[TIP] De-Noise Using Multiple Seed Values

Created a nice video, document or a translation of another document? Share it here for licensed users.
DavePSB
Licensed User v1.5
Licensed User v1.5
Posts: 1291
Joined: Wed May 12, 2010 8:56 pm
Location: France
Has thanked: 13 times
Been thanked: 16 times

I was experimenting with SSS and was looking for ways to improve render speed/quality. I ran across something interesting so thought I'd flag it here in case it has other applications. I know there are other denoising filters, software, and techniques, but this one seems to preserve detail and edges quite well.

I studied a particular scene and found that by rendering the scene 3 times, each for 2 minutes, but with a different seed value, and then combining the three renders (overlaying in postpro), the resulting image was much much better than rendering the scene for 6 minutes using a single seed value.

Here are the images:

The first one is the result from letting Presto MC go with default settings and Seed=1 for 2 minutes.
PrestoMC_2min_Seed1.png
I have two additional images, which I won't upload due to 3 image limit, that show the result of Seed=2 for 2 minutes and Seed=3 for 2 minutes. These images look virtually identical to the first image in terms of quality and graininess, but differ slightly on the pixel level due to the effect of the seed.

Here is the result of rendering with Seed=1 for 6 minutes. There is little improvement over the first image, which was rendered for 2 minutes.
PrestoMC_6min_Seed1.png
And here is the result of overlaying the three 2 minute images (Seed=1, Seed=2, Seed=3) using overlay weights of 100%, 50%, and 33% respectively. These weights are important and are a function of how many images are being overlayed. The weights might also be called opacity depending on the postpro app. No other adjustments were made regarding image brightness etc.
PrestoMC_2min_100Seed1_50Seed2_33Seed3.png
Well, I thought it was interesting!
Win 10
i7-9700K,3.6GHz,32Gb,
RTX2060Super, 8Gb, 442.19
SUMake2015 (15.3.331)
TheaForSU (1.5.09.547.1462)
Thea Studio (1.5.09.1462)
SUMake 2017 (17.2.2555)
TheaForSU (2.2.954.1860)
Blender 2.79, Plugin 1.5.8.760.1455
Blender 2.8x
Revised 14 Feb 2020
User avatar
SandroS
Licensed User v1.5
Licensed User v1.5
Posts: 3062
Joined: Wed May 20, 2009 8:20 am
Has thanked: 1 time
Been thanked: 43 times
Contact:

DavePSB wrote:Well, I thought it was interesting!
it is :)

isn't this a similar technique used when taking pictures of the night sky?

there's a pretty big difference between the straight 6 min render and the 3 x 2min one.

have you done any other tests?

i'm pretty sure i tried to do this a while back but didn't see any real difference at the time, but i think the problem is i never changed the seed value... i'll definitely have another go when i find some time.

i wonder if Giannis or someone else from the team could comment on whether there's something good to take from this and, if so, whether it's possible to automate it rather than have to do it all manually :?:
User avatar
rogerpenna
Licensed User v1.5
Licensed User v1.5
Posts: 615
Joined: Thu Jun 27, 2013 4:58 am
Location: Neu Hamburg, Brazil
Has thanked: 0
Been thanked: 0

DavePSB wrote:Well, I thought it was interesting!
i wonder if Giannis or someone else from the team could comment on whether there's something good to take from this and, if so, whether it's possible to automate it rather than have to do it all manually :?:
maybe Thea could render areas with SSS multiple times (like 10 times), for 1 second with each different seed and automatically join all results together.
User avatar
rogerpenna
Licensed User v1.5
Licensed User v1.5
Posts: 615
Joined: Thu Jun 27, 2013 4:58 am
Location: Neu Hamburg, Brazil
Has thanked: 0
Been thanked: 0

here are my tests with 1 SSS ball rendered 3 times, 30 seconds each, with 3 different seeds

seed 1, 2 and 3
Image

just one seed, two seeds overlayed and three seeds overlayed
Image
User avatar
rogerpenna
Licensed User v1.5
Licensed User v1.5
Posts: 615
Joined: Thu Jun 27, 2013 4:58 am
Location: Neu Hamburg, Brazil
Has thanked: 0
Been thanked: 0

here another test... fifty seconds render on the left versus five 10 seconds renders with 5 different seeds on the right. (ps, I see no difference in setting layers transparency as 50, 33, 16 against setting them 100/4=25% (main layer 100%, then layers above 25% each)

Image


notice that noise wise, it´s almost the same. HOWEVER, the 50 seconds render on the left has a bluish tint (I used RED absorption and PURPLE scattering) that is mostly absend on the 10 seconds renders.


