[TIP] De-Noise Using Multiple Seed Values

[rogerpenna]

I made only one of the tests using the wrong averaging... and noticed no difference. The others I used the 100-50-33 method (did not to go further in any except one that I used 5 seeds)


new test of the living room above, with longer render times... 3 seeds each rendered 2 minutes x 1 seed rendered 6 minutes


animated gif (Despite the loss of quality of the gif, it´s still possible to see areas with differences in quality)

ceiling looks better in the averaged image, but the LED tape lights from behind the TV seem better (imho) at the 6 minutes render.

[rogerpenna]

here, I see no difference between the SSS Cogs with 100x50x50 (layer opacity) vs 100x50x33 (scene is divided by 3x4 squares, the ones with 100x50x50 over it are obviously using that layer opacity method, the others without nothing written are using 100x50x33.

and well, I actually forgot the draw the squares... but you can't even notice any seams on it anyway, and that´s the point, right?



http://i.imgur.com/vllt2CM.png

[rogerpenna]

here Dave, this was the image I tested using different layer opacities, because I did not knew how to progress beyond the 33% opacity and I had used 5 seeds.

So... a pizza image showing layer opacity as 100-25-25-25-25 (Wrong method) and 100-50-33-25-20 (correct method, following your link)

can´t see any difference, sorry

[rogerpenna]

ok... some important considerations... the link provided is obviously for photography, not renders. And it never mentions that the averaging method results in a more noise free image than photographing for the same amoung of time in a single shot.

the averaging method is suggested to be used by cameras that have restrictions in the amount of time you can leave the shutter open, so you are FORCED to take like 3 20 second shots instead of 1 60 seconds shot. Or for shots where you want to capture low light but still needs a moving object to stay sharp, and stuff like that.


that said, there IS a slight improvement in noise reduction in some of the tests. So it must be investigated further.

in the same way, I guess this method can be used if for some reason the Thea user cannot render for a long time... but can make several shorter renders.

Example... leave it to render while you go to lunch, but when you come back to office you need to work on the computer and must stop the render.

Well, instead of rendering a 3 hours render, how about 3 1 hour (lunchtime) renders?


Using the averaging method, you might get about the same result as the 1 hour render.


That ALONE is a very good use for this method, even if we reach the conclusion the averaging method doesn´t produce better images than a single render.

I mean, they are AT LEAST of comparable qualities. So exchanging longer renders for shorter renders, WHEN NEEDED, might be a good option.

[DavePSB]

EDIT - posted the following before receiving your last reply. I think we are on the same page...


Well, 100/50/50 is certainly better than 33/33/33, but I feel the explanation and theory presented in the link I posted earlier is sound, so might as well use 100/50/33 no?

I have now done some testing on the Gear scene and in fact do not see such a dramatic difference between my 6 min version and my 3 x 2min 100/50/33 version. Perhaps I saw more of a difference in my first flesh colored test because the combination of the scale and different lighting caused a more difficult solving environment for the SSS calcs. Only a guess.

I suppose we have to say that in some cases, this stacking of different seeds may not be worth the trouble. But for those must-have scenes that will simply not clean up, perhaps the seed-stacking thing might be worth a try. Certainly, in my first test, it was dramatic.

But use the right weights!

Thanks again for the interest and tests.

[rogerpenna]

Don't worry. Despite not seeing difference in the tests with layer weighs following the correct methods or not, I will follow it when needed.

Anyway, as I said in my post you saw later, this method absolutely HAS its uses. Forget comparisons to see which is better. Image quality seems to be SIMILAR, and that means the method is useful for people who, for time constraints, need shorter render sessions...

--------------
I am curious what would be the result of hundreds of different rendered seeds. But that would need to be automated, both the rendering as also the post pro averaging

[rogerpenna]

Btw, for more efficient methos of reducing noise, chech the Neural networks" thread at Corner Bar
https://thearender.com/forum/viewtopic.php?f=24&t=17470

[yglemarec]

men, very interesting tests !!!
i will surely make a test before my next project
thanks a lot!