Page 9 of 13 FirstFirst ... 7891011 ... LastLast
Results 161 to 180 of 245
  1. #161
    Man! I'd really love to get my hands on a patched build!

    P.S.: I just tried to apply the patch in Git myself but I get this error msg:

    error: patch failed: intern/cycles/device/device_cpu.cpp:324
    error: intern/cycles/device/device_cpu.cpp: patch does not apply
    Last edited by SteffenD; 29-Jun-17 at 09:30.



  2. #162
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,953
    You try to apply in hash 839bf11 (that means, no denoiser available)
    I have built it on Linux, but it is a bit tedious to do CPU tests if you do not have really fast CPU (at the moment it is not implemented in GPU). Adaptive Sampling is not like the other features on this thread where it comes to having low noise renders in no much time and you can test these features on humble hardware. If I understand well it, Adaptive Sampling is more about achieving optimal results in optimal times, not necessarily very short times.
    Last edited by YAFU; 29-Jun-17 at 09:56.
    Be patient, English is not my language.



  3. #163
    I'm still scrolling through the extensive lists and tests here and this looks extremely promising.

    Grotesque speedups sometimes without any visual degradation. He also posted the mse and ssim ( mean squared error and structural similarity measurements ). I would seriously love to test this myself.



  4. #164
    Member lsscpp's Avatar
    Join Date
    May 2006
    Location
    Firenze
    Posts
    2,057
    What I miss is a comparison at same render time: plain vs adaptive.
    The original render takes hours but deciding what is 100% quality sounds arbitrary here.
    Everything's relative. Even saying "Everything's relative".



  5. #165
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,953
    Correct me if I'm wrong. What I understand is that with AS if you want to get no noise at all, you should set samples at least the minimum value needed for the more noisy material to render without noise. For example in BMW27.blend scene the more noisy part is the one corresponding to the headlights of the car. So you must set samples with a minimum value of at least the value necessary for this part to render without noise, you assume this is at least 100 square samples. So for me that would be the minimum value of samples to do tests with and without AS, obtaining similar images but with less render time when you use AS.

    Well, I can not spend days doing tests with 100 square samples on CPU, but here one with 30 square samples:

    No AS, 30 square samples:
    http://pasteall.org/pic/show.php?id=116819

    AS: 0.01 - 100 - 1 - 0
    30 square samples
    http://pasteall.org/pic/show.php?id=116820

    I guess AS would also be useful for getting a more uniform noise level in the whole scene, which is better for denoiser.
    Be patient, English is not my language.



  6. #166
    Member Ace Dragon's Avatar
    Join Date
    Feb 2006
    Location
    Wichita Kansas (USA)
    Posts
    28,360
    If some of the tests are credible (see the ones with very short times yet good quality), then it means we may have a performance increase to the point where the rendertime isn't that far above the time to denoise the image when rendered to a state where it can work well (the original time is mentioned near the top of the page).

    This could be even more of a game changer than I expected, and could perhaps even make the Cycles engine one of the fastest solutions in the industry.
    Last edited by Ace Dragon; 29-Jun-17 at 13:09.
    Sweet Dragon dreams, lovely Dragon kisses, gorgeous Dragon hugs. How sweet would life be to romp with Dragons, teasing you with their fire and you being in their games, perhaps they can even turn you into one as well.
    Adventures in Cycles; My official sketchbook



  7. #167
    Member
    Join Date
    Oct 2012
    Location
    UK, England
    Posts
    542
    I played with per material sample rates a while back for exactly this reason when just rendering without adaptive samples.

    A good help here could well be introducing sample rates per material with this (certain mats always need more samples than others with cycles, it wouldn't be difficult to build a guide list of what sample counts on average are needed in certain situations. e.g External rendered, Internal, physical size in scene etc etc).

    But still as many others have said it's a very basic system right now, Lots of Lukas denoise stuff could be leveraged here.



  8. #168
    Member Ace Dragon's Avatar
    Join Date
    Feb 2006
    Location
    Wichita Kansas (USA)
    Posts
    28,360
    When it comes to per-material samples, we need to know the root cause of why some materials need a lot more samples than others.

    Many times, the material may produce a lot of noise due to a few specific shading components in the node tree (such as heavy use of SSS, translucency, and rough refraction and reflection). What we need is a feature where certain closures get more samples than others based on their tendency to produce a high level of noise (regular path tracing would see cleaner material rendering as a result without the much slower method of giving every component a sample for each AA pass).