I guess the smaller render time never got any ray of light from the deeper purple layers?
User avatar
rogerpenna
Licensed User v1.5
Licensed User v1.5
Posts: 615
Joined: Thu Jun 27, 2013 4:58 am
Location: Neu Hamburg, Brazil
Has thanked: 0
Been thanked: 0

test with Thea SSS Cogs scene

three 30 seconds seeds
Image

vs 90 seconds render
Image


not much render imho... :think:


edit: all these images are bigger at imgur. Is thea forum reducing linked images sizes?

here, the links for bigger images
http://i.imgur.com/VfRtIuZ.png
http://i.imgur.com/WBupnSx.png

looking the direct links, the image finishing uZ (90 seconds render) actually looks cleaner than the image finishing Sx (3 30 seconds renders) :think:

which is funny, because the first ball tests looked CLEANER with the several overlayed renders.
User avatar
Frederik
Thea A Team
Thea A Team
Posts: 5581
Joined: Fri Jul 18, 2008 2:24 pm
Location: Denmark
Has thanked: 43 times
Been thanked: 100 times

Excellent tests, rogerpenna...! :clap: :thumbup:

I agree that the 90 sec render actually looks cleaner than the 3x30sec render...
Cheers
Kim Frederik

"If you have no critics you'll likely have no success..."
-Malcolm X
User avatar
rogerpenna
Licensed User v1.5
Licensed User v1.5
Posts: 615
Joined: Thu Jun 27, 2013 4:58 am
Location: Neu Hamburg, Brazil
Has thanked: 0
Been thanked: 0

another test... this one for lights and shadows...

5 seeds of 30 seconds vs 150 seconds test
Image


I think the 5 seeds look better than the 30 seconds... but not by much.
User avatar
rogerpenna
Licensed User v1.5
Licensed User v1.5
Posts: 615
Joined: Thu Jun 27, 2013 4:58 am
Location: Neu Hamburg, Brazil
Has thanked: 0
Been thanked: 0

Frederik wrote:Excellent tests, rogerpenna...! :clap: :thumbup:

I agree that the 90 sec render actually looks cleaner than the 3x30sec render...

now, what conclusion can we reach? In my spheres test, the 3 different seeds results were BETTER than the longer time with 1 seed. However, the longer render time gave a more precise result (the bluish tint)


the lights and shadows test there is little difference, but several seeds result was better... and the SSS Cogs show a better result with the longer render time.


My hypothesis:

the results of different seeds are better WHEN a long render sees no longer difference in quality as minutes pass. Like that scene where you leave it rendering, returns 10 minutes later and sees no difference? Maybe that scene would get better rendering 6x10 minutes seeds, than 1x60 minutes...

must check the theory with longer renders.
DavePSB
Licensed User v1.5
Licensed User v1.5
Posts: 1291
Joined: Wed May 12, 2010 8:56 pm
Location: France
Has thanked: 13 times
Been thanked: 16 times

Thanks for the extensive testing, but please, you MUST use the correct weightings. You can't just set all the weights to the same value. Perhaps the sphere scene was not the best to highlight these differences - its a bit dark. But regardless, please read this article - he explains it quite well.

http://www.cambridgeincolour.com/tutori ... -noise.htm

I believe the way you are weighting it is actually putting way too much weight on just one of the images (the one at the top of the layer stack (talking post-pro now). So you are seeing a lot of the noise from that image, and not much of the averaging that you should be seeing.

@Sandros - I haven't tried any other tests yet, but I think the same improvement will be seen in TR1 and TR2. I just ran out of time.

EDIT: I meant the SSS Gears, not Mechano..

@Roger - Perhaps try the Gears SSS seen using three seeds, but weight them 100, 50, 33. Follow the PhotoShop procedure outlined in the link above. See if it makes a difference. If I can find the time, I'll have a go, and also document the technique a bit better. I should have augmented my original post with a little more information.



I agree that if there is something to this, it might be made an option in the appropriate render engine, and perhaps the darkroom, dare I say it, could do the weighting and postpro behind the scenes and present the final image to the user. If nothing else, I feel another batch program coming on...
Win 10
i7-9700K,3.6GHz,32Gb,
RTX2060Super, 8Gb, 442.19
SUMake2015 (15.3.331)
TheaForSU (1.5.09.547.1462)
Thea Studio (1.5.09.1462)
SUMake 2017 (17.2.2555)
TheaForSU (2.2.954.1860)
Blender 2.79, Plugin 1.5.8.760.1455
Blender 2.8x
Revised 14 Feb 2020
Post Reply

Return to “Tutorials & Translations”