Smart noisethr for focusing on noisy areas with RT/progressive ?

[Subburb]

Hello !

i like playing with the noise thr interactlively with RT.
showing mask, I start with 0.005 then wait, observing RT speeding up when mask progressively disappear, as RT is focusing only on remaining Red parts.
Btw, "track mouse while rendering" is a GREAT idea with progressive mode...

i've seen today that the new version of mawxell can focus rendering via a bitmap as mask.

but why not try better than that ?

let's imagine a way, where in Vray progressive, noise thr starts at 0.005.
Then let's cook until 90% of image have reached noise thr 0.005.
Than, like "by hand", but automatically, change noise thr to 0.004.
Again, wait until 90% of pixels have reached noise thr 0.004.
Then, do the same for 0.003, 0.002, 0.001, etc...

Maybe i'm wrong, but i'm pretty conviced that it will help the engine to focus power on problematic areas...
At Least it will help to keep the same grain ratio over the image...
Or perhaps adaptive sampling do already the same kind of stuff ?

Just an idea.

[Art48]

Uhm, am I mistaken or is that the whole purpose of noise threshold? When VRay reaches a certain threshold per pixel it will stop sampling there and sample more on areas which arent there yet?

[super gnu]

yes as far as i can tell this suggestion just means the first bits to finish would have a noise threshold of .005 and the harder bits would end up with a progressively
lower noise thresholds. i must admit im struggling to see the benefit..

having said that, it was me who was asking for a mappable noise threshold years ago, so i could use a falloff map or something to force vray to concentrate samples on a distant forested horizon that was giving me grief.

ive now been convinced that any adaptiveness to the noise threshold doesnt make a lot of sense, unless you want some areas more noisy than others, and even then, its not exactly hard to deliberately make an area noisy in vray using other methods.

[vlado]

Technically it is possible to organize things so that more efforts are put in noisier areas first, so that the noise levels are more uniform across the image, but it's a bit tricky to get it right and I haven't gotten there yet.

Best regards,
Vlado

[Subburb]

I can figure easily how complex it must be...
That's why i thinked to this very simple stuff...
Of course i understand that there's no time to gain here, but, as you say, perhaps a more regular grain level over the image while rendering.

RT have been greatly improves in 3.1, LC as secondary is really improving performance.... And observing Vraysamplerate element in reatime is quite instructive, too.
I've seen you've worked to the noise thr part as well. Now it everytime end up correctly.

Sorry for this 2 cents post.... But sometimes, it's important to push the wall, just to see if the wall can fall, we never know

[3LP]

I though of that few weeks ago, so I definitely second this one as well as I'm struggling to get noise free renders with RT.

When the noise is ok on 90% of the image, to get the last 10% at the same noise lvl, it lakes sometimes 10 times to render-time than what took the first 90%.
By that time, the first 90% is dead sharp (I don't need that) and it just seems that the resources are not used the way I really want.

If it would be possible to get the whole image to go by steps in noise, IE : every 0.002 noise threshold, it would be awesome, and a massive gain in render-time.

Stan

[vlado]

If it would be possible to get the whole image to go by steps in noise, IE : every 0.002 noise threshold, it would be awesome, and a massive gain in render-time.

It won't really change render times that much - after all, the same amount of samples are needed for a clean image no matter what order you decide to compute them in. The only thing that changes is how the image looks at the intermediate stages.

Best regards,
Vlado

[3LP]

I understand,
Let's take a quick example (not real data) :

Render RT as it's now :

10 min render : 50% noise 0.01, 40% noise 0.02, 10% noise 0.05
20 min render : 50% noise 0.005, 40% noise 0.01, 7% noise 0.05
60 min render : 80% noise 0.001, 5% noise 0.01, 5% noise 0.05

Render RT with steps :

10 min render : 100% noise 0.04
20 min render : 100% noise 0.03
40 min render : 100% noise 0.02

For the time used, it's way more effective to have a consistent noise across the whole image in stead of having few bits racing though and few jogging behind.

1) You might never need to have that 0.005 or 0.001 noise on those pixels, meaning you "wasted" processing power that could have been used to help cleaning those nasty bits of noised areas.
2) You might want/need to stop your render at any time, it's not that cool to have a "half" render done, I rather have the whole thing consistent
3) It's really hard to asses how long it will take to render your expected quality. With production vray, when you set noise 0.01 and it's rendering in 1h you know that if you put 0.02 it will take 15min and 0.005 will take 4h. Each time you divide or multiply by 2 the noise, you divide or multiply render times by 4. Easy, predictable and critical in production.

If this makes it one day, it would also be great to have in the render dialogue the average noise the render is at for the moment. In the production render it says ie : 256 passes. Unless I'm not aware off, how does they relate with noise? It would be great to have : "noise thrs : 0.07" and 10 sec after 0.068, etc

Hope my logic make sens.

Stan