Lens Distortion compositor node artifacts #37462

New Issue

Anton Ilmyanov · 2013-11-15T01:11:54+01:00

Anton Ilmyanov commented

2013-11-15 01:11:54 +01:00

System Information
Windows 7 64

Blender Version
Broken: Blender 2.69, 2.68, 2.67

Short description of error
Lens Distortion node produces blur artifacts looking like a grid. The size of the grid changes when you adjust Dispersion and/or Distortion parameters. It's most noticeable in darker and noisy areas, especially in animations as you play them fullscreen. It doesn't completely go away with higher sampling.

Exact steps for others to reproduce the error
Render the attached .blend and zoom in.
lens-distortion-test.blend

Screenshot

**System Information** Windows 7 64 **Blender Version** Broken: Blender 2.69, 2.68, 2.67 **Short description of error** Lens Distortion node produces blur artifacts looking like a grid. The size of the grid changes when you adjust Dispersion and/or Distortion parameters. It's most noticeable in darker and noisy areas, especially in animations as you play them fullscreen. It doesn't completely go away with higher sampling. **Exact steps for others to reproduce the error** Render the attached .blend and zoom in. [lens-distortion-test.blend](https://archive.blender.org/developer/F27329/lens-distortion-test.blend) Screenshot ![lens-distortion-artifacts.png](https://archive.blender.org/developer/F27331/lens-distortion-artifacts.png)

Anton Ilmyanov commented

2013-11-15 01:11:54 +01:00

Changed status to: 'Open'

Anton Ilmyanov commented

2013-11-15 01:11:54 +01:00

Added subscriber: @AntonIlmyanov

gandalf3 commented

2013-11-15 22:16:59 +01:00

Added subscriber: @gandalf3

Brecht Van Lommel commented

2013-11-22 18:42:13 +01:00

Added subscribers: @Jeroen-Bakker, @monique

Brecht Van Lommel commented

2013-11-22 18:44:24 +01:00

Added subscriber: @brecht

Brecht Van Lommel commented

2013-11-22 18:44:24 +01:00

I can confirm the issue.

At first I thought this might have been because of the particular random number sequence cycles is using but it appears to be not the case, same happens for blender internal and other random noise.

I can confirm the issue. At first I thought this might have been because of the particular random number sequence cycles is using but it appears to be not the case, same happens for blender internal and other random noise.

Lukas Tönne was assigned by Sergey Sharybin

2013-12-04 16:28:34 +01:00

Sergey Sharybin commented

2013-12-04 16:28:34 +01:00

Added subscribers: @LukasTonne, @Sergey

Sergey Sharybin commented

2013-12-04 16:28:34 +01:00

@LukasTonne, would you have any idea here? If not, re-assign to me and i'l ltry to find a clue :)

Lukas Tönne commented

2013-12-04 17:31:44 +01:00

I can't decipher the math of that node easily. Just from the effect it has i suspect there is a problem with accumulating pixel values: depending on the subpixel coordinates it could lead some smoothing-out of small scale noisy variations, then if these offsets in relation to the integer pixel coordinates change it could lead to the observed grid pattern. But that's little more than a gut feeling ...

Would be great to have references to papers or at least some topics/keywords in such code which would allow you to get an overview of what it is supposed to do. Otherwise understanding cryptic formulas with one-letter variables and no comments is close to impossible and makes it dead code that nobody will ever be able to fix or improve ...

I can't decipher the math of that node easily. Just from the effect it has i suspect there is a problem with accumulating pixel values: depending on the subpixel coordinates it could lead some smoothing-out of small scale noisy variations, then if these offsets in relation to the integer pixel coordinates change it could lead to the observed grid pattern. But that's little more than a gut feeling ... Would be great to have references to papers or at least some topics/keywords in such code which would allow you to get an overview of what it is *supposed* to do. Otherwise understanding cryptic formulas with one-letter variables and no comments is close to impossible and makes it dead code that nobody will ever be able to fix or improve ...

Lukas Tönne commented

2014-01-03 12:06:13 +01:00

I've been investigating the ScreenLensDistortionOperation and cleaning/restructuring the code to make it more understandable. Unfortunately it seems like the problem causing this bug is rooted in the accumulation method the algorithm uses to sum up color values for distorted pixels. This is done by evaluating the input image over a range of pixels defined by a non-linear factor, which i assume causes this sort of unequal noise smoothing. I don't know enough about the supposed sampling algorithm to fix this easily.

