sampling is buggy in 2.77 #48418

New Issue

john hannovick · 2016-05-13T09:41:52+02:00

john hannovick commented

2016-05-13 09:41:52 +02:00

System Information
Win7 x64

Blender Version
Broken: 2.77a
Worked: 2.76

Short description of error
Low sample (16 and lower) images are noisy on edges

Exact steps for others to reproduce the error
bug sampling edges 2.77.blend
Open attached file in 2.76 and 2.77a, compare you get this in 2.76

and this in 2.77a

**System Information** Win7 x64 **Blender Version** Broken: 2.77a Worked: 2.76 **Short description of error** Low sample (16 and lower) images are noisy on edges **Exact steps for others to reproduce the error** [bug sampling edges 2.77.blend](https://archive.blender.org/developer/F310501/bug_sampling_edges_2.77.blend) Open attached file in 2.76 and 2.77a, compare you get this in 2.76 ![no bug 2.76.png](https://archive.blender.org/developer/F310502/no_bug_2.76.png) and this in 2.77a ![bug 2.77.png](https://archive.blender.org/developer/F310504/bug_2.77.png)

john hannovick commented

2016-05-13 09:41:52 +02:00

Changed status to: 'Open'

john hannovick commented

2016-05-13 09:41:52 +02:00

Added subscriber: @john_hannovick

NikoLeopold commented

2016-05-16 22:40:27 +02:00

Added subscriber: @NikoLeopold

Sergey Sharybin commented

2016-05-17 10:05:20 +02:00

Added subscriber: @Sergey

Sergey Sharybin commented

2016-05-17 10:05:20 +02:00

Can't reproduce the issue. Some questions:

Does it happen with latest builds from builder.blender.org?
Does it happen with both Gaussian and Blackman-Harris filters?

Attaching system-info.txt generated with Help -> Save System Info would be handy too.

Can't reproduce the issue. Some questions: - Does it happen with latest builds from builder.blender.org? - Does it happen with both Gaussian and Blackman-Harris filters? Attaching system-info.txt generated with `Help -> Save System Info` would be handy too.

mathieu menuet commented

2016-05-19 09:06:38 +02:00

Added subscriber: @bliblubli

mathieu menuet commented

2016-05-19 09:06:38 +02:00

I can confirm the bug. But the attached file with 16 samples make it less clear, at 4 sample it is obvious that edges are very noisy compared to 2.76b. Note that you may have to zoom on highDPI screen to see the difference.
Changing from Blackman-Harris to Gaussian doesn't solve the problem.
2.77:

2.76:

I can confirm the bug. But the attached file with 16 samples make it less clear, at 4 sample it is obvious that edges are very noisy compared to 2.76b. Note that you may have to zoom on highDPI screen to see the difference. Changing from Blackman-Harris to Gaussian doesn't solve the problem. 2.77: ![2.77.PNG](https://archive.blender.org/developer/F313581/2.77.PNG) 2.76: ![2.76.PNG](https://archive.blender.org/developer/F313582/2.76.PNG)

Sergey Sharybin commented

2016-05-19 09:41:27 +02:00

@bliblubli, what's your OS, compute device and system-info.txt?

mathieu menuet commented

2016-05-19 10:09:43 +02:00

Win 7 x64, CPU and the sys-info:
system-info.txt

Win 7 x64, CPU and the sys-info: [system-info.txt](https://archive.blender.org/developer/F313587/system-info.txt)

john hannovick commented

2016-05-21 14:21:11 +02:00

Tested it on many PC with windows:

It happens with blackmanharris and gaussian
It happens with latest build
It happens on CPU and GPU

Bliblubli is true. It's easier to see the problem at 4 spp than at 16.
The noise is most visible on edges, but it also affect faces. It's a pretty big regression actually regarding noise.

Tested it on many PC with windows: - It happens with blackmanharris and gaussian - It happens with latest build - It happens on CPU and GPU Bliblubli is true. It's easier to see the problem at 4 spp than at 16. The noise is most visible on edges, but it also affect faces. It's a pretty big regression actually regarding noise.

mathieu menuet commented

2016-05-22 16:09:14 +02:00

After a lot of reverts and checkout of old revisions, I nailed the problem. https://developer.blender.org/D2021 fixes the regression.

mathieu menuet commented

2016-05-22 16:13:00 +02:00

Of course, it makes rendering with HDR noisier. A better fix would be to detect "generated sky" and "HDR" to activate MIS only for HDR.

mathieu menuet commented

2016-05-22 16:15:37 +02:00

Another point is: Why does MIS break edges antialiasing?

