Noisy edges? Focusing sampling to edges?

[Micha]

Hi,

I like the V-ray denoiser since the edges are kept and not blurred. I ask me couldn't be possible to focus the sampling to this edges? Large areas without details are quite easy to clean per denoiser but edges are not so good looking. My hope is that focusing the sampling to clean up the edges only could help to get better result in a quite short time. So, why shot more samples per finer noise limit to areas where no problem exist? So in my example a noise limit of 0.05 is good enough for areas, but for edges 0,01 could be useful.

(attached example - GPU rendering)

[Karol.Osinski]

I am sorry if this might sound too obvious but do you have an AA filter ON?

[Micha]

No, I don't like the blurry look of the AA filter. But it's caused by the denoiser, which keeps geometric details undenoised. That's ok, but undenoised areas needs more samples than fully denoised flat surfaces.

(It looks like I posted the request in the wrong section, I thought I wrote at the render theory section.)

[seandunderdale]

Cant you use an AA filter that isnt blurry? And lower the AA filter size? Lanczos at 1.5 will give you kwispy results. Personally I find anything rendered over 6k needs gaussian at 2.2ish to get a photographic result or it looks to sharp and cg.

[Micha]

AA is an other story. The V-Ray denoiser clean up areas very nice, but edges are undenoised. Here an example where I used AA lanczos 2. It could be great if the calculation power could be focused to the crisp edges now.

[^Lele^]

I see what you mean.
In essence, the normals' discontinuity at edges will prevent the denoiser from blurring them (as the discontinuity signifies geometric variance), so what you're asking is to invert the process at sampling time, dedicating more effort where said discontinuities are, rather than elsewhere.
The thing is, however, that your two cases are very special, in that they lack the texture detail which could be associated to other type of shading (f.e. archviz interiors.), and Albedo (say, diffuse and specular) is also taken into consideration, and denoised less where contrast is higher, so to speak.
Add all the other effects contributing to denoising choices (f.e. defocus or motion blur amounts), and you'd rapidly get into a case of sampling everything better.

In general, the approach to denoising should be centered about the worst-case scenarios, not the optimal ones.
In other words, ensure those edges are good after rendering, and then you'll be able to denoise the rest effectively, while still saving some time overall.
As you're doing it now, the base image is too noisy and that noise amount still shows through where the denoiser acted -correctly- less.

[Micha]

Right. What I want to say is - I see a possibility to increase the quality by a small amount of render time if the sampling would be adapt to the final denoised result and not the RGB pass only. I could be nice, if the user could choose between "RGB" pass or "denoiser" pass optimized sampling. I know, we will not find a solution now and I would be glad if the suggestion could kept in mind for the future development.
At the moment I need a much stronger noise limit to get the small edges clean like the areas. The areas would be more sampled than needed. Here I see potential.

Thank you,
Micha