Chaos Corona Forum
General Category => Gallery => Work in Progress/Tests => Topic started by: maru on 2016-03-01, 12:43:54
-
Note: Currently only a few lucky ones were handed the new daily build with experimental adaptivity and denoising implemented. The features are not production-ready yet, and we are still testing them to see how and how much they can help.
Let's post our experiments here to show how awesome this new features are, and to share each other's observations.
Example 1 - both adaptivity and denoising enabled. The scene features DOF, volumetric fog, glossy materials, car paint. In this case the results are simply great. Despite the denoising, you can clearly see the pattern on the tire, or on the grid next to the lower halogen.
(http://i.imgur.com/90e59MJ.jpg)
(http://i.imgur.com/C8H3xPY.jpg)
Example 2 - adaptivity on vs adaptivity off, no denoising. Generally the results are good as the noise is more uniform and much weaker in some areas. You can however see that there is slightly more noise visible in the background. As I understand - this is caused by having some areas of the image refined at the cost of some other areas.
(http://i.imgur.com/JR02VKI.jpg)
(http://i.imgur.com/MpTJkpx.jpg)
Comparison - this is based on the second set of images (adaptivity only).
(http://i.imgur.com/sIEp9SJ.jpg)
I am also attaching all the images if anyone wants to inspect them with no additional compression.
-
In the last crop showing white paint corner, did you attach the images correctly ?
Because adaptivity made the overall noise levels worse. The other examples show minor improvement of details (rather then less noisy they are better articulated).
-
Looks kinda ok. Are there controls over denoising strength? Do you have maybe a more agressive tests? Just curious how much can it be pushed etc..
-
In the last crop showing white paint corner, did you attach the images correctly ?
Because adaptivity made the overall noise levels worse. The other examples show minor improvement of details (rather then less noisy they are better articulated).
Yes, that one is correct. As I wrote - I think this is because the noise level becomes more uniform, so given the same time, some areas had to "sacrifice" quality so that other ones can be improved. I think this is also important for the denoising to work correctly.
@Lacilaci: yes, there are controls for denoising strength and adaptivity error tolerance, but the UI is not yet finished, so I don't think it should be discusses. There are also some bug reports about the denoising already, so it has to be improved before any official tests.
-
A crop from a scene that Cecofuli posted few years ago as prime example where out-of-focus reflections wouldnt clear up. Only adaptivity.
-
This last test is very, very interesting. Also in V-Ray it was hard to slove bright highlight reflection with strong DOF! Great result! O_O
-
A crop from a scene that Cecofuli posted few years ago as prime example where out-of-focus reflections wouldnt clear up. Only adaptivity.
It seems that worst case scenorios with sharp reflections in Dof are no longer worst case scenarios. Tho it does not make big difference on the images that does not have wird noise problems. Am i right?
-
actually it really helps even on images that previously looked OKish, as long as there are some differences (shadows vs. bright areas, interior/exterior, ...). Obviously it wont work if you render only an empty wall ;)
-
I wounder how this will work out with fumeFX. As long as if the scatering for the smoke (or even without the scatering) is turned on rendertimes geting dramaticly long but fume itself clears from noise in one pass.
-
Great thread!
maru, in your first set of images, what render settings was? Number of passes is almost identical with adaptivity and without, does that mean that adaptivity almost took no place?
-
maru, in your first set of images, what render settings was? Number of passes is almost identical with adaptivity and without, does that mean that adaptivity almost took no place?
Shamefully, I don't know what were the render settings as I have already altered the scene in every possible way and overwritten it. I am pretty sure the settings for the nice one with denoising were GIvsAA 64, LSM 1, adaptivity 6, denoising 1.
About the number of passes - I think it's more complicated. I noticed that usually the number of passes with adaptivity enabled is actually lower, but the overall quality is better.
-
My tests...
1° 5 min without any trick..
2° 6 min adaptivity on default setting
3° 6 min adaptivity on and full denoise on 0.9
4° 24 min without any trick.. 201 pass
-
in this scene adaptivity can't work perfectly but i noticed a improvement on glossy reflections..
denoiser it's amazing, It should be used moderately but i love it :)
-
where is the adaptive setting?
-
just ran a test with one scene i am working on currently, this is awesome! denoising is so nice to use
-
My tests...
1° 5 min without any trick..
2° 6 min adaptivity on default setting
3° 6 min adaptivity on and full denoise on 0.9
4° 24 min without any trick.. 201 pass
6 min/ 24 min difference. 4X better rendering?
-
6 min/ 24 min difference. 4X better rendering?
Nope, rendering is the same (if you don't use adaptivity, that is), it's only does denoising (bluring/smudging) pass on top.
-
Here is a short test.
Motion Blur (camera + geometry) + Depth of Field
Render time = 2 minutes
(http://i.imgur.com/Zt1roqP.jpg)
(http://i.imgur.com/2pGiNNf.jpg)
(http://i.imgur.com/zNVLJbq.jpg)
(http://i.imgur.com/pchBqaX.jpg)
With heavy MB and DOF, which is obvious, adaptivity has a very low, almost neglegible impact.
