New adaptive method and optimisation prepass early results

[ben_hamburg]

Actually, as exciting as all of that new stuff already sounds, the idea about the depth based sampling would immensly help my day to day workflow even more...so I second that!

[vlado]

Actually, as exciting as all of that new stuff already sounds, the idea about the depth based sampling would immensly help my day to day workflow even more...so I second that!

In what ways would it help?

Best regards,
Vlado

[joconnell]

Tbh map based sampling would be handy if you know for a fact that you're going to be blurring areas afterward which would smooth out any grain but for non dof renders I'd say it could be fatal for animation - giving far away things less importance, if you're taking depth based as meaning that, could cause a tonne of issues - small objects moving slowly off in the distance are going to end up being the really sparkly things in your renders!

Ultimately everything gets sampled in 2d "screen" space by the anti-aliaser so not sure if depth based sampling would be an advantage, especially not for anything that won't use dof or heavy post production after!

[^Lele^]

Well, I have yet to see a noisy render defocused in the backdrop not look like a blurred, noisy render, but i can always blame the compers i worked with, for that.
And as John mentioned, it may even work for a set of very specific scenarios, but even then, exclusively for stills: it would not work with animation at all. (or companies like Innobright wouldn't invest so much money in denoising algorithms, if a blur sufficed.)
Yet, i'd love to see a mockup of the idea and approach, down to the Frischluft result (you can render a clean foreground, a noisy background, and comp them together to show the result).
I may be able to come up with some synthetic test myself, but i'd prefer the onus of demonstration stayed with the one that suggested it, for a chief reason: i may well be going off topic, and test the wrong scenario.

Perhaps this hasn't been very apparent as of yet, but the new advancements being worked on aren't meant to "undersample" a thing, rather to balance the workload TOWARDS A CLEAN RENDER.
Stopping early to a noisy result because of some arbitrary fact happening after the render is done, as far as my render senses, and sense of lexicon is concerned, is by definition on the arbiter to figure out, not the tool to provide for so specifically.
Besides, in this way you'de be locked into a very specific CoC in post, which would defeat entirely the purpose of going post for DoF: you wouldn't be able to "show" the noisy parts in focus, so to speak, as they came out noisy.

In other words, what's stopping you today from using your tools creatively and getting the job done in passes?
What is it that's currently missing from your toolset that VRay could address?
Are you aware that with a tiny bit of scripting you could already modify the object's subdivs to be correlated to whatever else in the scene (reference object, or camera, for instance, and then multiply the local subdivs by the normalised distance from it.)? Coarse and prone to errors as it may be, it would surely be enough to at the very least set up a test such a scenario, and get a better idea of how it behaves, to then share it so others can debate it with data in hand, rather than just hypoteticals.

As for the linked thread you started, all that's taken care of by the prepass (ie. the single blade of grass you mentioned) when rendering with progressive mode.
But it still won't choose for you what's your subject and where you want the stuff to be cleaner sooner.
That's the mouse-follow task, on timed renders (read: the aforementioned human arbiter).

Thankfully, i'm not the one having to call this types of shot (or likely, VRay today would only render planes and spheres.), yet my doubts remain, and i wanted them on record (so you'll be able to call me names later.^^).

[jenya.andersson]

This is a very cool idea, is there a video of this in action, compared to what we using now?

[Badbullet]

I may be able to come up with some synthetic test myself, but i'd prefer the onus of demonstration stayed with the one that suggested it, for a chief reason: i may well be going off topic, and test the wrong scenario.

I would be happy to do some tests for a proof of concept. One issue, I'm leaving for my honeymoon tomorrow, and won't return for over two weeks. I'll be playing catch up at work after that while moving into a new house. I'll try to sneak in some tests in that period, but it will probably be in a months time. Just so others don't think I'd be abandoning my want of a noise threshold map.

[Alex_M]

When can we see some more demos of the new sampling method? I'm eager to see comparisons before-after on real world examples - interiors, exteriors, studio renders etc.