Intel's powerful denoise library's latest update has been out for less than a month, but we already have it in the Blender master branch, which means it is available in the latest Blender 3.0 alpha builds.
#OpenImageDenoise 1.4 is out. Check out the new features in this major update to Intel's #opensource CPU render denoiser, used in #B3D, #Unity3D and #VRay, including support for normal and albedo maps and #AppleSiliconhttps://t.co/zKYG2u4MKM#rendering #archviz #VFX #gamedev pic.twitter.com/FIJIxjcqHG
— CG Channel (@theCGchannel) May 24, 2021
OIDN was already a favorite among Blender's denoising options, with the library arguably doing a better job at detail preservation than its NVidia OptiX denoiser counterpart, at the cost of more ram usage and slightly slower performance. This latest major update makes the denoiser even more performant, reducing memory usage, adding the ability to denoise normal and albedo maps, and finally doing a much better job at fine detail preservation. Check it out in action by downloading the latest Blender daily build from the link below:
5 Comments
I prefered optix in all of my projects since it was was available because of much less 'boiling' in animations - i'm doing vfx stuff, cg backstage and so on. Every time I use oidn to denoise render passes 'just in case', and every time I had to use them, the result was worse then clean optix.
In my practice there was only one case, when oidn worked better: very early morning on the sea with branched path teaching gave me less boiling then optix.
Is it in Cycles X branch as well?
Hi, I am wondering this as well. I only use the cycles-x builds now. However I never liked Open Denoise as it removes smaller details, but would like to try out this new version.
great to see development in this area!
my main wish for the openimagedenoise library is that they eventually figure out some sort of temporal coherence for animations.
OIDN was always the worst option for me. I'm a product designer and NLM worked the best for me in every single case, with sample counts of 500-750 NLM preserves all low frequency detail while OIDN just blasts every low frequency detail out of orbit at the same sample size.