Wish: denoiser adpated sampling

[Micha]

Hi,

if I zoom in a rendering which is denoised by the V-Ray denoiser than I see noise around geometric details. I understand that the denoiser try to keep edges and other details. But it doesn't look good, most if a quick render is done at raw noise threshold or a very high res image is needed and render times should kept low. So my idea, could the sampling not be focused to this critical area? Wouldn't be a hugh time saver only to add more samples to this critical areas which are a fraction of the image?

I know, this isn't a GPU related feature, but I don't know where to ask for core features independent from the finally used 3D software.

Best-
Micha

[^Lele^]

Explore the options of the offline denoiser, you may find what you seek in terms of performance management.

Geo is not going to be denoised across discontinuities, as that'd be a mistake (i.e. artifact).
You may find better ways to get a more consistent denoising by omitting some added render element.
This said, verify the normals for that geo, as it may be those are bad (f.e. with coincident, z-fighting faces) and throwing off the algo.

[Micha]

Geo is not going to be denoised across discontinuities, as that'd be a mistake (i.e. artifact).

I think this is the point. The scene objects are ok and I don't used so much render elements, only the mat id color.

It could be good if the noise at edges and grooves could be eliminated by an enhanced sampling. Since the plain surfaces are easy denoised and good looking after a short time, the calculation time could be focus to critical details. Like an adaptive sampling, but now with taking the final denoised image in account. OK, the undenoised result could not profit by this way, but if someone knows, the denoised image will be finally used only, than he could enable an option like "adaptive sampling based on denoiser". I expect a hugh speed enhancement on this way - adding more samples where the denoiser "fails".

EDIT: I forget to write the reason why I got the idea. The problem is, if I need to render very large image like for example a 360° panorama, than it works quite good to use the AI based enlarger AI Gigapixel. So, I can render a 6000x3000 pano within a good time during a tight deadline and enlarge it to 12000x6000 within a few seconds. Only the noise fragments cause problems.

[^Lele^]

The noise level RE already instructs the denoiser on where the noise is higher, be it geo edges, or else.
You may just need to render the image with a lower N.T and higher max sampling, that's all, if you're sure the geo is fine and not making a mess of the normals.

[Micha]

Here an example from a very large rendering which needs 50min per 2x2080ti. There are artefacts around geometry details. OK, if I would push the NT than I would get a finer sampling for the whole image, but only an additional sampling against the artefacts would be great.