    I believe Isscpp proposed something similar once and it appeared like a good idea (but I think the focus first of all should be for the adaptive sampling itself to get into master, let's not bog it down in a major way by asking for feature creep).
    Sweet Dragon dreams, lovely Dragon kisses, gorgeous Dragon hugs. How sweet would life be to romp with Dragons, teasing you with their fire and you being in their games, perhaps they can even turn you into one as well.
    Adventures in Cycles; My official sketchbook



  9. #169
    Member lsscpp's Avatar
    Join Date
    May 2006
    Location
    Firenze
    Posts
    2,057
    Originally Posted by YAFU View Post
    Correct me if I'm wrong. What I understand is that with AS if you want to get no noise at all, you should set samples at least the minimum value needed for the more noisy material to render without noise. For example in BMW27.blend scene the more noisy part is the one corresponding to the headlights of the car. So you must set samples with a minimum value of at least the value necessary for this part to render without noise, you assume this is at least 100 square samples. So for me that would be the minimum value of samples to do tests with and without AS, obtaining similar images but with less render time when you use AS.

    Well, I can not spend days doing tests with 100 square samples on CPU, but here one with 30 square samples:

    No AS, 30 square samples:
    http://pasteall.org/pic/show.php?id=116819

    AS: 0.01 - 100 - 1 - 0
    30 square samples
    http://pasteall.org/pic/show.php?id=116820

    I guess AS would also be useful for getting a more uniform noise level in the whole scene, which is better for denoiser.
    did you manage to make a build? which os?
    Everything's relative. Even saying "Everything's relative".



  10. #170
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,953
    Kubuntu Linux 14.04, not portable build.
    If people who build for Windows want to know how to apply patches in a version that is not the current master, here 'lukasstockner97' instructions:
    https://blenderartists.org/forum/sho...=1#post3099381

    This patch should be applied on hash "839bf11", so 'git checkout 839bf11'
    https://developer.blender.org/D2662
    Remembering this only works for CPU.
    And as I said above, this version does not have denoiser.
    Be patient, English is not my language.



  11. #171
    Member cekuhnen's Avatar
    Join Date
    Mar 2003
    Location
    Grosse Pointe Woods, Michigan, USA
    Posts
    11,983
    Originally Posted by wikifry View Post
    principled shader: 2x faster, minimum
    denoiser: less samples, may be between 200 and 1000

    really fast
    principled shader is considered faster?
    3D Design Generalist - Faculty Industrial + Interior Design
    Chair Interior Design - Wayne State University



  12. #172
    Member
    Join Date
    Jul 2017
    Location
    UK
    Posts
    66
    Sorry if it's already been stated. But will it be able to detect when rendering an area of the scene that is just background or an HDRI texture and containing no geometry, and simply render that square at super speed instead of running thousands of pointless samples? I witnessed mitsuba render able to do this trick..



  13. #173
    Member
    Join Date
    Oct 2012
    Location
    UK, England
    Posts
    542
    Ive done people a Experimental build from master with the extra experimental features and the new CPU (only at mo) adaptive sampling patch. Ive got a slow CPU so haven't been able to test this yet, Can someone with a fast ass CPU test this experimental build that adaptive sampling is working, I had to heavily hack it to work in current master.

    New Experimental with Adaptive Rendering: https://mega.nz/#!shZ2XbaB!x_QIQCcGH...pmMCk4YeJNhrXw
    Patch details for adaptive: https://developer.blender.org/D2662



  14. #174
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,953
    Hi 3DLuver.
    Is it possible to apply the hacked Adaptive Sampling patch to the current master? Could you share the patch?

    Edit:
    If so, does that mean that you can use AS and denosier at the same time?
    Last edited by YAFU; 04-Jul-17 at 17:33.
    Be patient, English is not my language.



  15. #175
    Member
    Join Date
    Oct 2012
    Location
    UK, England
    Posts
    542
    Originally Posted by YAFU View Post
    Hi 3DLuver.
    Is it possible to apply the hacked Adaptive Sampling patch to the current master? Could you share the patch?
    The patch was so old it had no chance of patching, So i decided just to see if it worked to try to recode the .diff by hand. Everything compiled and seems to work but ive not got a good enough CPU or understanding how to test the system so thats where you guys come in. I could try to make a .diff for you but there's lots of other patch's in my experimental builds and some of my own code changes like custom micro jitter for CMJ sampler and some other bits.

    When i try things out or add new things unfortunately I dont bother making individual changes and then commit which would allow patch role back to create a specific feature .diff because it just wastes my time.

    If the Build ive provided does work ill adapt it for Opencl, which is why ive done this test build. Have to know the cpu is working first and im on an old i5 2.66ghz. No point me trying to test this patch on my CPU

    Edit: Also yes in theory this has sobol and CMJ micro jitter scramble, and denoise. It's a very experimental attempt but people need to test it who know how to use the adaptive sampling patch. I dont

    Even if on Linux ive heard others say my windows builds work perfectly in wine, So linux users can try that.
    Last edited by 3DLuver; 04-Jul-17 at 17:50.



