Best Practices for Noise Reduction

I've been using Resolve Studio for some time now to edit footage from a Sony A6500. I shoot in 4K Slog in camera and import it into Resolve using 1920x1080 as the timeline resolution and DNxHR HQX as the optimized media format.

Everything has been running smoothly, but I've only been doing simple edits and color grading. I've recently shot some footage that had a lot of noise, so I tried to use Resolve's noise reduction tools and that's where I ran into problems. I've created two layer nodes and a layer mixer (Composite Mode set to "Add"). One of the nodes I've desaturated and in the other I've removed the luma component. I then apply noise reduction to each node in turn, with higher noise reduction settings on the chroma node.

As soon as I do this, I start getting "Your GPU memory is full" warnings and things become very sluggish, especially when playing the noise reduced clips, and I also see red flashes during playback of those clips. Trying to render output is also problematical as the render fails in the clips that have noise reduction applied, usually with GPU errors, as soon as the renderer starts processing the first clip with noise reduction.

In the Motion Effects, Temporal NR, Frames field, I have the number of frames set anywhere from 2 to 5. Higher settings makes the issues described above worse. What is the best setting for this parameter to maximize noise reduction without killing performance?

Are there any other settings I should change to help performance and reduce GPU out of memory errors when using noise reduction?

I have a Nvidia RTX-3080 GPU and am running v461.92 of the Nvidia Studio driver. Is there another version of the driver that's more robust with Resolve Studio 17.1.1?

My system's specs:
DaVinci Resolve Studio 17.1.1 Build 9
Windows 10 Pro build 19042 (64-bit)
AMD 32-core Threadripper CPU
128 GB RAM
Various NVMe SSD and 7200 RPM hard drives
Nvidia RTX-3080 GPU
Nvidia Studio Driver version 461.92

the best setting for NR is *off*

least worst is to key the area absolutely needing to mashed into mush, and leave the as much of the rest of the image un-f'dup as possiable, then use as little NR as you can get away with

temporal NR works by loading all the frames into GPU mem, so a 5fr setting means you will need 11frames held in GPU memory at one time, and that's a bit of a challange for a GPU with only 8g Vram

i rarely need to go past 2fr, as mentioned above, i key the area that's problematic, no sense crapifying the skintones if they are above the noise floor for example

i put the NR near the last node so i know what the minimum i really need to destroy is

one thing i try before trashing the image with NR is to add noise/grain in the shots before and after so the scene feels matched and the nosiy shot does not stick out

a few time in the last few years i've needed more than Resolve can deal with elegantly, and have turned to Nucoda's Clairity in those cases, meaning a round trip

i can't think of a singe shot in the last few features i've graded that i needed to use NR anyway, think last time i touched that control was sometime last fall

production could save them selves a ton-0-$ and use cheap prosumer lens to get everything into soft mush rather than shooting with great glass and then mashing into crap with NR... end game is much the same - soft mushy images on screen

Wow, Dermot, your reply really got me to thinking about NR and the necessity for it. Going back and looking at my rendered footage, both before and after NR, has convinced me that you're absolutely right--NR is just a band aid that actually makes the footage look worse. Thanks for setting me straight.

the best setting for NR is *off* least worst is to key the area absolutely needing to mashed into mush, and leave the as much of the rest of the image un-f'dup as possiable, then use as little NR as you can get away with

I think Dermot is right: there's always a price to pay for NR, and losing detail and creating more artifacts is not always worth the trade-off.

If I'm working with film negative, usually what I see is more noise in highlights, so I tend to pull a soft luminance key on just the highlights and knock out about 7-8dB of noise, and that's usually enough to get it under control. Skies are a good example.

With digital material, the noise often comes in the shadow areas, so I might qualify the lowlights and do the NR there, experimenting with Chroma-only SNR vs. Y-only SNR.

And if I see a lot of blue-channel-only noise (which does happen for various reasons), I do a Splitter-Combiner node, remove the R & G channels, and just apply Y-only SNR to the Blue channel only. I can go up to 18-20dB before I see any real problems, and it makes the pictures a lot more tolerable.

But I do all this NR on a selective basis, and I'm not a believer in using it all the time. I see a lot of "experts" on the net who advise doing NR in the first node in order to cache it, but the flaw with that is you're noise-reducing the shots that don't need it, plus you're noise-reducing the shots that do need it. And as the o.p. points out, this slows down renders and uses up more system resources.

Take a good hard look at the picture before deciding you really need NR. A lot of today's cameras are really clean -- provided you light at decent levels and have enough fill -- and don't need NR as much as you might think.

I NEVER use multiple instances of noise reduction. It is unnecessary and impacts performance too much.

1. I watch the clip IN MOTION, AFTER COMPRESSION to decide if it REALLY needs NR. (It usually doesn't.)
2. No need for two instances. You can control luma and chroma processing right within the controls.
3. I typically use a luminance mask as a super fast way to limit the NR to the shadows where the noise lives.
4. I actually put this node near the very beginning of the chain and cache it. Then any subsequent adjustments don't impact performance or require re-rendering.