Progressiv quasi full adaptive rendering based on IM+LC+PQMC

[Micha]

Hi,

yesterday I try to find an easy&fast-all-for-one-setup and ... I found one.
My goal was to find a workflow, that allow to render "every" scene with one setup and without subdiv settings of materials ... and fast must it be. Now call me a fool ... ... and try it.

Befor I explain and bring confusion, here my settings (at vray for rhino).



Standard method adative 1/4 0.85 ....


Quasi full adaptive IM


Like I have shown here, it's possibe to set Vray in progressiv mode per full image bucket size and "full" adaptive rendering. In combination with PQMC the results are:

Quasi full adaptive IM + progressiv final pass stoped after 13min


Quasi full adaptive IM + progressiv final pass stoped after 10min


Yesterday, after I tested the basic IM options, I wrote this at the VfR forum:

Conclusion:
The goal is to find a way, that beginners and not-daily-user get a option setup that allow to render clean and fast images in less render time.
A first step is done with Vlados unversal method - an full adaptive mode based on QMC/LC. The calculation is slow, but bring stable solutions - slow, mainly for low noise.
My idea is it, to find a full adaptive setup based on IM/LC. The IM calcualtion use approximations and is the base for fast rendering - the user set an image quality level (noise level) and not abstract subdivs everwhere.
My experiments show me:

(1) the full adaptive QMC sampler (1/100 amount 1 noise 0.01) together with IM cause sometimes artefacts (interiors)
(2) artefacts - because the noisethreshold 0.01 is to low for the IM calculation
(3) a noiselevel of 0.005 and lower makes, that the IM calculation works clean and fast
(4) a noiselevel of 0.005 is very strong for the final pass, but can be disarm by an adaptive amount 0.95
(5) no setup of the material subdivs is necessary

Here the example scene compared with the classical manual settings:


I have tested the option setup with a few scenes and it works fine. Manual settings are slower or only a little bit faster. At simple exterior scenes the speed of the full adaptive mode is a little bit slower (50s vs. 70s), but at interior the adaptive mode improve the speed. For interiors works a noiselevel of 0.005 and for exterior a noiselevel of 0.01 fine.

This method could be improved, if the noislevel of final pass and IM pass could be independent or work better together. Than the adaptive amount could be set at 100% at noiselevel above 0.005 (interiors).

Now I hope, after a whole day of test, nobody come and say: it's totaly nonsense. The method is not perfect, but a step in an interesting direction or?

PS: A now I hope I can bring this method together with progressiv rendering ala PQMC.

And here two interior renderings with Quasi adaptiv IM (QAIM):

... "use for glossy rays" of the LC enabled


I'm sure, not anything is perfect - the adaptive amount could be set at 100%, the memory usage is high and the two buckets (if a dual core is used) could cause different noise levels at the buckets, but I'm sure, this problems can be solved. After Vlado has done the first step with his universal method here my second step for a fast universal method.

[studioDIM]

Cleaner now, but i still don't think i grasped it properly: you want a way to have the image stop mid-way a qmc AA pass (and sampling in case of glossies and shadows) to have an image which is of acceptable quality.
So, should we keep the small buckets, vray could display each iteration pass of the QMC sampler?
Is that why you use N (where N is the number of active threads) buckets and would want them to "refresh" once a minute or so?

Lele

[Micha]

Yes, the QMC process can stoped at mid-way and you will see the shooted samples. Tested and works fine here.
If this method should work proper, we need a refresh once a minute or more or less.
I used two bucktes, because my machine is a dual-core, so both are used. Better it would be, if Vlado could code something, that no bucktes are needed and the QMC use all CPUs to sample the whole image ala PPT.

General note: I'm not sure I say it: the quasi adaptive IM can be used independent from the progressiv QMC.

[Micha]

progressiv ... I hope the max framebuffer works like the rhino: if you cancel the rendering at the progress window, the current buckets show the sampled samples. This is like it works here allways.

[studioDIM]

the vray framebuffer seems to be the exact same across the apps, works the same here.

Lele