Page 2 of 2

Re: Simple glass with whisky and caustics

Posted: Wed Jan 27, 2021 7:41 am
by Martini
This is LuxCore's beauty pass (DENOISED output)
LuxCore DENOISED output
LuxCore DENOISED output

This is what I get when I recombine all AOVs manually (denoised individually)
All AOVs denoised individually and recombined
All AOVs denoised individually and recombined

You can hopefully see that there is some highlight detail missing from the PhotonGI Caustics Cache (visible through/inside the ice cubes, for example), and it feels 'cold'/artificial.


This is the Caustics AOV when I try to denoise it. You can probably see that the denoiser does not do much to reduce noise on the caustics AOV.
Caustics AOV denoised
Caustics AOV denoised
Another issue is that the caustics from Light Tracing are not visible through the glass, liquid, or ice.

Re: Simple glass with whisky and caustics

Posted: Wed Jan 27, 2021 12:13 pm
by Dade
Martini wrote: Wed Jan 27, 2021 7:41 am This is the Caustics AOV when I try to denoise it. You can probably see that the denoiser does not do much to reduce noise on the caustics AOV.
The source of the problem is in having OIDN not really trained to denoise caustics, etc.

Have you tried to use clamping to reduce the noise on Caustic AOV ?

Re: Simple glass with whisky and caustics

Posted: Fri Jan 29, 2021 5:16 pm
by Martini
Dade wrote: Wed Jan 27, 2021 12:13 pm The source of the problem is in having OIDN not really trained to denoise caustics, etc.
I believe I have found the solution: Pixel Filter! I noticed that the pixel filter is by default in BlendLuxCore set to Blackman-Harris at 1.50 pixels. Even with 1024 light-path spp (thanks for the recent fix for that, by the way - I am Beheaded Kamikaze on GitHub :) ), the denoiser was not able to reduce the noise at all. It looks better (due to having more samples) but you could clearly see the noise, and the caustic pass did not look any different after denoising.

When I changed the pixel filter to None, the denoiser was able to completely remove the noise, even with a low number of light path samples like 128. :D