Denoiser Battle Royale: Testing Top Denoisers for Cinema 4D Renders

May 16, 2018

Can the built-in 3D denoisers in Redshift and Arnold beat out my favorite 2D denoisers when it comes to 3D animation? We test out Optix AI, Noice, Altus for Redshift, Red Giant Denoiser III, and Neat Video.

Before we get started, I have to get one thing out of the way. This little experiment entirely ignores denoising as it relates to still images. This is a motion design site, right? No one can argue the usefulness of denoising still images, but the tall order here is cleaning up moving images.

Denoising images is nothing new, artists and filmmakers have been using plugins for After Effects, Nuke, and Fusion for years. But now that denoising is showing up in Cinema 4D with artificial intelligence behind it, maybe it’s time to take a closer look.

Denoisers being put to the test:

  • NVIDIA Optix AI
    • Free with Arnold
    • Free with Redshift
    • NVIDIA GPU Only
    • Requires no extra AOVs
    • Denoises in the IPR
  • SolidAngle Noice (Autodesk)
    • Free with Arnold (Built-in)
    • Very slow (9 hours for this render)
    • CPU Only
    • Requires specific AOVs
  • Innobright Altus for Redshift
    • $149.90 annually for Redshift
    • AOV setups done automatically
    • Limited Controls
    • Dual pass looks better but takes longer
    • Denoises in the IPR
  • Red Giant Denoiser III
    • $199 for After Effects, Adobe CC
    • Quick to process
    • No OFX
  • Neat Video
    • $74.90 for After Effects, Premiere, Final Cut Pro
    • $99.90 for Avid
    • $249.90 for OFX (Nuke, Fusion, etc)
    • Relatively quick to process
    • Tons of control

Denoising grainy GI renders is a convenient way to push fewer samples into your scene and save time by denoising the images after the fact. There are a ton of players out there that are either built-in to 3D renderers, have standalone apps, or that you can access in your favorite compositing application.

The newer denoisers vary in technique from accessing unique AOVs that the renderer provides (Altus and Noice), using machine learning AI (Nvidia Optix), and the 2D denoisers use temporal and spatial magic to predict and remove noise in your footage.

Algorithms and machine learning all working to denoise that killer render your client is clamoring for. Mathematics YAY!

If you’ve seen any of my recent tutorials, you’ve probably guessed that my primary renderers are Arnold, Redshift, and Octane. At the time of writing this both Arnold and Redshift ship with two built-in denoising options and Octane has one on the way.

As a long time user of denoising plugins in After Effects and Fusion, I set out to see if any of these built-in denoisers could dethrone my go-to 2D denoisers. Would there be a new denoising champion in this battle royale?

Enough of this noise, let’s get to the results.


Denoising Arnold Renders

Up first, I’ll show the results of denoising Arnold renders. Here is a look at a RAW render without any denoisers applied.

Download Arnold Scene Files

Get the Arnold scene files by clicking on the button below.

GET ARNOLD SCENE FILES
 

Now let’s look at the Arnold denoiser results.


Nvidia Optix AI with Arnold:

Nvidia’s Optix AI denoiser is popping up everywhere lately. It seems like they are relatively good about providing access to renderers to their “Ultron-like” system. The problem is, most companies are using a version of Optix that was trained on iRay. I think we will see many renderers training the robots on their own style of grain and we’ll see better results soon.

For now though, Optix should be used strictly for stills. It has a difficult time with motion and often gets blotchy while the machine tries to predict what is noise and what is texture.


Autodesk Noice with Arnold:

Noice certainly has the best name of all these denoisers, so hat’s off Autodesk! Noice is one of those denoisers that require specific AOVs to function and it also needs to be run after your render has completed using a command line function or through the Denoising GUI in C4DtoA. So you won’t see just how far you can push it until you’ve fed it your EXR sequence.

It’s also EXTREMELY slow. This 300fm 1024×1024 sequence took 9 hours to compute in Noice. It also soaked up 99% of my CPU resources turning my machine into simmering brick of uselessness for an entire day. Not noice. However, the results were very noice (ok last one).

It did a great job recognizing texture detail while eliminating GI noise. However, it did struggle with glossy refraction noise where the AOVs could not help.


