Denoiser testers wanted

[Raph4]

My test scene.
OpenCL : GPU + CPU


n_threshold_005_13m.jpg
noise threshold: 0,005
render time: 13m
denoiser: off



n_threshold_001_1h1m57.jpg
noise threshold: 0,001
render time: 1h1m54s
denoiser: off

* great but crazy render time


n_threshold_005_denoise_18m10.jpg
noise threshold: 0,005
render time: 18m10s
denoiser: on (default setting)

[super gnu]

good to see there is virtually no difference between 0.001 and the denoised 0.005..

-apart from, strangely the reflections on the black metal in the foreground.. sem to be removed by the denoiser!?

[vlado]

-apart from, strangely the reflections on the black metal in the foreground.. sem to be removed by the denoiser!?

Yes, this is odd, never seen it before in my tests... but other than that, looks good.

Best regards,
Vlado

[vlado]

hmm there are two nightlies there, one 3.3005, but with a higher nightly number ( 26776) and the other, created later 3.3501 nightly number 26775. which is the one i want?

The 3.35/26775 one.

Best regards,
Vlado

[vlado]

But maby you can cut the volumetrics from beeing denoised so we can denoise the environment and the shadows the volumetrics throw.

Volumetrics seemed to work in my tests (I was using just slight environment fog though); if you have a scene where they don't, it would be best to get it to me at vlado@chaosgroup.com so that I can check it out and maybe improve it.

Best regards,
Vlado

[super gnu]

i recon volumetrics is one of the things that the denoiser could be -great- for! getting a clean volume fog is quite slow, and any softening caused by the denoiser could be forgiven

[vlado]

i recon volumetrics is one of the things that the denoiser could be -great- for! getting a clean volume fog is quite slow, and any softening caused by the denoiser could be forgiven

Yes, especially in the multi-pass version; I just need to make a few adjustments there (like ignore world position and normals; they make no sense for the atmospheric pass).

Best regards,
Vlado

[Raph4]

another test, CPU only.

ikea_scene


ikea_scene_denoise (default setting)



Denoise finishes nicely the scene.
I think better is using custom preset for "proper" render without other little issue.

Denoise really save the time !

[super gnu]

ha.. i must admit on my monitor i cannot see any difference between the denoised and non-denoised versions. non -denoised looks just fine to me.

[Moriah]

Here's some results from a project i'm finishing. Can't show the entire thing but here's a crop. First 2 images were made using default preset, at 0.01 noise, you can see it eats the texture detail by a lot. The other 2 are the final render, same noise level, but Low preset on denoise. Seems better, but still eats bit of the texture.

So far so good, but with noisy textures, for animation, it eats some level of detail all the time... Maybe it could be improved by somehow looking at the diffuse? Anyway i'm pretty happy with the results. Still have to test with fine hair.

[vlado]

Maybe it could be improved by somehow looking at the diffuse?

It is already looking at the diffuse element, but I can probably boost this a little bit for the release. You can also try to denoise by elements - this preserves texture detail a lot better.

Best regards,
Vlado

[super gnu]

It is already looking at the diffuse element, but I can probably boost this a little bit for the release. You can also try to denoise by elements - this preserves texture detail a lot better.

Best regards,
Vlado

does this "denoise by elements" provide a single denoised output as with the default setting? i didnt touch it as i assumed that was for people who wanted to rebuild the image from denoised passes.

if so, why is it not the default? is it much slower?

[Moriah]

It is already looking at the diffuse element, but I can probably boost this a little bit for the release. You can also try to denoise by elements - this preserves texture detail a lot better.

Best regards,
Vlado

Maybe a checkbox for texture-heavy renders that can provide a more significant boost you're talking about? Not sure what are the cons of boosting it. That way people could choose according to their needs.
Denoising by elements doesn't work for our case because the compositors are working from the RGB Color, not reconstructing it.

[Ihno]

Here we go:
Firstly I must say I previously used a way to high ColThresh of 0.08.
After Reading through this thread agan, I did this tests with a ColThresh of 0.025
So my results are way better than before.


With the volumetric mode I previously had a clearly visible line in the smoke where the ground plane ends in the backgound.
That is probably related to the nomal and position elements, which are not recognizing the smoke.
Can't see it in this test, but I think it would be probably visible in Animation.
That issue will probably disappeare if you set it to ignore the world position and world normal elements, so this is almost Fixed.

All in all i'm quite happy now with the results of the denoised Vol Mode.
It eats details in the color variation of the smoke and it would be great if you could reduce that.
But that is maby something we would have to accept


In the GeoVol Mode the horizon line problem dissapeared because the smoke is recognized in these passes.
But it is mutch slower than the 0.005 render and it has more Problems with cleaning up areas where thin smoke covers shadows.
Whitch can be seen on Bottom of the right cloud.


Here is the image without denoise and a ColThresh of 0.005 it took 3:12 Min


This images are rendered in the vol mode of the simulators and took 3:01 Min
I didn't re render it to get the Rendertime without the Denoiser.



And this images are rendered in the GeoVol mode of the simulators and took 5:46 Min





In my chase using the Denoiser with scenes like that makes no sense.
As you can see it takes almost the same time to do a 0.005 render compared to a 0.025 with denoiser in the vol mode.
And the 0.005 has way more details in the clouds.
But this is a very minimalistic scene and it probably would be a huge timesaver if I'd put this clouds in a complex scene.

Edit: Vlado let me know if you still want the scene for your tests.

[vlado]

Yes, I do want the scene. But why are you rendering in geometric mode at all?

Best regards,
Vlado