Blender 2.81 Alpha: Denoiser Comparison
Prof. Michael Klein writes:
I made a test series with an existing LEGO scene in Blender 2.81 to compare the Cycles Denoiser with the Intel Open Image Denoiser. I started with 1 sample and doubled it 9 times up to 512.
Please feel free to download the picture wth 5 x 10 HD pictures with a resolution of 9.960 x 11.160. The preview picture in this posting is downscaled. I saved it as JPG in Photoshop with maximum quality. The file is 48 MB.
Here’s a short summary:
The Cycles Denoiser looks better when it comes to higher sample rates. It keeps important subtle texture details (eg. bump maps) but even at 512 samples there are still quality issues in the glass parts.
The Open Image Denoiser is not able to keep those texture details but looks better at lower sample rates. At higher samples rates the picture seems still blurred at the cost of missing texture details.
Interesting is that the Denoise Albedo layer will add artifacts/noise to some parts of the picture and makes it looking worse. There’s not much a difference between a denoising without and with the Denoise Normal pass in a still picture. Maybe this will be different for animations? Also the head of the Minifigure looks not as good as the Cycles Denoiser result.
It would be interesting to see a) how must the render time is decrease with the RTX / Optix Cycles patch and b) add a another column that shows the OPTIX denoiser output.
This is a great comparison and very helpful. Thanks for taking the time to put it together!
You’re welcome.
From my perspective (informed but not professional) I can’t see a significant quality different at high sample rates and at lower sample rates there is no contest that the non-Cycles denoiser is far more effective.
I do agree with you. I made tests and the details are beginning to be blurred from 32 downwards.
From 8 to 1, the time begins again to expand, maybe because there is more load on the OptiX.
Well, the least I can say is that it was not useful for me, because there were many details in my scene.
I guess that for other scenes with less significant details, it’s worth it. I’ll continue with the traditional denoiser.
There are quality differences in the glass material and it looks like the Cycles Denoise can’t fix it. The Open Image Denoiser is doing a better job here. I haven’t tested another option to use the Denoiser Data passes with the external use of the Cycles Denoiser (this is not implemented yet as a node in the compositor) to see if this makes a difference here.