With 1 sample:
Without MIS (pre 2.76.10)

With MIS (after 2.76.10)

Another point is: Why does MIS break edges antialiasing? With 1 sample: Without MIS (pre 2.76.10) ![sky.png](https://archive.blender.org/developer/F314295/sky.png) With MIS (after 2.76.10) ![sky with MIS.png](https://archive.blender.org/developer/F314296/sky_with_MIS.png)

Bastien Montagne commented

2016-06-01 16:11:51 +02:00

Added subscriber: @mont29

Lukas Stockner commented

2016-06-02 10:05:08 +02:00

Added subscriber: @LukasStockner

Lukas Stockner commented

2016-06-02 10:05:08 +02:00

As far as I can see, this is no bug, but simply how path tracing and World MIS works:
If you have no world MIS enabled, each path samples the lamp in the direct light sampling step amd then bounces and hits the background. Both give pretty similar results each time, so you get a clean image.
With World MIS, Cycles now sees two lights to sample (the lamp and the World) and will choose randomly between the two, producing darker pixels when it sampled the World.
It's unfortunate, but that's just how the algorithms work, no bug happening there.
So, how to "fix" this? The easiest option would be to use branched path tracing and enabling sampling of all lamps.

As far as I can see, this is no bug, but simply how path tracing and World MIS works: If you have no world MIS enabled, each path samples the lamp in the direct light sampling step amd then bounces and hits the background. Both give pretty similar results each time, so you get a clean image. With World MIS, Cycles now sees two lights to sample (the lamp and the World) and will choose randomly between the two, producing darker pixels when it sampled the World. It's unfortunate, but that's just how the algorithms work, no bug happening there. So, how to "fix" this? The easiest option would be to use branched path tracing and enabling sampling of all lamps.

mathieu menuet commented

2016-06-13 10:07:58 +02:00

I agree it's how MIS works, but thing is it should be smartly activated. Just changing default like that make some old files look suddenly really noisy without the user knowing why.
There are already conditions added by Thomas to deactivate MIS in the code for uniform color background, adding the generated sky exception would do the job.

I agree it's how MIS works, but thing is it should be smartly activated. Just changing default like that make some old files look suddenly really noisy without the user knowing why. There are already conditions added by Thomas to deactivate MIS in the code for uniform color background, adding the generated sky exception would do the job.

Sergey Sharybin commented

2016-06-13 10:58:38 +02:00

This is a documented change in 2.77 - [x]. There's also a do-versions code which should keep files saved in previous releases to render exactly the same way. If there's file which was saved in previous release and renders differently in new one -- this is to be reported and fixed.

Excluding generated sky from MIS is not really correct, there might still be strong highlights on it.

https://wiki.blender.org/index.php/Dev:Ref/Release_Notes/2.77/Cycles#Settings

This is a documented change in 2.77 - [x]. There's also a do-versions code which should keep files saved in previous releases to render exactly the same way. If there's file which was saved in previous release and renders differently in new one -- this is to be reported and fixed. Excluding generated sky from MIS is not really correct, there might still be strong highlights on it. - [x] https://wiki.blender.org/index.php/Dev:Ref/Release_Notes/2.77/Cycles#Settings

Sergey Sharybin commented

2016-07-27 10:25:50 +02:00

Changed status from 'Open' to: 'Archived'

Sergey Sharybin closed this issue

2016-07-27 10:25:50 +02:00

Sergey Sharybin self-assigned this 2016-07-27 10:25:50 +02:00

Sergey Sharybin commented

2016-07-27 10:25:50 +02:00

Wouldn't really consider this a bug. Just for a simple scenes you can disable MIS. And defaults we'd better be enabling based on heavy production shots.

Thanks for the report anyway.

Wouldn't really consider this a bug. Just for a simple scenes you can disable MIS. And defaults we'd better be enabling based on heavy production shots. Thanks for the report anyway.

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

sampling is buggy in 2.77 #48418