LuxCoreRender Forums

This might be normal behavior but if there is an object currently in a far end of shadow of an object, that area won't get denoised. Not sure if noise samples are too far inbetween or there is a correlation but denoiser won't do anything with those areas unless they are very low in noise.

In my test example I needed over 2500 samples for the area to get properly denoised and it's still kinda not perfect.

10samples 10samples denoised 50samples 50samples denoised 500samples 500samples denoised scene: Is this normal? Or maybe oidn related issue?

That is quite strange, are you using "tiled denoising" ? It may be a bug introduced with that patch, I will check.

nope, this is normal denoising. I haven't even seen tiled denoising in blender actually

I can reproduce the problem but I can not see any error on our side.

I will upgrade the binaries to the latest Oidn version than, if the problem is still there, I will report behavior to Intel.

Dade wrote: ↑Thu May 30, 2019 7:37 am I can reproduce the problem but I can not see any error on our side.

I will upgrade the binaries to the latest Oidn version than, if the problem is still there, I will report behavior to Intel.

Good, by the way, this can also happen in a more complex environment.. It always kinda seems related to very soft shadow. I assumed that maybe there is some pattern that oidn recognizes in that part.

lacilaci wrote: ↑Wed May 29, 2019 8:57 am denoiser won't do anything with those areas unless they are very low in noise.

I've got similar behaviour in this scene download/file.php?id=4471.
It's Glossy Translucent mixed with Null, with homogeneous multiscattering volume.

I did expereince some issues as well (though not this one) when there were completely black background areas in my image.
Have you tried computing albedo and average shading normal AOVs as well? I had the feeling that helped the algorithm.

CodeHD wrote: ↑Thu May 30, 2019 2:20 pm Have you tried computing albedo and average shading normal AOVs as well? I had the feeling that helped the algorithm.

They are always enabled in BlendLuxCore if the denoiser is used.

B.Y.O.B. wrote: ↑Thu May 30, 2019 2:28 pm They are always enabled in BlendLuxCore if the denoiser is used.

Hmm ok, it seems I was just wrong with that feeling.

I just tried that scene I had in mind again (indeed no difference when including albedo + nromal aov), and it seems to depend also on where I focus in that case, with the result failing completely when my object is defocused (which unfortunately is the scientifically correct case for that instrument ):

focused case (works):
defocused case (doesn't work):

I think it's OIDN 'failing' to achieve human anticipations... did observed similar issues on occasions & also with NVidia's denoiser... more so, all over noticing critiques about this tech. Human adapted, tolerance raised...

And a reminder, during NVidias first AI filtering/denoising demo it was noted that this tools are meant for better, faster, more efficient previewing!!! not final HQ renders.