OIDN Bad quality result and analysis of the problem to see if we can get it fixed :)

[juangea]

Mmm I tried using the Cycles Denoise Normal and the Cycles Denoise Albedo and the result is the same one, no improvement.

I think the improvement in quality may be related to when the denoise is applied, being it applied to different light components may be the trick for the result.

In fact I think I just understood the thing, since we multiply the color pass over the light components, but the color pass is not denoised, because there is no need for that, a big part of the detail is regained there, check these pictures, this is the Diffuse Direct Pass + Diffuse Indirect Pass, notice the blurred zones:

And now this is the Diffuse Color Pass Alone (Keep in mind that this pass does not need denoise):

And this is the result of multiplying them (Dif Dir + Dif Indir) * Dif Color:

You can see how A LOT of detail is there BECAUSE of the Dif Color pass, not because of the light components.

So I think there relies the "trick" of all that, I think if we can recreate this in Lux, then we can get the same results.

[Dade]

You can see how A LOT of detail is there BECAUSE of the Dif Color pass, not because of the light components.

I think this how denoising only the (estimation) of the incoming light and than applying the BRDF (i.e. compute the material reflected light). This is based on the idea that Montecarlo integration has noise in light estimation while material evaluation is noise free (not entirely true for glossy materials but it is mostly true in general).

This is somewhat like rendering in "black and white", denoise and than compose with the Albedo pass. The idea is to denoise (i.e. blur) only the lighting, not the material color details.

It is quite a trick, it doesn't really remove the denoise artifacts, they are just "hidden" (human eye far more sensitive to albedo variations than light blurring).

This may be a viable solution, I'm trying to figure out how it could exactly work.

[juangea]

Awesome, implementing something like this means that the denoise visual quality will improve a lot, even when the "correctness" is not exactly there