Compositor: Glare Node artifact in OpenEXR file with half float color depth #84167

New Issue

炎臻 · 2020-12-27T06:10:12+01:00

炎臻 commented

2020-12-27 06:10:12 +01:00

System Information
Operating system: Windows-10-10.0.19041-SP0 64 Bits
Graphics card: GeForce GTX 980/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 460.89

Blender Version
Broken: version: 2.92.0 Alpha, branch: master, commit date: 2020-12-26 22:50, hash: 7916c59304
Worked: Unknow

Short description of error

Extremely black shows up in the result after applying Glare node.
Below is an OpenEXR file viewed in blender render result with half float color depth and Glare effect:
Glare effect worked in the purple circle and looks fine.
But the actual file which saved as PNG has artifact:

Exact steps for others to reproduce the error

{icon fa-plus} Here is a zip containing the OpenEXR file above and a reproducible blender file I built up.
{F9531547} The 4026.exr is the file which gets artifact after saving as a PNG and 0001Full Float.exr ,0001Half Float.exr are two files rendered from the "RepoduceScene.blend".
The RepoduceScene.blend: Then they three are composited in CompositingGlareNode.blend: In my practice,with OpenEXR format and half float color depth,I found the artifact occurs where the scene is extremely bright.In the reproducible scene a point light with 500MW is lighting the cube from an angle.
And if the color depth is set to full float the artifact is gone: I don't know whether this is a limitation of color depth and color space conversion or a bug.
While the repoduced EXR files gets artifact in viewer,original one(4026.exr) only gets artifact after it is saved as another format (in my case:PNG,JPEG,JPEG,BMP)),the render result viewer in blender shows fine:

**System Information** Operating system: Windows-10-10.0.19041-SP0 64 Bits Graphics card: GeForce GTX 980/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 460.89 **Blender Version** Broken: version: 2.92.0 Alpha, branch: master, commit date: 2020-12-26 22:50, hash: `7916c59304` Worked: Unknow --- ## Short description of error Extremely black shows up in the result after applying __Glare node__. Below is an OpenEXR file viewed in blender __render result__ with __half float__ color depth and Glare effect: ![Viewed from build-in render result.png](https://archive.blender.org/developer/F9531498/Viewed_from_build-in_render_result.png) Glare effect worked in the purple circle and looks fine. But the actual file which saved as PNG has artifact:![ArifactResult.png](https://archive.blender.org/developer/F9531513/ArifactResult.png) --- ## Exact steps for others to reproduce the error {icon fa-plus} Here is a zip containing the OpenEXR file above and a reproducible blender file I built up. {[F9531547](https://archive.blender.org/developer/F9531547/DataPackedup.7z)}![filelist.png](https://archive.blender.org/developer/F9531569/filelist.png) The __4026.exr__ is the file which gets artifact after saving as a PNG and __0001Full Float.exr__ ,__0001Half Float.exr__ are two files rendered from the "RepoduceScene.blend". The __RepoduceScene.blend__:![RepoduceSceneThumb.png](https://archive.blender.org/developer/F9531605/RepoduceSceneThumb.png) Then they three are composited in __CompositingGlareNode.blend__:![CompositingGlareNodeThumb.png](https://archive.blender.org/developer/F9531630/CompositingGlareNodeThumb.png) In my practice,with OpenEXR format and __half float__ color depth,I found the artifact occurs where the scene is extremely bright.In the reproducible scene a point light with 500MW is lighting the cube from an angle. And if the color depth is set to __full float__ the artifact is gone:![CompositingGlareNodeFullFloat.png](https://archive.blender.org/developer/F9531639/CompositingGlareNodeFullFloat.png) I don't know whether this is a limitation of color depth and color space conversion or a bug. While the repoduced EXR files gets artifact in viewer,original one(__4026.exr__) only gets artifact after it is saved as another format (in my case:PNG,JPEG,JPEG,BMP)),the render result viewer in blender shows fine:![CompositingGlareNodeOriginalOne.png](https://archive.blender.org/developer/F9531655/CompositingGlareNodeOriginalOne.png)

炎臻 commented

2020-12-27 06:10:12 +01:00

Added subscriber: @PrettyFireNOI7

Robert Guetzkow commented

2021-01-01 14:15:14 +01:00

Added subscribers: @Sergey, @rjg

Robert Guetzkow commented

2021-01-01 14:15:14 +01:00

Changed status from 'Needs Triage' to: 'Needs Developer To Reproduce'

Robert Guetzkow commented

2021-01-01 14:15:14 +01:00

I can reproduce the behavior. The issue appears to be that your file produces values that exceed the range of half-float which is +/- 65.504 (IEEE 754). The maximum RGB value in the 0001Full Float.exr file is 40852116.0. It seems that the floating point arithmetics result in -nan instead of nan, which produces pure black instead of white in the image.

