Page 39 of 90 FirstFirst ... 2937383940414989 ... LastLast
Results 761 to 780 of 1790
  1. #761
    Hi Lukas, did you forget us?



  2. #762
    Member Ace Dragon's Avatar
    Join Date
    Feb 2006
    Location
    Wichita Kansas (USA)
    Posts
    28,555
    Dito; If you look at the commit logs you will find that Lukas has been occupied with other Cycles development projects for the time being (though the roadmap I posted indicates that denoising for stills may get committed to master before the year is over).
    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



  3. #763
    Member MirceaKitsune's Avatar
    Join Date
    May 2011
    Location
    Romania, Bucharest
    Posts
    559
    Originally Posted by Ace Dragon View Post
    Dito; If you look at the commit logs you will find that Lukas has been occupied with other Cycles development projects for the time being (though the roadmap I posted indicates that denoising for stills may get committed to master before the year is over).
    I wonder if I should start placing the bets for when this denoiser will be abandoned before it makes it into master. After the adaptive sampling branch and other precedents, I have adopted the "I won't believe it until I see it" approach here.



  4. #764
    Member Ace Dragon's Avatar
    Join Date
    Feb 2006
    Location
    Wichita Kansas (USA)
    Posts
    28,555
    Originally Posted by MirceaKitsune View Post
    I wonder if I should start placing the bets for when this denoiser will be abandoned before it makes it into master. After the adaptive sampling branch and other precedents, I have adopted the "I won't believe it until I see it" approach here.
    Lukas reported that it is likely for the denoiser to make its way into master because he is now being paid by Tangent Animation to finish the project among other things.

    Do note that the studio is the one that is actually dropping Maya in favor of Blender (redirecting costs that would otherwise go to licenses to Blender development).
    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



  5. #765
    Member MirceaKitsune's Avatar
    Join Date
    May 2011
    Location
    Romania, Bucharest
    Posts
    559
    Originally Posted by Ace Dragon View Post
    Lukas reported that it is likely for the denoiser to make its way into master because he is now being paid by Tangent Animation to finish the project among other things.

    Do note that the studio is the one that is actually dropping Maya in favor of Blender (redirecting costs that would otherwise go to licenses to Blender development).
    Oh... I didn't know that actually. This is definitely nice to hear, and it totally gives more hope for it to be finished!



  6. #766
    Member
    Join Date
    Mar 2014
    Location
    Germany
    Posts
    345
    No, I did not forget anybody
    Yes, it is still intended to go into master soonish.

    Oh, and @Ace: Theory Animation, not Tangent (Yes, I used to mix them up as well)



  7. #767
    Originally Posted by Ace Dragon View Post
    ... he is now being paid by Tangent Animation to finish the project among other things.
    Originally Posted by lukasstockner97 View Post
    No, I did not forget anybody
    Yes, it is still intended to go into master soonish.
    These are really very good news!



  8. #768
    @lukasstockner97 You're awesome!



  9. #769
    2d denoise with use of openCV and 5 lines of python code : https://plus.google.com/+PrinioNikos/posts/6kcK5BGMkBh
    Maybe it can be of use for development of the current beta denoiser

    the actual openCV denoiser as in opencv documentation http://lumiguide.github.io/haskell-o...nCV-Photo.html
    fastNlMeansDenoisingColoredImg.png



  10. #770
    Member
    Join Date
    Mar 2014
    Location
    Germany
    Posts
    345
    Yep, I've seen that - it's just a classic NLM filter, nothing fancy

    However, NLM can indeed be used in combination with the first-order regression filtering that my branch is based on, and in fact I've been working on that for the last few days (NLM weights are one of the three improvements from the recent Disney paper). For scenes that are already pretty clean, the difference isn't really noticeable, but for quick noisy previews they can help a lot - so stay tuned

    Oh, and I think I also found the underlying issue that caused the black spots on Windows: Denormals. The Singular value decomposition is implemented using Jacobi rotations, and it turns out that the implementation used in the reference code (and therefore also in my code) depends on denormals being disabled (by setting them to zero, aka flush-to-zero).
    On Linux (and OSX iirc), Blender is built with -ffast-math, which enables flush-to-zero. On Windows, we use /fp:fast, but that doesn't enable flush-to-zero afaics. So, by enabling it manually from the code, the issue should finally be gone.
    Last edited by lukasstockner97; 14-Nov-16 at 09:04.



  11. #771
    Member 3pointEdit's Avatar
    Join Date
    Sep 2005
    Location
    Brisbane, Australia
    Posts
    8,741
    So no chance of breaking out NLM as a stand alone node for the compositor to correct non 3D media?

