, and Start Sample sits right below. If you are still choosing between Blender's engines, the comparison is in Eevee versus Cycles.
Which Denoiser to Pick
Automatic uses GPU-accelerated denoising when supported and prefers OpenImageDenoise over OptiX. OpenImageDenoise is Intel's AI model and usually delivers the best quality. OptiX is NVIDIA's: it requires an NVIDIA GPU ticked under Cycles Compute Device, and its advantage is supporting older cards where OIDN no longer runs.
Passes, Prefilter and Quality
Passes offers None, Albedo, and Albedo and Normal: use at least Albedo, because with None the denoiser blurs fine detail on noisy input. With OIDN two extra controls appear: Prefilter (None, Fast or Accurate: None is the quickest and assumes noise-free passes) and Quality (High, Balanced or Fast). Useful trick: the viewport runs perfectly on Fast.
The Compositor Route
Enable the Denoising Data pass and use the compositor's Denoise Node, which uses OpenImageDenoise: it takes the noisy image plus optional Denoising Albedo and Denoising Normal inputs, with an HDR checkbox to preserve colours outside the 0 to 1 range. You gain control over where the image is cleaned, but it performs worse if you apply other effects first: clean first, composite after. AOVs and render passes have their own walkthrough in compositing with render passes and AOVs.
Animations: Where the Flicker Comes From
OIDN denoises one frame at a time: it is spatial filtering, with no memory of neighbouring frames, so the result can shift between frames and flicker on playback. Intel announced in early 2026 that OIDN 3 will add temporal denoising built on motion vectors and depth, planned for the second half of 2026 (CGChannel); today the official series is still 2.x, with v2.5.1 as the latest release.
Noice in Arnold does consider multiple frames, which makes it the better starting point for animation work in Maya. In Blender, if you denoise per frame, render 8 or 10 frames and watch them in motion, because flicker only shows up on playback. Then it needs assembling: Blender's video sequence editor.
Recommended Workflow, Step by Step
In Maya with Arnold: check which denoiser imager is active; if you go the OptiX route, confirm NVIDIA GPU, driver and free memory; compare before and after on a low-sample test render; for sequences, enable Output Denoising AOVs and Merge AOVs and render to EXR; run Noice and keep the raw EXR.
In Blender with Cycles: turn on Denoise for viewport and render; pick Automatic or OpenImageDenoise; leave Passes on Albedo and Normal; test at 50 percent resolution; and for animation, check 8 or 10 consecutive frames before launching the sequence.
The golden rule is identical in both engines: a denoiser buys you time, not infinite quality. Which GPU performs where is covered in GPUs for rendering in 2026, and when a sequence no longer fits on your machine, cloud render farms are the natural next step.
Conclusion
A denoiser does not replace samples: it cleans the image that was already rendered. In Maya it lives under Imagers rather than Render Settings, which explains why OptiX looks missing when the real issue is hardware. In Blender it is a checkbox plus the right passes, and for animation temporal denoising is what matters: Noice handles it well in Arnold while OIDN does not yet.
To see it in action, the platform has three curated video tutorials: Blender Guru's denoiser showdown, Ethan Davis's ultimate guide to denoising in Blender and a walkthrough on how to speed up Arnold renders and remove fireflies in low light. They need no signup: the point of this space is learning without distractions.