Interior sampling problem with automatic sampling - still noisy

[Jiri.Matys]

Scene sent.

[^Lele^]

Yep, got it.

The black noise level is a great cue, along with sampleRate: your problems stem from the post-processing: you expose up, burn highlights, and then contrast.
That means that what the renderer did, and what you get in the VFB aren't the same thing anymore.
You will simply need a lower NT than usual (and possibly a higher number of max subdivs) to compensate for the post contrast (which in the area you mention lowers blacks and raises whites, so it makes the noise higher than it ought to be.).

Attached a simple contact sheet: the top bit is a crop without any post, the bottom the same, after the post you had active in the VFB.
Regardless of how i get to the numbers (and their poor formatting. the gizmo expects bigger images.), you will notice that the lower one has a higher maximum (V, the last of four numbers, in each line), and a higher average: meaning more noise.



Let me know if this helped.

[Jiri.Matys]

Sorry, I look at those two images, load it to PS to be sure, and top image has more noise then bottom.

[Jiri.Matys]

But I know what you saying. I think I use vfb like anybody else, nothing special. This is what I done: rise exposure little bit (its same like change camera settings so I think no big deal), then drop highlights to limit burn in sun (everyone doing this right?) but contrast is lower after change highlights so to compensate it I turn contrast little bit up. I think didnt do anything drastical like change contrast three times higher or something and VFB is still in 32bit. So I think I can do this but maybe I am wrong.

But if I dont do this post, image look different and not so good to me. So how you would change the scene/lighting/post to get "nice" image like with post but without it Or any other suggestions please if I do something wrong in the scene? I will really appriciate. You can reply on my mail, thank you Lele.

[^Lele^]

Well, let me go with you through it, as i haven't comped a thing for work in ages.

The burn you can forget almost entirely, as it happens on the higher ranges (so nothing to do with the mid-to-lows of the furniture).

Exposure alone, particularly in the amount you used it (0.2 stops if i recall right) isn't a problem either, as while it does stretch the pixels' differences (and so enhances "noise"), you used it fairly sparingly.
Exposing up, however, does enhance the pixels differences, as it multiplies the original values by the same amount, so while they stay identical relative to one another, their absolute difference grows.
For example, have pixels A and B of values 0.25 and 0.75 respectively exposed up one stop.
Their intensity doubles, the ratio always 3:1, but the difference between them also doubled, from 0.5 to 1.0.
The pixels now are twice as far apart.
In your case, a 0.2 stops up means it's like if you multiplied the values by roughly 1.15, or 15%.

For contrast, things are even worse, even when used sparingly, as it's an operation designed to precisely increase (well, or decrease, ofc.) the differences between pixels.
Leaving the maths aside, look at the graph below (from "The Art and Science of Digital Compositing", no less.) showing a 33% contrast operation centered on 0.5:



If this was the case, the two pixels are now 0.0 for A and 1.0 for B, again their difference doubled (with just a 33% contrast! And 25% would have been enough as well, ofc.), with the aggravation of loss of data for the A pixel (at 0.0, it's dead to any other post-recovery of the image.).

So, possible fixes:

When I had to feed a pack of hungry compositors with sequences which would need a stronger than usual post-treatment (say, a lightning flash brightens the Cg, and it's added in post due to the timing being not yet decided.), and *well before* 3.3 was out, the only possible way out was to render the whole image brighter, so to help the sampler do more work.
This hack was needed because the Noise Threshold expressed an absolute color value.

Today, the best possible way to do this is simply to lower the Noise Threshold further: you want to gain the image that looks good as it is by one stop in post? Well, Halve your Noise threshold. 2 Stops? Make it a quarter, and so on.
This is possible because the Noise Threshold is now expression of the allowed percentage of variance between samples.
It goes without saying that this (just like the pre-3.3 hack above) will cost time and money, as V-Ray will simply have to do (a lot) more work before calling it done.
This covers exposure only, however.

