Page 4 of 4

Re: Simple Test Scene

Posted: Fri Nov 29, 2019 4:59 pm
by Dade
Dade wrote: Fri Nov 29, 2019 12:25 pm
Magog wrote: Fri Nov 29, 2019 12:21 pm
Dade wrote: Fri Nov 29, 2019 11:12 am @Magog, can you post a test scene with just a single couch pillow (if the major problem with the denoiser is there) ? So I can check if there is some problem with Oidn and where it comes from (or is just a limit of the Intel software, at the end of the day all denoisers have some limit).
https://drive.google.com/open?id=1ovyDS ... jvJ6KeWhVN
Thanks, checking.
It looks like the problem is in the albedo pass of cloth material:

cloth.jpg

See the heavy moire effect of cloth material on the albedo AOV ? It seems to drive crazy Intel Oidn. This could be explained with a lack of filtering on the albedo pass and I may be able to fix it.
In the meantime, you could use a normal image map instead of the cloth material and it should fix the Oidn behavior.

Re: Simple Test Scene

Posted: Fri Nov 29, 2019 10:05 pm
by Magog
Thank you so much for your support.
Tomorrow I will try to do new tests with different maps.

Re: Simple Test Scene

Posted: Sat Nov 30, 2019 10:06 am
by Magog
This morning I tried to replace the textures.
At the first test, done at 500 samples, OIDN denoise and 1920 * 1080 (200%) the image is perfect output!
However, changing resolution (130%) I found myself again the planed textures and without details.
Changing some material for, with the same initial settings I noticed a further loss of details.
Also tested with BDC tests.
Below are the screenshots.

Re: Simple Test Scene

Posted: Sat Nov 30, 2019 2:20 pm
by Dade
Magog wrote: Sat Nov 30, 2019 10:06 am This morning I tried to replace the textures.
At the first test, done at 500 samples, OIDN denoise and 1920 * 1080 (200%) the image is perfect output!
However, changing resolution (130%) I found myself again the planed textures and without details.
Changing some material for, with the same initial settings I noticed a further loss of details.
Also tested with BDC tests.
Below are the screenshots.
I have looked a bit into the problem and, at first, I was thinking that the moire was a Albedo pass specific problem than I noticed it is present in the normal rendering too:

r1.jpg

Indeed, this is quite detrimental when the denoiser is applied:

r1-denoised.jpg

Increasing the pixel filter width from 1 to 3 seems to solve the problem:

r3.jpg

And once the moire effect was gone, the denoiser starts to do a pretty good job:

r3-denoised.jpg
Pixel filter is a setting available in the Blender scene panel. The default value is 1.5 but it was set to 1 in this scene; 3 was working fine for me but your mileage may vary with a different scene.

Your test seems also to suggest that rendering at an higher resolution and than scale down the result of the denoiser may help too (i.e. the denoiser will have more details to work with).