sobol gives nonuniform noise #41143

New Issue

lopata · 2014-07-22T13:51:30+02:00

lopata commented

2014-07-22 13:51:30 +02:00

System Information
win7 x64

Blender Version
blender-2.71-2e43617-win64.zip 87M Tue Jul 22 02:19:39 2014

I do not know, this is bug or feature of algorithm, but in this scene on left side always more lighted. Position of light source into box is irrelevant. Multijitter pattern works normal

sobol_bug.blend

**System Information** win7 x64 **Blender Version** blender-2.71-2e43617-win64.zip 87M Tue Jul 22 02:19:39 2014 I do not know, this is bug or feature of algorithm, but in this scene on left side always more lighted. Position of light source into box is irrelevant. Multijitter pattern works normal ![sobol.jpg](https://archive.blender.org/developer/F99054/sobol.jpg) ![multijitter.jpg](https://archive.blender.org/developer/F99056/multijitter.jpg) [sobol_bug.blend](https://archive.blender.org/developer/F99058/sobol_bug.blend)

lopata commented

2014-07-22 13:51:30 +02:00

Changed status to: 'Open'

lopata commented

2014-07-22 13:51:30 +02:00

Added subscriber: @lopataasdf

Bastien Montagne commented

2014-07-22 14:04:31 +02:00

Added subscriber: @mont29

Sergey Sharybin self-assigned this 2014-07-22 17:04:36 +02:00

Brecht Van Lommel commented

2014-07-26 12:20:49 +02:00

Added subscriber: @brecht

Brecht Van Lommel commented

2014-07-26 12:20:49 +02:00

This is a known issue, the Sobol pattern currently has some correlation between pixels in some cases. If you take more samples both will converge to the right result, I would not consider this a bug at the moment.

Perhaps this could be improved by making it so the Cranley-Patterson rotation is different for different dimensions, so you would do something like cmj_hash(*rng, dimension/2) and use that instead of *rng in path_rng_1D. I've tried that before to solve a correlation issue in another scene but it didn't help then, maybe in this case it does.

Sobol is using the same sequence in every pixel, just randomized a bit. Correlated Multi Jitter has a different sequence in every pixel which avoids correlation issues.

This is a known issue, the Sobol pattern currently has some correlation between pixels in some cases. If you take more samples both will converge to the right result, I would not consider this a bug at the moment. Perhaps this could be improved by making it so the Cranley-Patterson rotation is different for different dimensions, so you would do something like `cmj_hash(*rng, dimension/2)` and use that instead of `*rng` in `path_rng_1D`. I've tried that before to solve a correlation issue in another scene but it didn't help then, maybe in this case it does. Sobol is using the same sequence in every pixel, just randomized a bit. Correlated Multi Jitter has a different sequence in every pixel which avoids correlation issues.

Sergey Sharybin commented

2014-08-12 15:48:48 +02:00

Changed status from 'Open' to: 'Archived'

Sergey Sharybin closed this issue

2014-08-12 15:48:48 +02:00

Sergey Sharybin commented

2014-08-12 15:48:48 +02:00

@brecht, using hashed rng seems to help in this particular case, that even doesn't seem to introduce huge amount of slowdown.

But since it didn't work for the scene you've been investigating a while ago, this isn't a reliable solution, unfortunately. So would think of making CMJ more production ready (move it out from experimental, enable for GPU, likely optimize) is the way to go.

So thanks for the report, but to the TODO it goes. Stay tuned!

@brecht, using hashed rng seems to help in this particular case, that even doesn't seem to introduce huge amount of slowdown. But since it didn't work for the scene you've been investigating a while ago, this isn't a reliable solution, unfortunately. So would think of making CMJ more production ready (move it out from experimental, enable for GPU, likely optimize) is the way to go. So thanks for the report, but to the TODO it goes. Stay tuned!

blender-admin commented

2016-11-26 17:30:02 +01:00

This issue was referenced by blender/cycles@6bcca3d526

This issue was referenced by blender/cycles@6bcca3d5268ae41931cbe47840018a57f22fdc05

blender-admin commented

2016-11-26 17:30:02 +01:00

This issue was referenced by 9d50175b6c

This issue was referenced by 9d50175b6cc5860926f0974fafd9e8e7868eb805

Sergey Sharybin commented

2016-12-01 14:20:11 +01:00

Changed status from 'Archived' to: 'Resolved'

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

sobol gives nonuniform noise #41143