-
Using denoise you can have similar (but not identical) result if you compute the image for 24 minutes, but for me this is not a point.. denoiser can resolve, with adaptivity, that part of image with heavy noise.
-
20 min renderings with (top to bottom): adaptivity only, denoising with radius 0,5
(http://i.imgur.com/dBQLtyU.jpg)
(http://i.imgur.com/wV002dD.jpg)
10 min renderings with (top to bottom): adaptivity only, adaptivity + denoising with radius 0,5
(http://i.imgur.com/BmOuy9z.jpg)
(http://i.imgur.com/DuzDSWu.jpg)
-
hi there,
i have found that the aa mode plays kinda strong role on the outcome of denoise feature - when aa is off it does not work?! can anyone confirm nthis?
greetings
-
What do you mean by "aa mode"?
-
Scene that gave me the creeps is no longer the problem.
150 pases no adaptivity no denoising
(http://s11.postimg.org/t2fdgczzn/image.jpg)
150 pases adaptivity denoising
(http://s11.postimg.org/piodtfqyb/image.jpg)
-
Look like it's proved :- ) Looks great.
Anybody test it on animation ? Does it create some flicking artifacts ?
-
Anybody test it on animation ? Does it create some flicking artifacts ?
I rendered two short animations with adaptivity and denoising - didn't notice anything wrong.
-
Look like it's proved :- ) Looks great.
Anybody test it on animation ? Does it create some flicking artifacts ?
I will render 10s of this tesla animation tomorrow. As soon as i will instal DB to the rendernodes.
-
this is my screen, where is the adaptivity option to enable and disable?
-
Performance tab/ Dev./Experimental stuff/ ...bottom
-
Thanks!
-
hi guys,
with adaptivity and denoising, when I click stop the rendering progress it can't be stopped immediately just like usual ... I have to wait for few minutes and for big 4k px image it needs about more than 20 minutes to stop. my CPU is i7-4770.
do you guys experience this? is it a normal thing? if it is true, I think it would be better if after click stop there is a note pop out explain about denoising progress is starting. it makes the user know what's happening
-
What is the best use of 'Noise limit'?
-
hi guys,
with adaptivity and denoising, when I click stop the rendering progress it can't be stopped immediately just like usual ... I have to wait for few minutes and for big 4k px image it needs about more than 20 minutes to stop. my CPU is i7-4770.
do you guys experience this? is it a normal thing? if it is true, I think it would be better if after click stop there is a note pop out explain about denoising progress is starting. it makes the user know what's happening
Yes, denoising requires additional time to compute... and additional RAM too, unfortunatelly. And no No NO NOOOO to additional pop ups, we have enough of those already. Simple message in render progress window should be more than enough.
What is the best use of 'Noise limit'?
Same as pass limit or time limit, but in this case it will stop rendering process based on residual noise.
-
@Nekrobul: I am also curious about seeing this as an animation - especially if the car would turn - the reflections on its sides
Yep, denoising makes my PC want to explode. It sounds like a starting F16 during the denoising.
-
With my simple test, I have a very strange behaviors:
FULL HD - Internal resolution 2 - AA filters ON
Regular render 3.9 Million rays : 3 min 40 sec
ADAPT ON 3.1 Million rays: 4 min 24 sec
ADAPT ON + DENOISE 1.0 1.7 Million rays = 6 min 43 sec
FULL HD - Internal resolution 1 - AA filters OFF
Regular render 3.9 Million rays : 3 min 45 sec
ADAPT ON 3.1 Million rays: 4 min 24 sec
ADAPT ON + DENOISE 1.0 2.1 Million rays = 5 min 41 sec
-
Simple test. A 480px image. A) 100 pass, unfiltered 2:34 B) Adapt_Denoise_1r_0% 2:43 C) Adapt_Denoise_1r_20% :48
This will work great for my workflow. B) same time, better image C) 60% less time, comparable results
Did a 200 frame 480px animation of C) Adapt_Denoise_1r_20% :48, clean results
-
hi noah45, how do you give a percentage to th adaptativity?
-
hi noah45, how do you give a percentage to th adaptativity?
It's the noise limit % (below pass limit)
-
This is a crop of a rendering (zoom x4) which has never cleaned up, consisting of a recfractive glass, dof setup and 4 corona lights with visibility enabled except direct visibility.
1. 200 passes (you can render 500 passes it won´t get much better)
2. 200 passes with adaptivity enabled
3. 200 passes with adaptivity enabled and denoise 1px
Looks like it works very good in this situation.
Good Luck
-
I think it was a bug. Ondra already solved it )
Frood: i know very well your problem. I'm very glad to see how adapt is able to solve this situation!
-
hi guys,
with adaptivity and denoising, when I click stop the rendering progress it can't be stopped immediately just like usual ... I have to wait for few minutes and for big 4k px image it needs about more than 20 minutes to stop. my CPU is i7-4770.
do you guys experience this? is it a normal thing? if it is true, I think it would be better if after click stop there is a note pop out explain about denoising progress is starting. it makes the user know what's happening
Yes, denoising requires additional time to compute... and additional RAM too, unfortunatelly. And no No NO NOOOO to additional pop ups, we have enough of those already. Simple message in render progress window should be more than enough.
