Big Denoise test - Intel Open Image Denoise vs others!

In hindsight i probably should not have used Depth of Field in the render. Also note programs like Topaz AI Clear /AI Gigapixel work to reduce blur so negate the DOF effect as well as noise. Also AI Clear is fully adjustable.
If you see differences in color vibrance, its because of the RGB color gamut used in saving. it confuses me to no end and different websites cant handle certain wide color gamuts anyway. The duller ones got saved to regular SRGB i guess. And the more vibrant ones probably managed to get AdobeSRGB.

noisy image 128spp render time 2:26


With Blender internal denoiser default settings. Render time 2:30

Intel Open image denoise using Albedo and Normal data

AI Clear(using just the noisy image as the input)

AI Gigapixel(using the noisy image as the input)

AI Gigapixel + AI Clear

Photoshop reduce noise

Nearly Ground Truth 2000spp render time 28 mins.


2000spp + AI Clear

3 Likes

A simplified test.
Noisy image 64spp Render time about 1 min


Blender Denoiser

ODIN denoiser

3000 spp almost ground truth.(Render time 23 min)

3000spp +ODIN

1 Like

Hi, I think internal wins, OIDN smooth to much.
May you can check node setup from BA user (will add credits later).
It gave more details in my tests.

denoise_nodes.blend (124.0 KB)

You can append collection from your test scene but sun settings are not included.
May it is a better way to import but node group does not work, hm.

Cheers, mib
EDIT: Credits for node setup goes to @LordOdin

So that looks like you’re taking most of the render passes available in Blender and feeding them into ODIN’s and then combining them all at the end??
I recall node groups like this in the old days before blender internal denoiser.

Yes but I never can do it myself, here is the post from LO: Cycles Development Updates

In E-Cycles is an addon included mange the node setup in 3 quality steps, with/without SSS and some more. You have to click and it create the nodes.
I think something like this should be in official Blender, too.

Cheers, mib

1 Like

In your case, the comparison between the Intel denoiser and the built-in denoiser is more or less a wash.

The Intel denoiser loses detail in the yellow part, but the built-in denoiser destroys the highlights in the berries and in a couple of other areas (not to mention it’s splotchy). The built-in one keeps detail in the yellow part, but the plate details are preserved better with the Intel solution.

Well heres what i got with that big node group, i had to add one more color mix node to combine the albedo + normal nodes with all the other passes in order to get an output, i used ‘add’ as the mix type…


Here’s the big node group but with clamp on the last node stops the highlights being blown out i guess.

1 Like

Hi.
OIDN result for still images is impressive, not only in this image but also in my tests.

By the way, this is “OIDN”, not “ODIN”.
Nice tests!

Very interesting, thanks for sharing. AI Clear achieves quite impressive results. Topaz is rocking with their new AI series. I own Topaz Gigapixel AI and Topaz Adjust AI. I keep being amazed by the striking detail Gigapixel adds to image enlargements.

I’m wondering how NVIDIA’s OptiX denoiser holds up in this comparison. Did you try that too, or do you have no NVIDIA GPU?

Yea AI Clear /Gigapixel are really tuned for images that have only slight noise, rather than low sampled render images. I own both and i use on final renders when i have a great image but just want more sharpness i will use gigapixel.
I have not tried OptiX yet simply because with the Topaz software i did not need it, i only usually deal with low noise final render images. Yes im on Nvidia hardware.

1 Like

Could you render with high render samples so that it looks without noise (without using denoisers) to compare better?
If you don’t have powerful hardware, it would be enough to set a render border (Ctrl+B from camera view) to render a portion like this:

on its way in 20 mins. the center portion will be 10,000spp ground truth.

1 Like

By the way, if you’re using the Denoising Normal pass for the detail it gets, you will get a fairly correct looking pass if you multiply the color data by itself using a mix node.

In addition, you can try to eke out more detail by stringing that pass and the denoising albedo pass through the Color Balance node set to the ASC-CDL Mode. This technique is essential if your scene is volumetric in nature.

1 Like

Just had a look at the OIDN Denoise node in the Compositor of yesterday evening’s 2.81 master build, but I’m getting a ‘Disabled’ message:

image

I assume OIDN needs to be installed separately / manually?

Apparently only working in Windows builds at the moment.

But I’ve got Windows these days. Mac is sooo 2018 (at least for me). :wink:

Also note that all those images had a Blackman-Harris filter with the filter width at 2.00 rather than 1.50 so they maybe slightly blurry.

I’ve tried it with 200SPP at a larger resolution and it does not wipe out all the detail on the yellow sponge part. Overall verdict this is a fantastic tool, i can mix in a little of the noisy image with soft light mix mode if i want to recover some detail if its too denoised… So yea it really saves render time if you find the sweet spot.

64 Samples

Internal Denoiser - 16sec - 2GB Ram

ODIN Denoiser - 23sec - 5GB Ram

LordOdin Node-Setup - 55sec - 15GB Ram

1 Like

Hi, can you add more information about samples and may a sample with no denoise and one with more samples without denoise, please?
15 GB, really? Is it possible your system starts swapping on hard disk because of the long render time?
Will do some more tests looking on RAM usage.

Cheers, mib