Best noise reduction approach

[Sharlybg]

Are we feeding OIDN correctly in this scene ?

https://drive.google.com/file/d/1LkSh9u ... sp=sharing

[Sharlybg]

Are we feeding OIDN correctly in this scene ?

https://drive.google.com/file/d/1LkSh9u ... sp=sharing

Please Can someone download and check if this scene have OIDN passes correctly setup ?
Thanks in advance for your help

[Dade]

Please Can someone download and check if this scene have OIDN passes correctly setup ?

Yes, it does however what are you using for the rendering ? If you are using LuxCoreUI, this is running ODIN at each screen refresh (huge useless overhead). It is a lot better to have 2 image pipelines, the first one without OIDN and the second one with. So you can use the first one for screen refresh and run the second one only on demand (for screen refresh or file save).

[daros]

aniway it's wired...despite feeding OIDN with normal+ albedo, which should help to maintain texture detail...It seems impossible to use Oidn for final production. Even if i increase the resolution by 8 times, as suggested by Dade, the final area of the surface is blurred by oidn. Even if i increase rendering time by 8 times the Ai doesn't gets smarter. Oidn is very destructive.

The last picture shows an other bidir rendering engine (10 years old) on which we tried a different approach obtaining decent results. We are running the same rendering twice at half the sampling level but using a different noise seed. Then we pass the two results through a time-based video noise reduction tool (neat-video). The result is an in-between of the noise of each pixel. In our implementation the only noticeable defect is that light distribution is slightly less "evolved" because the lower sampling level. But, at higher sampling levels, it converges correctly to optimal result with no artifacts. By sure it's a silly approach but it helped us to increase rendering speed by roughly 5 times with almost no noticeable artifacts.

All tests are 3 minutes long and done on 4 Xeon Gold 6138 Cpu's


Luxcore+Oidn

Luxcore no Oidn

Other bidir engine with time based noise reduction.

Anyway i want to remark the fact that a speed up in bidir is truly needed. Please help!

[Continuum]

Just a guess, not enough pixels in the image for OIDN to distinguish features properly? Here is a 4K render (6min path tracing Blender).

[daros]

Yes probably...however a noise reduction algorithm is supposed to deal with the signal/noise problem.
If we have to render images at such high resultion in order to make the algorytm understand what is signal and what is noise...it means the algorhytm is cheap or at least not suitable for final production. Rendering a picture at 16 X the desired resolution means 16 X rendering time. I prefere to render at original resolution 12 times longer.

[Dade]

Have you disabled Pixel filtering in LuxCore ? Intel OIDN is trained without pixel filtering and BlendLuxCore takes care of that kind of settings.

[Sharlybg]

Have you disabled Pixel filtering in LuxCore ? Intel OIDN is trained without pixel filtering and BlendLuxCore takes care of that kind of settings.

Where can we find this parameter in Blender ? And does it make sense on Bidir engine ?

[daros]

Have you disabled Pixel filtering in LuxCore ? Intel OIDN is trained without pixel filtering and BlendLuxCore takes care of that kind of settings.

Yes, we have disabled the pixel filters as always wen using Oidn

[Sharlybg]

Have you disabled Pixel filtering in LuxCore ? Intel OIDN is trained without pixel filtering and BlendLuxCore takes care of that kind of settings.

Yes, we have disabled the pixel filters as always wen using Oidn

Ahh ok it is this same blackman harris filter stuff. i didn't see any improvement with/without. will test again and see.