Big Denoise test - Intel Open Image Denoise vs others!

Rendered the animation with all three node-setups for comparison.

Still hard to see, but the 1st or 3th option the video shows is the best.( So in your post that is the 2nd or 3th).


Great, thanks a lot for testing the different node setups, much appreciated!


720p is not enough to make a judgement.

1 Like

Metin_Seven are you able to re-post the bottom picture with the entire green node(Fram…) showing and the nodes to the right of it? I think I like it the best. Thanks

1 Like

Hi, I believe we just had a chat on Blender Market. Do you have your render somewhere here?
If Metin_Seven has EV EXpress he can show the screenshot with the Frame Node. Because that is one of the Node groups that come with that addon. But it is not related to denoising, so I cropped it. But here it is:
So for other readers; Don’t mind the Frame, ASC-DCL and Post node.
But I have no idea how I got this. Maybe from blender developers page or so. And I don’t know if it’s good or not.


Hi @Peetie,

I’m wondering if there’s a particular reason why you’ve disabled the HDR option in your Denoise node setup?

I’ve noticed a difference in the Denoise output, even if the image is not a HDRI.


I thought I am not using and HDRI but lights only, so didnt check that. But wasnt sure if it was related to an HDRI or to High Dynamic Range in the scene.

1 Like

Thanks. Here’s an image I’ve made for anyone wanting to quickly make use of OIDN in Blender:


Hi Metin_Seven,

the normal OIDN (Denoise-Node) setup destroys fine structures.

Look here (ODIN Denoiser - 23sec - 5GB Ram)

Thanks, I know, but I prefer a quick ‘n’ easy OIDN setup I can remember, not a complex node setup I need to append from an external file each time I want to render. :slightly_smiling_face:

I am rendering some video’s out with the latest node-setup shown, 50 samples. And I want to compare that with the OIDN denoise 5GB ram. What node-setup is that, can we download it somewhere? I checked the post 19, but I don’t see a download there.

This should be the one.. Good luck and please keep us posted.

Currently RAM usage depends on the size/resolution of the output image, and the amount of Denoise nodes used in the compositor. The latter should not be that way, it is a reported problem that devs already know.
I think that what is labeled there as 5GB is one of the simple nodes that we already know for OIDN. The other (15GB) is LordOdin node setup which has already been shared in this thread as a node group by NewVisitor (the one whose link has just been shared by Metin_Seven)


An option/setting for tiled denoising would also make sense to control the memory usage as needed.

Is it documented or some other software making use of this?

It is really difficult for me to imagine that a technology that is basically guessing results, could give consistent results by handling different tiles separately, even with overlapping areas.

I don’t know whether other software is using it, maybe E-Cycles does.

I am using the tiling approach for my denoiser, such that everything can be executed on the GPU. That’s why I know it works. If your intuition is that neural network based denoisers are basically guessing, I can totally see why you are skeptical. I am not sure what makes you believe that though.

Someone on YouTube apparently got some spectacular results using OIDN, and he did it by simply using the generic normal pass for the normal data and DiffCol for albedo.

In some testing, it does appear that not using the passes hidden behind ‘denoising data’ solves a couple of issues with smaller details (in spite of not showing reflected or refracted data).

1 Like

It is reasonable to assume that it was also trained with noisy normal and color passes. That’s why it is not surprising.

As I’ve mentioned before, I don’t get the best results with those node setup in most of my tests. Even you can see it for yourself in BMW27 scene that I have shared here:

Also result is usually more blurred than with denoising passes.

Manually processing on separate tiles for example with render border would be an acceptable test?