What is the best use of 'Noise limit'?
Same as pass limit or time limit, but in this case it will stop rendering process based on residual noise.
ah yes romullus, pop up can be an annoying thing. but I think a note for denoising progress is still a need wherever it placed.
-
ah yes romullus, pop up can be an annoying thing. but I think a note for denoising progress is still a need wherever it placed.
Absolutely. Let's hope it will be fixed before final release.
-
I think it was a bug. Ondra already solved it )
Really? I only know of two resolved issues regarding noise and IES lights lately, but let´s see!
Good Luck
-
Been testing the denoise, and i have a question. Is it possible to apply the denoise only to light passes? If you prioritise AA sampling, and the denoise could take care of the direct and indirect light noise?
-
hey!
this is bomb :)
just tested it on current projects, showed the collegues and they are freaking out (almost as much as when i first introduced the mighty corona to them)
the only feature i would ask for - because this would drive it to perfection:
we are using a lot network rendering, so we are not able to "tweak" the denoising amount manually on rendertime - and
maybe it can be useful to blend the denoise amount in different areas in the picture differently.
therefore i think a render-element with denoise amount of 1 would be very great to blend it manually
with the un-denoised beauty pass in post.
edit:
the CShading_Components can be used for that.
with all components checked you get an un-denoised beautypass :)
perfect!
lasse
-
Hi. Did some testing too. I used my usual test-file (Definline Modernist Showroom).
A) Is a 100 pass render without adaptive and denoiser. Rendertime: 11min. 42 sec. Corona version: 1.4 daily 2016-03-02
B) Is a 50 pass render with adaptive and denoiser (radius: 0,5). Rendertime: 5min. 24sec. Corona version: 1.4 daily 2016-03-06
I've tested other scenes ass well, and i find that for medium sized renders in most cases a denoising radius at 0,5 will make it possible to render only the half amount of passes as before (without adt & denoiser), without loosing important details.
It's great times for Corona users :)
-
6 min ... Max 2013 + 1.4 beta ... 4790K + 16 GB ram ...
but ... what is the sample error ?? dead pixels ?? noise amount ??
-
but ... what is sample error ?? dead pixels ?? noise amount ??
Yes, it's a measure of noise in the picture.
-
14 min ... Max 2013 + 1.4 beta ... 4790K + 16 GB ram ...
too noisy in mirror's reflection !! :\
-
too noisy in mirror's reflection !! :\
Here is reflection pass ...
-
14 min ... Max 2013 + 1.4 beta ... 4790K + 16 GB ram ...
too noisy in mirror's reflection !! :\
This topic is for showing personal experiments with adaptivity and denoising features. To post anything with post production applied here, makes no sense at all. Especially when that post productions so horribly ruins whole image.
-
Will the noise reduction feature take into consideration distance from camera and object/texture size/detail? So it doesn't confuse small intended geometry/texture detail with noise.
What's the approx release date of 1.4 by the way? I'm just starting an unreal engine project, and would be tempted to hold fire if it's in the next week, so I can take advantage of render to texture.
-
Will the noise reduction feature take into consideration distance from camera and object/texture size/detail? So it doesn't confuse small intended geometry/texture detail with noise.
What's the approx release date of 1.4 by the way? I'm just starting an unreal engine project, and would be tempted to hold fire if it's in the next week, so I can take advantage of render to texture.
As far as I read denoising papers... No, that information is not relevant at all.
You have known constants, normals after all the bump mapping and displacements stuff, albedo after the texture shading pipeline and coverage after the image has been sampled.
And the unknown constants, which is lighting. Since it is unknown we have to apply a large as possible filter kernel to average the samples, as supposed to average by sampling.
Let's define energy. Energy is where there is high variability of known stuff. More albedo noise, like grainy textures, high frequency normals increase that energy. This is the part where we cannot average and thus we chose the filter kernel low. As supposed to solid color, solid normals, where we have low energy coming from known constants, there we blur the crap out of the random lighting calculation. Lastly the local noise level is being used as a multiplier for the filter kernel aggressiveness. The coverage is being used as a cut-off not to blur edges.
As you see the (abbreviated) logic does not contain any info about camera position or textures per say, but it is solid in understanding what is detail and what is not.
The lead dev said here: https://forum.corona-renderer.com/index.php/topic,11763.msg75973.html#msg75973
, that they aim at 16.05 as a release date. a realistic +/-a week is to be expected.
-
What's the approx release date of 1.4 by the way? I'm just starting an unreal engine project, and would be tempted to hold fire if it's in the next week, so I can take advantage of render to texture.
Be aware, that RTT won't support projection mode in v1.4 If it's important to your project, you can wait no more and look into other options for baking until Corona brings support for projection baking.
-
Thanks for the clarification guys.