  16. #176
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,953
    Ok. I also do not know much about AS, so let's hope someone with more knowledge will try it
    Also I can not understand exactly what would be the correct way to use it along with denoiser. For me the most important use of using them at the same time would be to configure AS in the next way : "Render with many samples those parts where there are more noise, and leave all the noise level in all the image similar to the less noisy zone/area". That way you would get a uniform noise pattern/level in all the image, where you could then apply denoiser.
    Be patient, English is not my language.



  17. #177
    Member
    Join Date
    Oct 2012
    Location
    UK, England
    Posts
    542
    Originally Posted by YAFU View Post
    Ok. I also do not know much about AS, so let's hope someone with more knowledge will try it
    Also I can not understand exactly what would be the correct way to use it along with denoiser. For me the most important use of using them at the same time would be to configure AS in the next way : "Render with many samples those parts where there are more noise, and leave all the noise level in all the image similar to the less noisy zone/area". That way you would get a uniform noise pattern/level in all the image, where you could then apply denoiser.
    Well when looking through the code i get the impression because this is a very simple system it ties directly into cycles sampler, in theory it should be a matter of the adaptive sampling all happens as normal before the denoise pass is even run, so it should make a difference.

    Ive done some tests and nothing crash's with using adaptive with denoise at the same time so thats encouraging, but as i had to hack the .diff to even work im not 100% sure it's even affecting the render. People on the .diff page say it's more about rendering at high sample rates but getting lower render times compared to no adaptive. But i render only on GPU as i only have an i5 2.66 with 12 gb main mem, but i have a AMD Firepro W9100 with 16GB GDDR5. And i just dont have the time or patience on my granny CPU to realy be able to test this.

    You did some test's above, Did you find your render times were shorter? If you could repeat the same thing if it did at least were know the code is running as should in the background.



  18. #178
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    2,953
    I've done a couple of tests without knowing anything about what each configuration item means. I think that at same samples, always times have been shorter with AS than without using AS, and obtaining very similar resulting image. But always the noisiest part is still the noisiest part (obviously less noisy the higher are the configured render samples).
    What I wish I could configure for example in BMW27.blend scene is: "Render with many samples that part of the headlights of the car where there is more noise, and not waste time rendering with many samples the body of the car where there is little noise". And somehow then be able to set an even level of noise throughout all the image.

    Edit:
    Ok, reading diff comments again, Brecht said:
    'For example BMW no adaptive vs. adaptive with similar render time. There it's doing basically what you'd expect, distributing noise more evenly.'
    So I just must be setting it wrong. What mattered most to me was to be able to get an uniform distribution of noise to save time and to get better result with denoiser. And it seems that it works, I just do not know how to set it up
    Last edited by YAFU; 04-Jul-17 at 19:44.
    Be patient, English is not my language.



  19. #179
    Member lsscpp's Avatar
    Join Date
    May 2006
    Location
    Firenze
    Posts
    2,057
    Originally Posted by 3DLuver View Post
    Ive done people a Experimental build from master with the extra experimental features and the new CPU (only at mo) adaptive sampling patch. Ive got a slow CPU so haven't been able to test this yet, Can someone with a fast ass CPU test this experimental build that adaptive sampling is working, I had to heavily hack it to work in current master.

    New Experimental with Adaptive Rendering: https://mega.nz/#!shZ2XbaB!x_QIQCcGH...pmMCk4YeJNhrXw
    Patch details for adaptive: https://developer.blender.org/D2662
    Many thanks. I was waiting for a build to test this patch.
    Anyway, it doesn't seem to work here. No matter the settings I use (from 10 to 0.0001 in the Norm) i get the same quality and the rendertime is just slightly more than regular render with adaptive sampling disabled.
    (Path tracing + Cpu rendering)

    edit: I even tried exactly some of the settings from this page

    Am I missing something?
    Last edited by lsscpp; 05-Jul-17 at 04:40.
    Everything's relative. Even saying "Everything's relative".



  20. #180
    Yesterday I managed to checkout the very old master 839bf11 and apply the diff thanks to this posting by YAFU: https://blenderartists.org/forum/sho...=1#post3213068

    The results were shockingly good (and always faster than the non-adaptive renderings) and although the settings still are kind of "cryptic" I could figure out what to tune to get faster / grainier or slower / cleaner images. The noise pattern looked very even especially with very small bucket sizes.

    Now I just compared it to the 3DLuver build (patch applied to the current master) and can say that something obviously went very wrong. The "hacked" version doesn't really react to the "norm" value as expected (which IMHO should be something similar to a threshold in other render engines). Even setting it to something like 10.0 doesn't give me a super grainy but fast render like in the older patched master.

    Maybe it would be best if Milan Jaros, the original author of the patch would update his work.
    Nonetheless this looks super promising and can save loads of render time not only on super fast CPUs but also and especially on older ones.



Page 9 of 13 FirstFirst ... 7891011 ... LastLast

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •