Composited cycles render passes not matching combined output #103224

New Issue

Peter Goodman · 2022-12-14T19:40:08+01:00

Peter Goodman commented

2022-12-14 19:40:08 +01:00

System Information
Operating system: Microsoft Windows [Version 10.0.19042.2251]
Graphics card: Nvidia Quadro RTX 4000 (driver 512.36)

Blender Version
Broken: 3.3.1 and 3.4.0
Worked: Unknown

Short description of error
Carrying out manual composition of the Cycles render passes as described in the Blender documentation is not resulting in the same result as the combined output.

Fig 1 shows the blender manual description of how passes are combined for the final output (edited by me with how I think the volume passes are included - the documentation doesn't explain this).

Fig 2 shows my compositor node tree that matches that description.

Fig 3 is the compositor output...

...and Fig 4 is the combined noisy image output. If you click on Fig 3 or 4 to expan the image view then use the left/right arrows on your keyboard you can flick between these two images to clearly see the difference.

Neither are denoised to eliminate that as a contributory factor (I used 4096 samples to keep noise low). Fig 5 highlights the regions where I'm seeing significant shading differences...

Figs 6,7,8 are closer views of these regions.

The key issues (shown in Fig 5) are the missing shading around the sphere rim, and missing shading on the floor panels at the larger rib regions between the panels. The missing shading around the sphere rim is exacerbated when denoising is applied and those notches become more pronounced. All other regions appear to be the same to me.

A potential clue: the issue might be originating with the DiffDir and/or the DiffInd passes, because the sphere rim highlight 'notches' and the darker regions on the floor panels are both visible when those two passes are multiplied in the compositor node tree (see Fig 9 showing viewer output from that multiply node).

Exact steps for others to reproduce the error
Load the attached blend file and render in cycles (caution - its currently set to 4096 samples... you dont need to render it with as many samples to see the issue!). Compare the 'noisy image' output and the compositor output.
cycles_compositor_combine.blend

**System Information** Operating system: Microsoft Windows [Version 10.0.19042.2251] Graphics card: Nvidia Quadro RTX 4000 (driver 512.36) **Blender Version** Broken: 3.3.1 and 3.4.0 Worked: Unknown **Short description of error** Carrying out manual composition of the Cycles render passes as described in the Blender documentation is not resulting in the same result as the combined output. Fig 1 shows the blender manual description of how passes are combined for the final output (edited by me with how I think the volume passes are included - the documentation doesn't explain this). ![fig1_blender_manual.PNG](https://archive.blender.org/developer/F14057767/fig1_blender_manual.PNG) Fig 2 shows my compositor node tree that matches that description. ![fig2_compositor_nodetree.PNG](https://archive.blender.org/developer/F14057769/fig2_compositor_nodetree.PNG) Fig 3 is the compositor output... ![fig3_composite_4096samples.png](https://archive.blender.org/developer/F14057771/fig3_composite_4096samples.png) ...and Fig 4 is the combined noisy image output. If you click on Fig 3 or 4 to expan the image view then use the left/right arrows on your keyboard you can flick between these two images to clearly see the difference. ![fig4_noisyimage_4096samples.png](https://archive.blender.org/developer/F14057775/fig4_noisyimage_4096samples.png) Neither are denoised to eliminate that as a contributory factor (I used 4096 samples to keep noise low). Fig 5 highlights the regions where I'm seeing significant shading differences... ![fig5_significant_difference_regions.png](https://archive.blender.org/developer/F14057777/fig5_significant_difference_regions.png) Figs 6,7,8 are closer views of these regions. ![fig6_sphere_comparison.PNG](https://archive.blender.org/developer/F14057779/fig6_sphere_comparison.PNG) ![fig7_sphere_rim_highlight_problems.PNG](https://archive.blender.org/developer/F14057781/fig7_sphere_rim_highlight_problems.PNG) ![fig8_missing_shading_on_floor.PNG](https://archive.blender.org/developer/F14057783/fig8_missing_shading_on_floor.PNG) The key issues (shown in Fig 5) are the missing shading around the sphere rim, and missing shading on the floor panels at the larger rib regions between the panels. The missing shading around the sphere rim is exacerbated when denoising is applied and those notches become more pronounced. All other regions appear to be the same to me. A potential clue: the issue might be originating with the DiffDir and/or the DiffInd passes, because the sphere rim highlight 'notches' and the darker regions on the floor panels are both visible when those two passes are multiplied in the compositor node tree (see Fig 9 showing viewer output from that multiply node). ![fig9_multipled_DiffDir_and_DiffInd.png](https://archive.blender.org/developer/F14057785/fig9_multipled_DiffDir_and_DiffInd.png) **Exact steps for others to reproduce the error** Load the attached blend file and render in cycles (caution - its currently set to 4096 samples... you dont need to render it with as many samples to see the issue!). Compare the 'noisy image' output and the compositor output. [cycles_compositor_combine.blend](https://archive.blender.org/developer/F14057793/cycles_compositor_combine.blend)

