Noise reduction in linear or log?

[David Brown]

Out of curiosity, I did one small test and found that DaVinci's spatial noise reduction produced a similar result at a lower slider value after I transformed my log footage to linear on either side of the noise reduction node.

Does anyone know if this is a valid finding? Or whether it make sense at the math level? 

Does anyone use this approach?

Thanks! 

[Marc Wielage]

Me personally, I try to only do NR after the initial correction, about halfway through the node tree. You can cache there and then still make subsequent trims, keys, windows, OFX plug-ins, and so on, and it won't slow you down.

Having said that, when I encounter significant noise, we usually turn to Neat Video and render the whole show twice: once without any NR, and then a second pass with NR added on a scene-to-scene basis.

[Kai Klassen]

On 8/23/2021 at 7:36 AM, Marc Wielage said:

Having said that, when I encounter significant noise, we usually turn to Neat Video and render the whole show twice: once without any NR, and then a second pass with NR added on a scene-to-scene basis.

Hi, sounds interesting, may i ask what you do the two renders for and/or how you further process them?

 

I usually use denoise in the beginning of my grade and have that node cached. I mostly  work with log footage and denoise in log as well. If its heavy noise i also use neat video.

Edited August 25, 2021 by Kai Klassen

[Marc Wielage]

1 hour ago, Kai Klassen said:

Hi, sounds interesting, may i ask what you do the two renders for and/or how you further process them?

We use some temp SNR noise reduction while we're color-correcting the show, to give us an idea as to what it'll look with NR. Then we turn off the temp NR for the render to a mezzanine format like ProRes 444 or 444HQ.

Then we take the color-corrected (but no NR) mezzanine version and run it through Resolve again, with only Neat Video activated in a single node. We come up with 7 or 8 different settings for different kinds of scenes -- day interiors, day exteriors, night interiors, night exteriors, super-dark scenes, super-bright scenes, problem scenes -- and manually split the clips and add the NR-only correction from a PowerGrade bin. It won't run at speed, but we do before/after comparisons to make sure it looks good. If the shadows need to be adjusted -- they sometimes wind up a little high after NR -- we lower them. 

Once that's good, we render out what we call a "cc_NR" version (color-corrected noise-reduced), and that's what gets delivered as the final. We hang on to the "cc" (mezzanine) version in case there's any issues. This method has worked for at least 11-12 projects so far, including one I did last week. It does take more time, so it helps to have a fast computer. I just set up a whole stack of renders on the Mac Pro, kick them off at the end of my shift, and it chugs all night until they're done. I set up the OS to turn off the machine after X number of hours, knowing it'll be done by that time.

[Kai Klassen]

ok wow, thanks for that insight, i like that approach.

sounds like another half day of work on a cc, right? 

 

 

[Marc Wielage]

34 minutes ago, Kai Klassen said:

ok wow, thanks for that insight, i like that approach.

sounds like another half day of work on a cc, right? 

Technically, it is more time but we'll just stack up the renders and I'll hit the button on my way out the door at the end of the day. If it takes 6-7 hours to do them all, it doesn't matter: I'll be safely home in my bed. 

We tend to work in reels, so the next day I'll stitch them all together (assuming it's a feature), but I make sure I check off the "Bypass Re-encode When Possible" option is turned on on the Delivery page. Usually I can get a flattened single file out of the 4-5 files in faster-than-real-time by the moment I get back into the office, maybe 45 minutes for a 2-hour film in 4K ProRes 444.