    Great news on the black spot resolution, thank you.
    Cheers, David ___________"awesome in space and other places".
    ________Follow me on twitter @3pointedit
    _____________ Check the Youtube Channel for tutorials. Or my Sketchbook



  12. #772
    Member
    Join Date
    Mar 2014
    Location
    Germany
    Posts
    345
    That's possible as well, but I don't have a lot of experience with the compositor code so far...
    The NLM algorithm is pretty easy - basically, if you want to filter a pixel, you loop over all neighbor pixels. For each of them, you compare two small patches of pixels, one around the neighbor pixel and one around the center pixel. Based on the patch difference, you calculate a weight for the neighbor pixel. Then, by calculating the weighted average of all the neighbor pixels, you get the NLM-filtered center pixel.



  13. #773
    @lukas thanks for your answer

    @3pointedit not sure might happen some day maybe but an alternate way might be to import opencv under blender python.
    Besides lukas work i'm also interested into figuring out how to use opencv inside blender.



  14. #774
    Member YAFU's Avatar
    Join Date
    Mar 2013
    Posts
    3,031
    My experience is, any common noise reduction method applied to a single image eliminates a lot of detail, so it makes the image very blurry. You can notice that in any of those examples.
    For this to work, it must somehow be able to distinguish what is noise, and what is small detail. I have no idea of technical issues, but I think the options are to have it analyze and compare multiple consecutive frames like filters in VapourSynth thread does (render with random noise pattern is better for this I guess). Or something that works at render level, I guess some of that is what makes Lukas denoiser implementation. Really I have no idea, but I guess it can somehow detect which pixels are noise or fireflyes to know where apply noise reduction (inpaint or whatever)
    Last edited by YAFU; 14-Nov-16 at 11:08.
    Be patient, English is not my language.



  15. #775
    lukasstockner97: I implemented the Bilateral Blur years ago, the one that is in blender. I'm not sure if and how much ppl use it. The important part was I enabled a second input for 'Determinator' basically a normalmap + zbuffer or anything that inputs the information for the denoiser. Then if ppl want to filter 2d image(why would anybody want to do that in compositor..?) they can input it in both inputs. If ppl want to use some mix of passes from render they can. At the time I implemented it, it was a weak try for a denoiser...
    I still however don't see a point in this. I guess your patch works on sample level, and thus is far far better in every aspect than any possible postprocessing operation. In conclusion, I guess it's not logical to loose time with that.



  16. #776
    Member
    Join Date
    Mar 2014
    Location
    Germany
    Posts
    345
    Yes, a similar thing can be done (and is already done in my code) for the NLM filter - basically, you can use a separate image for calculating the patch differences. By splitting the samples in two halves (and therefore two images) and using each image for the weights to filter the other one, a lot of correlation can be avoided.

    My patch basically works by 1. producing a lot of extra "feature passes" like normal, Z, surface color etc. and then 2. running a first-order linear regression on that data (essentially, modeling the color as a function of the features). In order to preserve the details that aren't explained by features (like glossy highlights), you need to weight the pixels in the least-squares solver - and these weights are what's generated using NLM currently.
    That could actually be done in a compositor node as wel, but it'd 1. require a lot of inputs and 2. be slower (no CUDA etc.)



  17. #777
    Member Ace Dragon's Avatar
    Join Date
    Feb 2006
    Location
    Wichita Kansas (USA)
    Posts
    28,555
    Regarding Bilateral Blur, it happens to be one of those filters that can easily give you a trade to low-frequency noise unless it's almost gone to begin with. Another thing is that the lack of a reflected/refracted normal and Z pass also means it will blur the image a lot in areas where the detail is seen through such effects.

    And in case people don't see the point of Lukas' denoiser, it works at rendertime and is capable of smoothing a moderately noisy result while preserving all of the detail (especially when the half-window parameter is on higher values). When complete, it should be the same story for reflected and refracted regions of the image as well (and in terms of results makes bilateral blur seem like some cheap trick).
    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



  18. #778
    Originally Posted by Ace Dragon View Post
    And in case people don't see the point of Lukas' denoiser, it works at rendertime and is capable of smoothing a moderately noisy result while preserving all of the detail (especially when the half-window parameter is on higher values). When complete, it should be the same story for reflected and refracted regions of the image as well (and in terms of results makes bilateral blur seem like some cheap trick).
    I meant I don't see the point in doing extra work on a compositor node. However I'm not sure if you actually read my post, since your reply kind of repeats what I said.



  19. #779
    Member Ace Dragon's Avatar
    Join Date
    Feb 2006
    Location
    Wichita Kansas (USA)
    Posts
    28,555
    Also to Lukas; I am wondering now, since you have managed to fix the Windows NaN bug, when we can expect a new build with the latest enhancements that you have been working on (I know the Cycles roadmap indicates denoising for stills possibly being committed this year yet, but it's been a while since I've seen activity for the denoising branch in the commit logs)?
    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



  20. #780
    Member lsscpp's Avatar
    Join Date
    May 2006
    Location
    Firenze
    Posts
    2,069
    Hi Lukas, could you explain how adaptive sampling will work using denoiser' data?
    Everything's relative. Even saying "Everything's relative".



Page 39 of 90 FirstFirst ... 2937383940414989 ... 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
  •