Denoiser slowing down rendering? Tweet

[Andreas_Resch]

I did a few render tests using the PMC kernel to check the differences. Render speed is slower as expected. However, when the Denoiser is checked (Camera Imager), the render speed really goes down. Without Denoiser ON I'm at about 6Ms/sec. With the Denoiser ON, it goes down to about 1.5Ms/sec, which is significant. There's an older thread discussing the issue, but it's suspected, that GPU memory might be the issue. But from what I can see, that's no problem in my case. Any information why that could happen?

[J.C]

It's not memory issue. Unfortunately this is how it works and you have to live with this until new improved denoiser is available sometime with version 2023.2 or newer.
My suggestion is to disable denoiser when workin on a scene and enable it on during render time if you need it.

[J.C]

You can also use Blender's Intel open denoiser using compositor which does not slow down rendering.

[grimm]

Not a solution, just an observation. The PMC kernel works differently than the other kernels. You get fewer samples per second usually, but those samples do more work, so the convergence happens about the same time. The denoiser might be taking resources from the PMC kernel which doesn't effect the other kernels?

Jason

[J.C]

When using PMC kernel the denoiser won't work. It's better to use Photon Tracing kernel which gives you similar results faster and can be denoised.
Generally enabling denoiser in any enabled kernels will slow down rendering maximum up two 2 times.

[Andreas_Resch]

Hmm. But if the Denoiser won't work with PMC, why is the render throughput so much lower when it's enabled?

[J.C]

No idea. Could be a bug or something.

[Andreas_Resch]

I did a few more tests. And indeed, PMC doesn't seem to work with Denoiser. I wasn't aware of that. So I did some comparisons with Denoiser ON and OFF for both Path Tracing and PMC.

The times are taken after all the samples are calulated BEFORE the actual denoiser kicks in.

Path Tracing
DN OFF ... 48sec
DN ON ... 60sec

PMC
DN OFF ... 1m 30sec
DN ON ... 6m 36sec (more than 4x longer)

So the Denoiser seems to add time while calculating the samples. From the tests I did with Path Tracing, it might be around 20%. However with PMC, the times are significantly longer even if the Denoiser will not be used eventually. So that must be a bug.

One more thing that I've seen. Even when all the Denoiser checkboxes are OFF but the Denoiser Beauty pass is checked, the "denoising..." message appears in the status line (although the this pass will not be saved out). Maybe something to take care of as well.

@JC ... I've also tested the Blender denoisers (including the SuperImageDenoiser addon) on the PMC results and they aren't great. I might get better results with some denoising in Photoshop. But it was worth a try.

[frankmci]

I think this is a known problem they are working on. Search the last few weeks of forum posts an you should find some discussion of it.