Hi Ricky Sharp,
There is no mentioning that the Denoiser takes any “Noise Pattern” based analytics nor that it is a temporal function.
Manual:
This is where Denoiser comes in: this function attempts to reduce the noise via blurring.
Which excludes the idea of luminance and chrominance differentiated functions.
The manual states already a lot of hints how it works, for example, that the Denoiser uses more information than a standard 2D denoiser would:
Manual
Contrary to simple 2D filters, Denoiser can also use other criteria that can help it recognize edges. Multi-passes such as Albedo and normal shading, for example, can be evaluated, i.e., 3D information that is not necessarily visible in the final image.
This would be the first indicator that post denoising with this is not as effective as it could be based on the full data set of the scene.
The comparison with the standard and physical render clear another point:
Manual
Denoiser also works with the Standard and Physical renderers (albeit after the render process has been completed). Here, there are also functions that remove grainy results. For example, think about the Physical renderer with a low sample setting or QMC GI, or Area shadows, etc. Here, Denoiser can be of a lot of help as well. Please be aware of the fact that the results will not be as good as with ProRender. This is due to the fact that the Intel Open Image Denoiser expects a path tracer-typical noise, which older renderers cannot provide. Don’t let this scare you - very good results can be achieved here as well!
Given these quotes from the manual, it becomes clear that the quality of the denoising has dependencies. It is preferable during render in the ProRender, while it is a good idea to save the Raw image.
My personal view, since it requires an Anti Alias
Manual
(Catmull 0.5 and Box 0.5 for Standard and Physical renderers, respectively)
Any post denoising might lower the quality in the process. Which is, of course, a judgment call, based on that each image is different.
Is a complete noise-free image always wanted? Since we have moved more and more away from 8bit/channel renderings in the past 25 years, the dithering is less given for 10bit/c or 16bit/c channel material, of occurs. Perhaps the problem banding (after denoising) is something to consider, as some noise often helps here more than it causes trouble. Again, this needs to be evaluated carefully.
I think the best way for denoising is, to have a few tools available, and check which one works best in each case, some are great but slower, etc.
The differences, as you mentioned, are temporal, pattern analytic, luminance, and chrominance, as well as in size (i.e., noise v. grain).
In a nutshell, there is no general rule from my point of view, and to have the denoiser inside of Cinema 4d is undoubtedly a win.
Cheers