Lukas Tönne commented

2014-01-22 15:51:20 +01:00

Compare this to artifacts you get when scaling a noise texture (open image to see the effect):

This is just white noise, scaled (no power-of-2) and nearest-filtered. On a small scale the lens distortion does just that, it scales the image and the bad filtering produces these artifacts when applied to sharp per-pixel noise contrast.

Compare this to artifacts you get when scaling a noise texture (open image to see the effect): ![screen.png](https://archive.blender.org/developer/F72315/screen.png) This is just white noise, scaled (no power-of-2) and nearest-filtered. On a small scale the lens distortion does just that, it scales the image and the bad filtering produces these artifacts when applied to sharp per-pixel noise contrast.

Natan Dotan commented

2014-05-28 02:14:09 +02:00

Added subscriber: @NAD

Natan Dotan commented

2014-05-28 03:37:39 +02:00

It looks like this might be the result of the fact that it uses bilinear interpolation in the accumulation function (line 150 in COM_ScreenLensDistortionOperation.cpp). This could be what's causing discontinuities at the block boundaries. This seems like it might explain what's happening: http://libnoise.sourceforge.net/noisegen/index.html

@LukasTonne, could you send me your cleaned version of the ScreenLensDistortionOperation code? I'm new to Blender so it would help to have something more readable. Perhaps I might be able to help.

It looks like this might be the result of the fact that it uses bilinear interpolation in the accumulation function (line 150 in COM_ScreenLensDistortionOperation.cpp). This could be what's causing discontinuities at the block boundaries. This seems like it might explain what's happening: http://libnoise.sourceforge.net/noisegen/index.html @LukasTonne, could you send me your cleaned version of the ScreenLensDistortionOperation code? I'm new to Blender so it would help to have something more readable. Perhaps I might be able to help.

Lukas Tönne commented

2014-05-28 08:25:05 +02:00

@NAD: My cleanup changes are already in master, committed as part of 9c883a1. Before this it was basically one big clunky function with redundant code, i merely split it into smaller parts to clarify what is happening.

And yes, i believe the key to this issue is the accumulation method. Seems like some kind of aliasing issue, but i'm not an expert in sampling methods, so take that with a grain of salt. Thanks for looking into it anyway, hope you manage to find a solution. Let me know if i can help.

@NAD: My cleanup changes are already in master, committed as part of 9c883a1. Before this it was basically one big clunky function with redundant code, i merely split it into smaller parts to clarify what is happening. And yes, i believe the key to this issue is the accumulation method. Seems like some kind of aliasing issue, but i'm not an expert in sampling methods, so take that with a grain of salt. Thanks for looking into it anyway, hope you manage to find a solution. Let me know if i can help.

Natan Dotan commented

2014-05-28 20:15:54 +02:00

Thanks @lukastoenne. I'll start digging in this weekend. Will report back next week.

Kévin Dietrich commented

2014-06-10 18:46:30 +02:00

Added subscriber: @kevindietrich

Lukas Tönne commented

2014-07-08 18:19:24 +02:00

Changed status from 'Open' to: 'Archived'

Lukas Tönne closed this issue

2014-07-08 18:19:24 +02:00

Lukas Tönne commented

2014-07-08 18:19:24 +02:00

No solution in sight for this, putting it on the ToDo list.

Natan Dotan commented

2014-07-08 18:32:16 +02:00

@LukasTonne: Unfortunately I've been swamped with work lately and haven't had time to get to this, but I did a quick pass. Perhaps this will be useful to whoever gets to this first:
It's worth noting that the size of the artifacts is inversely proportional to the amount of dispersion used -- so doubling dispersion halves the size of the artifacts. It seems like it might be an interpolation problem in COM_MemoryBuffer.cpp (readEWA line 255). I still think http://libnoise.sourceforge.net/noisegen/index.html might be the answer.

@LukasTonne: Unfortunately I've been swamped with work lately and haven't had time to get to this, but I did a quick pass. Perhaps this will be useful to whoever gets to this first: It's worth noting that the size of the artifacts is inversely proportional to the amount of dispersion used -- so doubling dispersion halves the size of the artifacts. It seems like it might be an interpolation problem in COM_MemoryBuffer.cpp (readEWA line 255). I still think http://libnoise.sourceforge.net/noisegen/index.html might be the answer.

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

Lens Distortion compositor node artifacts #37462