Red Giant Denoiser III (After Effects) with Arnold:

What I love about Red Giant Denoiser III is its simplicity. Very few knobs to twist and reasonably good results on a wide variety of footage.

Red Giant’s Denoiser III performed well with the Arnold footage cleaning up a majority of the noise with no sweat. I did notice it struggled a bit in the shadows and that seemed to be a recurring issue.


Neat Video Denoiser with Arnold:

Don’t be fooled by the website, this plugin is pretty much magic. I love how the fact you can use it in both AE and as an OFX plugin in Nuke or Fusion.

In this test, Neat Video did a fantastic job smoothing out the noise while maintaining texture detail.

Overall, Neat Video outperformed the competition and remains my go-to denoiser.


Arnold Denoiser Ranking:
  1. Neat Video
  2. Red Giant Denoiser III
  3. Noice
  4. NVIDIA Optix

Denoising Redshift Renders

Before using a denoiser, here is a RAW render straight out of Redshift.

Download Redshift Scene Files

Get the Redshift scene files by clicking on the button below.

GET REDSHIFT SCENE FILES
 

Now let’s dive into the Redshift denoising results.


Nvidia Optix Denoiser with Redshift:

It’s not easy to get a grainy/noisy render in Redshift, so that was a challenge unto itself. You can quite effortlessly get clean results with this renderer, but hey, we had to give the denoisers something to work with.

Optix in Redshift is pretty much the same Optix as everywhere else. Trained on iRay renders in a secret lab and pushed out into the world. Great for stills, not so much for animation. Splotchy, unpredictable results and still quite noisy. Not great.


Altus Denoiser for Redshift:

So Altus has a standalone app that I’m sure has more bells and whistles to get better results and more control, but for this test, I used the current Redshift version which has fewer controls than the stand alone but enough to work with. This denoiser also has the option of using “dual pass” which is sort of like doing twice the samples (according to the folks at Redshift).

I was assured that the single pass would be sufficient, but the dual pass would most likely produce better results. I tried both and saw little difference in my test, so I opted for the faster single pass method.

Given that Altus uses Redshift AOVs to hone its results I expected similar results as Autodesk’s Noice. Boy, I was wrong. Not sure what was happening with Altus, but nearly all of the detail in the statue’s shoulder and leg were completely obliterated.

It did a slightly better job at removing noise than Optix, but it’s inability to hang on to subtle texture detail makes this one a definite pass for me.


Red Giant Denoiser III with Redshift:

I’d say that Red Giant Denoiser performed slightly better in this test than during the Arnold run. I think this might be because the Redshift version of this scene ended up being marginally brighter and it’s the shadows that Denoiser III struggled with in previous tests.

Here Red Giant Denoiser III did well in both denoising and texture preservation. Nicely done Red Giant!


Neat Video Denoiser with Redshift:

Neat Video once again clobbered the built-in denoisers by effortlessly eliminating all the noise AND holding onto all the subtle texture details. All without AI, AOV’s, or CPU burning render times, and Neat Video is much cheaper than Red Giant Denoiser.

That’s another win for Neat Video.


Redshift Denoiser Ranking:
  1. Neat Video
  2. Red Giant Denoiser III
  3. NVIDIA Optix
  4. Altus for Redshift

Summary:

I think it’s safe to say that the built-in denoisers, in spite of their robotic intelligence and use of AOVs, are no match for the tried and true 2D denoisers.

Optix AI has the most ground to make up here followed by Altus for Redshift. I think Noice has potential if they can speed it up, and perhaps get it working in the IPR. Until then, I’ll stick with rendering out my undersampled GI and letting Neat Video repeatedly save the day.

Thanks to Pixel Plow‘s render farm for helping us get these results! Stay tuned for a follow-up with the Octane denoiser upon its release.

Tutorial Focus:  , , ,
Software:  , , ,
Educator:  

18 Comments
  • Hi! Thanks a lot for the tutorial! Only in Greyscalegorilla we can found this information, I have one question what parameters do you set on Neat Video?

