Page 18 of 23

Re: A Kitchen. Why to prefer Metropolis for openCL

Posted: Wed Jul 03, 2019 6:20 am
by FarbigeWelt
kintuX wrote:
Tue Jul 02, 2019 9:09 pm
Yay... yup, seeing same deterioration in quality.
Pardon? I have only pointed towards current Sobol implementation. There is obvious something wrong with it. A good sampler delivers stochastic distribution of pixels but these horizontal tendencies are definitely not.
In contrast to your comment I see enormous improvements in quality and speed. This is nothing natural but expertise and excellence of our developers and the good technical feedback of testers.

Comparison of Metropolis Hybrid with BiDir

Posted: Wed Jul 03, 2019 10:54 pm
by FarbigeWelt
CPU BiDir is much slower than Hybrid. In fact 1000 samples took BiDir 5h41m compared to 14m45s for Hybrid. From the speed view BiDir is definitively out of scope. Maybe there are some advantages left if we look at the picture. But first one has to see the BiDir image is less sharper than Hybrid. Next there is this strange green overall tone. I am not sure about the reason. Maybe it is indirect light from a projected image outside the view. The glasses look oddly dark with BiDir, Hybrid glasses look more natural. Light cone intersection at the wall in the back look different. BiDir renders some additional caustics at the ceiling. One nice BiDir thing is are the caustics of the bottles. This last thing is the one of the only missing part in Hybrid picture.
Comparison of Metropolis Hybrid with BiDir, 1000 Samples, 1HD
Comparison of Metropolis Hybrid with BiDir, 1000 Samples, 1HD
Conclusion
Hybrid only requires some further caustics improvements and then there is obviously no meaning left for use of BiDir.

Re: A Kitchen. Why to prefer Metropolis for openCL

Posted: Thu Jul 04, 2019 2:30 pm
by kintuX
FarbigeWelt wrote:
Wed Jul 03, 2019 6:20 am
kintuX wrote:
Tue Jul 02, 2019 9:09 pm
Yay... yup, seeing same deterioration in quality.
Pardon? I have only pointed towards current Sobol implementation. There is obvious something wrong with it. A good sampler delivers stochastic distribution of pixels but these horizontal tendencies are definitely not.
In contrast to your comment I see enormous improvements in quality and speed. This is nothing natural but expertise and excellence of our developers and the good technical feedback of testers.
OK, so i expressed poorly... but regression is there, destroying the quality. Let's not get too biased ;)

Re: Comparison of Metropolis Hybrid with BiDir

Posted: Thu Jul 04, 2019 4:57 pm
by epilectrolytics
Good job with the comparison.
I'm baffled that there are such visible differences!
FarbigeWelt wrote:
Wed Jul 03, 2019 10:54 pm
Conclusion
Hybrid only requires some further caustics improvements and then there is obviously no meaning left for use of BiDir.
I don't think hybrid is meant to replace BiDir, instead it should accompany PGI for speedy caustics instead of caustic cache which became just too complicated.

Once BiDirVM is finished, that might be the ultimate replacement for classic BiDir as an unbiased engine with no blind spot for certain path categories.

Artificial Eye

Posted: Sun Jul 07, 2019 2:59 am
by FarbigeWelt
Artificial Eye sees Cube and mirrored Laser, openCL Hybrid, 500 samples
Artificial Eye sees Cube and mirrored Laser, openCL Hybrid, 500 samples
What can you see?
A red green checkered cube, a mirrored image of a little part of the cube, mist and a laser beam and both mirrored image.
Anything else?
Yes, the image you can see is the image of a lens thrown at a thin layer of matte translucent.
Pardon? You don't believe me? Look at the scene setup.
Screenshot  scene setup views from different angles.
Screenshot scene setup views from different angles.
Credits: Thanks to CodeHD's little lens add-on I was able to define the artificial eye.

Re: Different Works in Progress

Posted: Sun Jul 07, 2019 9:41 am
by CodeHD
Oh my it took me ages to understand exactly what is what in this image and how it corresponds to the scene objects :lol: :roll:

Nice to see the addon is being used :) I'll have to expand it a bit more soon, could use some extra features myself...

Rainbow Chains V0.9, the addon version

Posted: Sun Jul 07, 2019 2:44 pm
by FarbigeWelt
Since my last post of the last scripts' version I learned again a lot with coding, debugging, looking up commands' use or script examples. It is a lot of fun for me playing with the parameter sets of each of the four patterns. To achieve a simple thing like the color ramp from the lowest to highest object z-position I had to split the main part, the two loops, into two parts to be able to calculated shading according to z-positions. And because for one pattern stepping with natural numbers was not fine enough I had to replace for-im loops by while loops. Also I replaced if: else: if: by if: elif: because it is shorter and easier to read. I just did not know elif during my the first steps. Meanwhile the script can be installed as add-on from the attached zip folder, enabled, disabled and removed. As long as no object is removed from the script generated object array the script should run reliable. :mrgreen:

If some one likes to add its own patterns read and follow the next lines.
There are 4 pattern presets. One can add further patterns with code adjustments.
To add further patterns extend the code close to each of the following FIVE comments.
  • Add your own pattern name above between
    Add your own pattern parameters above between
    Add your own pattern method below
    Add your own pattern selection below
    Add your own pattern method call below
Search for above comments in the code and extent as descripted.
Do not remove the comment lines for obvious reasons.
Rainbow Chains V0.9.blend.zip
Rainbow Chains V0.9
(20.97 KiB) Downloaded 39 times

Rainbow Chains, the four presets, animtaed

Posted: Sun Jul 07, 2019 4:24 pm
by FarbigeWelt
Four animated GIFs of the four presets of the previously posted Blender add-on.
Cloud
Cloud
SinCos
SinCos
Gauss
Gauss
Boid
Boid

Eagle statue made from stone

Posted: Fri Jul 12, 2019 6:13 am
by FarbigeWelt
Thanks to provisory I found this nice model of an eagle on the page of [http://threedscans.com/info/]Three D Scans Info[/url]
The stone looks not finished yet but is close to the look I like to get.
Eagle, openCL hybrid, 2000 samples OIDN 0.95
Eagle, openCL hybrid, 2000 samples OIDN 0.95
An area light is responsible for the bright light. This light is also responsible for a rather noisy render, especially below one of eagle's wing. I guess the scene require more to 5000 samples to be clean without denoiser.

Re: A Kitchen. Why to prefer Metropolis for openCL

Posted: Sun Jul 14, 2019 6:30 pm
by epilectrolytics
FarbigeWelt wrote:
Tue Jul 02, 2019 5:54 pm
Sobol as implemented currently produces scan line similar patterns.

As assumed OIDN is not any fan of this Sobol-Lines too. OIDN develops more or less creative patterns instead of denoising the image.

I switch to Metropolis: well distributed noise patterns. Even with as low as 250 samples, OIDN will have the perfect meal.
Thanks very much for your systematic testing!
When I ran into the same problem I remembered your post, switched from Sobol to Metropolis and got OIDIN to work again.
Then I tried the random sampler and it worked as well!
Seems the scan lines are a Sobol-only issue?

Also I discovered that the problems OIDN has with volume scattering are solved by disabling Sobol.