Page MenuHome

3.0 alpha: Cycles viewport: Adaptive Sampling with Denoise degrades as samples increase
Closed, ArchivedPublic

Description

System Information
Linux Mint 20.2 Cinnamon
3400G

Blender Version
Broken: blender-3.0.0-alpha+master.e81533b25e6b-linux.x86_64-release.tar.xz
Worked: blender-2.93.4-linux-x64.tar.xz

Short description of error
For difficult to light scenes, Adaptive Sampling with Denoise begins clean, then counter-intuitively degrades as samples increase.

Exact steps for others to reproduce the error


Open the above .blend file and switch the Viewport to Rendered.

48 samples

96 samples

256 samples

1024 samples

Compare 256 samples sans Adaptive Sampling

Notes:
To help highlight the issue, this scene is intended to be difficult to light.

Testing a Noise Threshold of .01 (same as Render) helps, but only by delaying the issue.

For faster CPU renders, Viewport Pixel Size 2x was used.

Event Timeline

I know it's counterintuitive, but I do not consider this a bug at this point.

Due to how noisy this scene is, it needs a much higher minimum number of samples. Otherwise, there will be structural artifacts in the render which the denoiser will recognize as actual features in the image.

We plan to further improve the adaptive sampling to avoid such structural artifacts in more cases, but the existence of this behavior is in itself not a bug.

Thanks Brecht,

Further to the above; is there a specific reason for the Viewport denoising default of just 'Albedo'?

All CPU rendering tests performed so far, using 'Albedo and Normal' seem very positive, with barely a performance difference.
'Albedo and Normal' also appears to help offset the issue mentioned in the first post, since artefacts don't appear until higher sample counts.

Kind regards.

It's mainly because of some issues with the normal pass and OptiX denoiser and extra memory usage. We may change the default still.

Thank you again!

In case helpful; yesterday eve, tested the Adaptive Sampling issue.blend with 2.9.3, while Viewport rendering (even with a maximum AS setting of 1.0000), there's the expected minor shading fluctuations as samples refine, but no major artefacts as seen with 3.0.0.

Please understand this is in no way a criticism, your Blender contributions and incredible coding skills are appreciated far more than you may ever realise!