  • hallo chad,
    thank you so much for making this comparison. highly appreciate your work
    ….weeks before I watched a presentation of Innobright Altusaltus via/for redshift renderer and I was kind of impressed so to speak but now after your amazing video its absolutly disappointing. yeah its almost crap and unusable (the texture blury, washed out ect.) even with all the parameters at default it should do a better job then this, obviously. neatvideo “2d” denoiser seems the only way to go. wow, just wow!

    • Thanks for checking it out! I’m hoping that the implementation of Altus improves and I’ve already received word that Arnold’s Noice is getting speed improvements. I guess I’ll have to revisit this in a few months 🙂

  • Thanks for great testing but I have a little request if possible; is it possible test Maxwell Renderer Denoiser? thank you

  • Hello Chad, thank you very much for this excellent comparison on denoisers, I am in a VR production with Arnold 4096 X 4096 px renders and I was hesitating which tool to choose for the denoising, I think you have made it very clear.
    I also liked the document you have done in Fusion to make the comparison, more clearly impossible, Thank you.

  • Interesting test, thanks! But what about the denoiser in Resolve? Any particular reason you didn’t consider it?
    Thanks!

    • I stuck to only tools I use in my workflow (After Effects and/or Fusion). I have yet to jump into resolve, but I do plan on it. I have heard good things about the denoiser in there.

  • robertw@walkersound.com May 17, 2018 at 2:44 pm

    Octane 4’s denoiser is pretty amazing. They haven’t built the hook into C4D yet, but you CAN export ORBX from C4D and then use the Octane 4 standalone to render. That is the workflow I used recently for some TV commercials.

  • I’m not sure how you can tell about preserving textural detail in such a simple scene. Hiw do all of these handle denoising a sharp fine detailed texture like a wallpaper pattern, behind a piece of glass that has a noisy shadow on it?

    While Neat Video is certainly still a great fallback for us on large features. Even our use of Neat requires far more work than this article suggests. You seem to run Neat on the beauty render and that is it and say that is the best result? With textures, reflections, and retractions, we still need to subtract the diffuse albedo (texture aov) and often the specular albedo (refection texture aov) if there is some, denoise the results and reapply.

    There is also, no geometry awareness in Neat, another thing this test scene fails to test. Hair, grass, scenes with millions on small geometries like x particle scenes. Neat can’t tell if the fine particles are render grain or intentional fine details.

    We also have to switch to an entirely additive compositing workflow instead of a subtractive workflow, which is far less accurate. So Neat while very good. Is a lot of manual work.

    The tests are great to see how it will handle simple shading and coherence, but in a weird way, the noise it gets rid of is actually the lower priority, while the details it can maintain are the highest. A simple scene like this can has little to lose so agressive results might be appealing.

  • Version 4 – XB2 (Test release1) for C4D is out!! Any tests?

  • Thanks Chad, This is great.
    I bought Neat video purely to de noise live action video, and never really thought about de noising 3D renders. Now my investment has doubled.

  • Im sorry, but ABSOLUTELY disagree on your negative outcome on Altus denoiser. Im using it in dual-pass mode, and its much better than all the others. (I tried them all as well) Your video sample on texture preservation is simply incorrect. I think you may do something wrong, maybe not giving enough samples OR not the same amount of sample to each passes.

    • It could very well be that something went wrong. This was an experimental build after all. Feel free to grab the scene file and give it a go (you can download it on this page). You’ll have to add your own subtle texture to the diffuse. Be sure to share any results here. Thanks!

  • Hi guys,

    have you looked into anything to do with Corona render for C4D.
    It’s got a great denoiser that can be modified along with a lot of other great features.
    The best bit is a really intuitive interface and ease of use compared to other 3rd party render engines.

    Would be interested to see how you guys tackle it.

  • Hi Chad!
    Great article and what a nice comparison!
    I use Neat video for quit a wile now in my 3D work
    and i’m glad to read that it safes the day for you as well!

    I was interviewed by Neatvideo about using their plugin in my 3D work and I’m sure they would love to share this article too!

    https://www.neatvideo.com/user-stories/geert-van-hooff

  • Leave a Reply

    Your email address will not be published. Required fields are marked *

    This site uses Akismet to reduce spam. Learn how your comment data is processed.





    Everyday Material Collection: Get 350+ Materials and Save $100 Before Friday 11/16SAVE $100