Peter Goodman commented

2022-12-14 19:40:08 +01:00

Added subscriber: @badbunny_uk

Peter Goodman changed title from ~~Composited render passes not matching combined output~~ to Composited cycles render passes not matching combined output

2022-12-14 19:41:04 +01:00

Germano Cavalcante commented

2022-12-15 19:02:11 +01:00

Added subscriber: @mano-wii

Germano Cavalcante commented

2022-12-15 19:02:11 +01:00

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

Germano Cavalcante commented

2022-12-15 19:02:11 +01:00

The file is quite complex.
One of the materials even has an invalid node!

But I can confirm the difference in the results and it seems to be related to the volume.

The lack of documentation about composing volume passes and this TO DO about it make me doubt if this is really a bug or some limitation.
#91290 (Improve volume render passes)

The file is quite complex. One of the materials even has an invalid node! ![image.png](https://archive.blender.org/developer/F14060064/image.png) But I can confirm the difference in the results and it seems to be related to the volume. The lack of documentation about composing volume passes and this `TO DO` about it make me doubt if this is really a bug or some limitation. #91290 (Improve volume render passes)

Peter Goodman commented

2022-12-15 20:15:49 +01:00

Hi, those materials are from a procedural material pack off blendermarket, I had looked inside a few of them but I confess they are beyond my current level of understanding.

To eliminate those materials from the equation, I've changed both the floor and sphere materials to principled BSDFs:
cycles_compositor_combine2.blend

The edge highlighting artefact is no longer obvious, but the sphere shading brightness is significantly different between the noisy image (brighter) and compositor (darker), but everywhere else seems almost identical. See attached images comparing the noisy image output (fig 10)...

.... to the compositor (fig 11) - again if you expand an image and left/right keyboard arrow flick between them to see the difference, again both not denoised (I was impatient and reduced the samples from the previous 4096 to 1024 for these!)...

If I remove the volume object, then the noisy image and compositor outputs look identical. So I think you are right that it's related to the volumetric shading passes, and possibly how they interact with transparent objects? I'm assuming maybe some component of the overall shading is not being allocated to any passes when they are all divided up?

So is it currently not possible to match the compositor to the combined output when volumetric shading is involved, and using the compositor for volumetrics is potentially vulnerable to visual artefacts? If this turns out to be the case, is there a workaround that could minimise the impact?

Also it would be really useful if the documentation could be amended to include the developer's thoughts on the best practice for compositing the volumetric passes, as well as to warn of this issue/limitation (again if this turns out to be the case).

Thanks for looking at my file, Pete.

Hi, those materials are from a procedural material pack off blendermarket, I had looked inside a few of them but I confess they are beyond my current level of understanding. To eliminate those materials from the equation, I've changed both the floor and sphere materials to principled BSDFs: [cycles_compositor_combine2.blend](https://archive.blender.org/developer/F14060187/cycles_compositor_combine2.blend) The edge highlighting artefact is no longer obvious, but the sphere shading brightness is significantly different between the noisy image (brighter) and compositor (darker), but everywhere else seems almost identical. See attached images comparing the noisy image output (fig 10)... ![fig11_principled_noisy_image.png](https://archive.blender.org/developer/F14060188/fig11_principled_noisy_image.png) .... to the compositor (fig 11) - again if you expand an image and left/right keyboard arrow flick between them to see the difference, again both not denoised (I was impatient and reduced the samples from the previous 4096 to 1024 for these!)... ![fig12_principled_composite.png](https://archive.blender.org/developer/F14060190/fig12_principled_composite.png) If I remove the volume object, then the noisy image and compositor outputs look identical. So I think you are right that it's related to the volumetric shading passes, and possibly how they interact with transparent objects? I'm assuming maybe some component of the overall shading is not being allocated to any passes when they are all divided up? So is it currently not possible to match the compositor to the combined output when volumetric shading is involved, and using the compositor for volumetrics is potentially vulnerable to visual artefacts? If this turns out to be the case, is there a workaround that could minimise the impact? Also it would be really useful if the documentation could be amended to include the developer's thoughts on the best practice for compositing the volumetric passes, as well as to warn of this issue/limitation (again if this turns out to be the case). Thanks for looking at my file, Pete.

Philipp Oeser removed the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Composited cycles render passes not matching combined output #103224