This is ~~quite~~ an edge case. Perhaps COM_GlareStreaksOperation.cpp or parts of the pipeline that process the results could handle this better, but this may be seen different by the VFX module owner @sergey.

I can reproduce the behavior. The issue appears to be that your file produces values that exceed the range of half-float which is +/- 65.504 (IEEE 754). The maximum RGB value in the `0001Full Float.exr` file is 40852116.0. It seems that the floating point arithmetics result in -nan instead of nan, which produces pure black instead of white in the image. This is ~~quite~~ an edge case. Perhaps `COM_GlareStreaksOperation.cpp` or parts of the pipeline that process the results could handle this better, but this may be seen different by the VFX module owner @sergey.

炎臻 commented

2021-01-01 18:05:02 +01:00

In #84167#1085343, @rjg wrote:
I can reproduce the behavior. The issue appears to be that your file produces values that exceed the range of half-float which is +/- 65.504 (IEEE 754). The maximum RGB value in the 0001Full Float.exr file is 40852116.0. It seems that the floating point arithmetics result in -nan instead of nan, which produces pure black instead of white in the image.

This is quite an edge case. Perhaps COM_GlareStreaksOperation.cpp or parts of the pipeline that process the results could handle this better, but this may be seen different by the VFX module owner @sergey.

Thank you.I understand it's an edge case and I had replaced the original file with Full Float EXR to fix the black area.
While as described(file 4026.exr ),I wonder why when viewing in Blender render viewer it "looks" correct but black shows up after saving to another format(like PNG).I think it would be less confusing if the render viewer has the matching result with the saved file.

> In #84167#1085343, @rjg wrote: > I can reproduce the behavior. The issue appears to be that your file produces values that exceed the range of half-float which is +/- 65.504 (IEEE 754). The maximum RGB value in the `0001Full Float.exr` file is 40852116.0. It seems that the floating point arithmetics result in -nan instead of nan, which produces pure black instead of white in the image. > > This is quite an edge case. Perhaps `COM_GlareStreaksOperation.cpp` or parts of the pipeline that process the results could handle this better, but this may be seen different by the VFX module owner @sergey. Thank you.I understand it's an edge case and I had replaced the original file with __Full Float EXR__ to fix the black area. While as described(file __4026.exr__ ),I wonder why when viewing in Blender render viewer it "looks" correct but black shows up after saving to another format(like PNG).I think it would be less confusing if the render viewer has the matching result with the saved file.

Sergey Sharybin self-assigned this 2021-01-11 16:16:36 +01:00

Sergey Sharybin commented

2021-01-11 16:16:36 +01:00

To me it seems that root of the problem is the glare node producing nan. Compositor shouldn't give nan numbers, as they cause all sort of issues. Full float EXR might be OK for your need, but in general nan will ruing someone's day eventually.

Will check on this issue deeper in a bit.

To me it seems that root of the problem is the glare node producing nan. Compositor shouldn't give nan numbers, as they cause all sort of issues. Full float EXR might be OK for your need, but in general nan will ruing someone's day eventually. Will check on this issue deeper in a bit.

blender-admin commented

2021-01-13 14:20:26 +01:00

This issue was referenced by 53bd58993e

This issue was referenced by 53bd58993e2f6d35452242762402c20343d6eef3

Sergey Sharybin commented

2021-01-14 15:02:40 +01:00

Have a deeper look and seems that to fix the root cause here we'd need to fix the half-float EXR saving.

Fix for this is coming shortly. After the fix the "bad" EXR images are to be re-saved though.

The Glare node might have some overflow issues, but so far with the given files i didn't see Glare to give nan/inf pixels.

Have a deeper look and seems that to fix the root cause here we'd need to fix the half-float EXR saving. Fix for this is coming shortly. After the fix the "bad" EXR images are to be re-saved though. The Glare node might have some overflow issues, but so far with the given files i didn't see Glare to give nan/inf pixels.

Sergey Sharybin commented

2021-01-14 15:19:42 +01:00

Changed status from 'Needs Developer To Reproduce' to: 'Resolved'

Sergey Sharybin closed this issue

2021-01-14 15:19:42 +01:00

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

Compositor: Glare Node artifact in OpenEXR file with half float color depth #84167

Short description of error

Exact steps for others to reproduce the error