Pixar’s RenderMan 25 will for the first time feature its in-house AI denoiser, and this is “temporally stable”. Translation: when run on animation frames, this de-noiser is stable from frame to frame. When the frames are run as an animation, there’s no strange wavyness, jitter, or edges popping from sharp to blurred and back.
The devs and artist at Pixar report this feature reduces render times “two to four” times, and it also “has CPU and GPU implementations”.
But ‘what use is this to hobbyists’, you might ask. Ah, well… there will be a free non-commercial edition of RenderMan 25 by the end of 2022. The free version is reported to lack only RenderMan’s “XPU” feature — which is Pixar’s “new hybrid CPU + GPU rendering engine” that many are calling the future of high-end rendering.
Thus it sounds to me like hobbyists could have a pro-level ‘temporally stable’ AI denoiser, free and highly trained on 3D CG frames, by the end of the year. And presumably it will be able to process a folder of animation frames produced with other software. Poser 12, for instance, which has a superb Intel denoiser for stills — but this is apparently not “temporally stable” for animation.
Google is also reported to be working on an AI image denoiser, but it’s still in the Labs. Presumably this will be free and open source when it appears. Part of the larger NeRF from Google, a one-click quick image enhancer.