noise reduction?

[The Unknown]

Hi,

is it possible to speed up the noise reduction with the following method:

1. start rendering a scene
2. store a image with sample per pixel 1000 and 1500
3. calculating the difference of each pixel
4. pixels with large difference will be rendered more (for a while)
5. repeat the algorithm after spp x


For example:
I have a room with a big window. The outside looks good and
needs no further calculation (no visible changes).
Is it possible to use the calculation for other pixels now.

[IanC]

That would be like importance sampling, which I'm pretty sure is already implemented.

[Zom-B]

That would be like importance sampling, which I'm pretty sure is already implemented.

I'm not so sure about this... maybe Ono could drop some lines about this topic.
Hereis an older Thread, that describes importance sampling and links to the related paper...

[deltaepsylon]

interesting... but i think noise reduction should be in the form of a separate as-easy-to-use-as-indigo app. Remember, indigo is a renderer, not a renderer-&-noise-reducer

[CTZn]

I don't quite agree, isn't Indigo a noise reducer AS a renderer ?

Object based sampling is an alternative, but adaptative would be best.

[Zom-B]

interesting... but i think noise reduction should be in the form of a separate as-easy-to-use-as-indigo app. Remember, indigo is a renderer, not a renderer-&-noise-reducer

He He.... the Idea behind this is to focus Render Power of Indigo on noisy parts of a Scene
(like for example caustics or SSS objects),
and NOT to do a denoise Filter or stuff

[Sebastian]

it woud be intresting if "you" can set to each pixel a render-weight so that the noisy areas of an image have a higher weight which causes more process time or more samples per seconds to them.

[CoolColJ]

yeah I agree, definitely some need for weighted rendering because caustics are the last thing to be fully rendered and they take forever!
Probably due to the fact that even if the rest of the scene is already clean they are still receiving samples, where they should be directed to the caustics area

I know in Maxwell, caustics show within the first few minutes and are already pretty full even then

[CoolColJ]

maybe something like delta sampling - whatever it is

http://www.hungrycat.hu/DeltaSampling.html

[CTZn]

Nice one, and the other papers in that site rock too !

[Zom-B]

maybe something like delta sampling - whatever it is

http://www.hungrycat.hu/DeltaSampling.html

As far as it was discussed in the old Forum, it seems to be an technique that is used for animated content...

[The Unknown]

Thanks for your comments.

I only want to calculate pixels with large difference between image1 and image2.
The final result can be noisy (perhaps of bump maps or so).
Is "importance sampling" not a little bit different?

[CoolColJ]

maybe something like delta sampling - whatever it is

http://www.hungrycat.hu/DeltaSampling.html

As far as it was discussed in the old Forum, it seems to be an technique that is used for animated content...

from 2 years ago, maybe it will be ready right about now

Thank you for your good words about Hungry Cat.

The pictures have been made without noise filtering or photoshopping or any post processing. (Except gamma-correction on some cases.) There you can see the pure unbiased output.

Actually computation time data say practically nothing, because Hungry Cat is not finished yet. The public version will be equipped with Delta Sampling, which avoids the brute forcing.

Best Regards

Csaba Kelemen

[Xman]

someone say,that engine is give up,no longer continuing to develop the research.

[CTZn]

Fund that about delta-sigma modulation, the word unbiased is here again (?)

http://www.sciencedirect.com/science?_o ... ab6ba7fb31

Can't pretend I'm understanding, but looks like delta-sigma modulation allows to oversample signal for fine noise analyse, then a band-pass filter is used to retain relevant color information (ie noiseless colors).

Someone's got a better explanation (or a correct one ) ?