Contrast is a very different beast, by nature, and is something you *really* don't want to abuse in post.
If a Compositor came to me complaining of low contrast in a sequence, I was always grateful (partly because they scared me.), as if the correction was made in post, by the time all was said and done that shot may have looked too noisy (some amount of unavoidable contrast will happen, with LUTs and effects piled up, and one doesn't want to add misery to it all.)
So the solution was to increase the contrast in the scene, directly.

Particularly for interior shots, and with 100 bounces like with the LC, it's *vital* to nail surfaces' albedo, along with light intensities.
As you're using an unmultiplied Sun, we can assume it's correct (if not, it's a bug. ^^).
In your specific case, if i had to nitpick, i'd say the wall shader is at least 50% brighter than it'd have any right to be, bouncing heaps more light around, without it having much chance to extinguish, thereby lowering overall contrast.
Likewise, the wood table has a very bright diffuse, and possibly other shaders (i miss a couple of plugins, like the tileGenerator, so i can't quite check.).

In general, lower shaders albedo, raise camera exposure to compensate, and you have obtained more contrast without post.

For your specific scene, you've got the glass pane still on the window, right in front of the portal, which surely isn't helping (it's a shadow-casting object.), albeit it being "phisically accurate" (it isn't, as no glass is a single atomic sheet.).
Further, if you light interiors with a sun and sky, and a high sun at that, you'd better raise the max ray intensity (or remove it altogether), as what you may save in rendertime, you lose *bigtime* in light entering your environment.

If you need to recover this particular scene, my suggestion is to start lighting with a gray 64 (or a lower 18%, if you feel fancy. The goal is to have a 0.5f gray as normally-lit colour after the display inverse gamma is applied, ie. at viewing time, so the real value would be 0.5^2.2, or 0.217f, 55 in 8 bit.), without "max ray int." active at all, and initially under a 6500k WB (your colour cast will be quite different without clamping).

Once you're happy with that, start with the bigger patches' shaders (ie. walls, ceiling, floor), and rebuild them with darker diffuse and reflection values, aiming for the sum of both components to never exceed 0.75 (broaaaaad rule of thumb.).
You will see you won't need such odd IoRs for reflections, not super-bright reflection values: the lights will provide all the range you'll want and more.
Raise or lower exposure as you like in camera, until you're happy with how white the walls look, and the rest will follow.

Notice that it's conceptually wrong to raise the specularity of shaders at rendertime, to then cull the resulting intensities with a (soft) clamp in post (see the Burn): you introduce enormous energies into the render phase, which will require time to extinguish, and then for good measure remove the hard work V-Ray had to do by clamping it right after it's done.
Much better, under a more controlled setup, would be to get the specular values right for the shader before the render started, so to avoid excessive energies (and consequent rendertimes) and the horrible Burn parameter, which has always the net result of *removing* contrast from the image.

[Jiri.Matys]

Thank a lot Lele. Something I know something I dont, something I dont understand

In a point of 3D and shaders, I put bright color to the walls for the first time for one reason - to scene be brighter. When I didnt do that, whole interior was darker, but areas where is sun those areas was still bright so I have to compensate burn more then now, but scene was darker then now. And I really like bright whole interior, but not washed out.

When I will use LUT or do post everywhere else, it will be the same, so I done it in VFB, its faster.

So when I lower albedo, I will have more contrast, ok, but I still have to compensate burn value, right?

Is there any chance you can change the scene, only in basics - lighting/camera/wall shader, no the rest.

sorry, i dont understand this: "If you need to recover this particular scene, my suggestion is to start lighting with a gray 64 (or a lower 18%, if you feel fancy. The goal is to have a 0.5f gray as normally-lit colour after the display inverse gamma is applied, ie. at viewing time, so the real value would be 0.5^2.2, or 0.217f, 55 in 8 bit.), without "max ray int." active at all, and initially under a 6500k WB (your colour cast will be quite different without clamping)."

Thank you once again.

[Jiri.Matys]

I send another mail. But if you dont have time or something, please let me know.