lines, glass and denoising

[lacilaci]

Lines/strips from sampling on glass:

The attachment lines.jpg is no longer available

lines+denoiser:

obviously not happening on cpu:

Now, if I remember correctly those lines have something to do with how gpu samples stuff and they disappear after a while...
They disappear very fast on every object/shader except glass reflections though... They can stay around for quite a while and denosier cannot handle them since it is pretty obvious pattern.

Now in a shot like this, with heavy presence of dark tinted glass this hurts performance a LOT.

So, is there something that I can do? Or can this be "fixed"?

[Dade]

Now, if I remember correctly those lines have something to do with how gpu samples stuff and they disappear after a while...

Yes it is due to Sobol "quasi-random" nature and how the work is partitioned in OpenCL (i.e. in strips of 16 pixels). The patterns should disappear like normal noise.

So, is there something that I can do? Or can this be "fixed"?

The most simple workaround is just to switch RANDOM sampler, it shouldn't suffer of the same problem because it is "true" random nature.

Do you have a scene where the pattern is still visible on high samples count ? I understand the denoiser related problem but the only scenario I can think too, where it is worth solving, is a fast rendering animation where you leverage the denoiser to use a low samples/pixel count.

It should never be a problem on final renders of still images.

[lacilaci]

It's mostly problem with previews, and random sampler actually solves that, thanks.

But with this scene it's hard to judge what's high sample counts, pattern is visible even at 500 samples but the glass is still noisy overall even with random sampler. So I guess I'm gonna need like 2000-3000 samples cause this glass is hard to render in general.

[FarbigeWelt]

So I guess I'm gonna need like 2000-3000 samples cause this glass is hard to render in general.

I think you guessed quite right. I estimate at least 1500 openCL samples to get a just acceptable image and rather 2500 to 3000 samples for a satisfying result.