Noise Reduction, Am I doing it right?

When doing noise reduction, I set it to the first node. After, I change the frames to 3, the motion estimation type to better, and the motion range based on the scene. For temporal I set chroma to 25 for less chroma noise and 40 when more chroma noise is present. For luminance, I set up to 50 if needed though anywhere above that leads to the noise looking like water flowing, a weird artifact.

For spatial NR, I set radius to large always and the luma and chroma to 1/2 the temporal value. I set the mode to Enhanced usually, though sometimes better looks better despite using more processing power. Faster looks awful and smears the image like no tomorrow. If I don't get the desired result, I start adjusting these values over the starting 1/2 temporal point until I get a good mix between detail and noise to avoid the image looking like a wax mess.

I keep to a one node process and I do know some do a complex tree just for NR but I do not have time for that.

Do you all find better better than enhanced sometimes, or am I going crazy?
Also, have any of you come across the flowing water side effect of temporal NR? I find it very strange and sometimes more distracting than the noise.

I use high denoise values because sometimes I shoot in crazy high ISOs on my Panasonic S5 (204800 in a dark forest.)

just accept a bit of noise... otherwise it get plasticky...

Even when shooting at extremely high ISOs?

I don't think I ever used any temporal NR in Resolve and even for lowlight scenes I typically have spatial luma set to 4 maximum (but usually 0) and chroma around 7.

I have done heavier denoises in the past using NeatVideo but I always try to do as minimal as possible. As Walter says, you have to accept most of it. If there was little light there's noise. We're used to it being like that and getting rid of it fully will always look nasty with huge loss of detail.

The only other proper way to address noise is to shoot with more light or a better performing low light camera/sensor.

I don't think I ever used any temporal NR in Resolve and even for lowlight scenes I typically have spatial luma set to 4 maximum (but usually 0) and chroma around 7.

I have done heavier denoises in the past using NeatVideo but I always try to do as minimal as possible. As Walter says, you have to accept most of it. If there was little light there's noise. We're used to it being like that and getting rid of it fully will always look nasty with huge loss of detail.

The only other proper way to address noise is to shoot with more light or a better performing low light camera/sensor.

I find that the temporal and spatial NR does a good job in resolve though. I understand on professional shoots never going above base ISO or something, but I like having the flexibility of not bringing around a lighting rig with me. I’m shooting on full frame with f/1.8 and f/1.4 lenses whenever in low light.

When doing noise reduction, I set it to the first node. After, I change the frames to 3, the motion estimation type to better, and the motion range based on the scene. For temporal I set chroma to 25 for less chroma noise and 40 when more chroma noise is present. For luminance, I set up to 50 if needed though anywhere above that leads to the noise looking like water flowing, a weird artifact.

I think this is a bad idea, because (generally) the first node directly from the source will either be a Log image, or it'll at least lack the contrast added by the rest of the grade.

I'm with Walter above: there's always a danger in adding too much noise reduction, where you wind up with a very "plasticky" image lacking in detail, sucking the life out of the image. Remember that there's always a price to pay for noise reduction. Less is more.

I use one of two methods:

1) with Resolve TNR/SNR, I usually will use a node somewhere in the middle of the node tree, so we've already built up a lot more contrast and the levels are reasonable. I generally make different settings on Luma and Chroma for SNR, and I think the Enhanced mode can deliver better results sometimes.

2) when I have the time, I'll render out the entire color-corrected project to a "visually-lossless" format (like ProRes 444XQ) in the delivery resolution, then create a new timeline with one node just for Neat Video. I'll chop up the timeline based on scene content -- day scenes vs. night scenes, interiors vs. exteriors, blown-out shots vs. underexposed shots -- and adjust the Neat Video NR accordingly. I keep at least a half-dozen presets of Neat and usually wind up backing off at least 40% for their versions of SNR/TNR, after using their analysis pass.

In some cases, I'll use a Resolve NR node targeted especially for a single color channel or just the shadows: there are ways to use the Blue-only channel, which (for me anyway) is where a lot of the noise lives in underexposed digital or film. Other cases, rather than use NR per se, I can "flatten" out shadow noise by qualifying a soft key on information below about 20ire, then using SNR Luma only. This at least "softens" the noise in those areas, which can help. Other times, I'll manually desaturate the shadows, which is great if they're contaminated with color noise. Every situation is different and requires some judgement and experimentation.

bit offtopic, was just wondering which camera is the absolute best in low light these days?

bit offtopic, was just wondering which camera is the absolute best in low light these days?

How much noise/low quality are you willing to bear?

I second denoising of the blue channel, at least more than the other two. Blue is where most of the noise happens in digital cameras, since silicon sensors are least sensitive for blue and much more for red or infrared. Quite the contrary to optochemical film. And then, or eyes resolve worst in blue, so we are noticing less of the detail loss.

How much noise/low quality are you willing to bear?

let's say as close as possible to zero noise

let's say as close as possible to zero noise

Somewhat difficult to answer. All sensors have a base sensor noise there is no sensor that doesn't have it. How visible it is depends on how the camera manufacturers engineer it and tweak the converters to optimize for whatever encoding they let it record to. What matters most for a low as possible noisefloor comes mostly down to sensor size and what it is you are actually recording. A scene that doesn't have any dynamic range whatsoever that is in low light will have to be enhanced gained by such an amount that you'll inevitably bring up the noise with it. But overall as it stands now I think Sony full frame sensors are best performing.

Getting good low light performance is important but not key for me. If something is dark, why grade it like it's bright/day.

Like I said before, if you have the tools/budget/option.. shoot with more light. You can bring down healthy exposure very cleanly provided nothing clips the sensor. Bringing something up brings up the noise floor.

If you want a music or sounddesign piece that is mixed/mastered quitely you also don't record it quitely. ETTR

This will get you on the right track my friend:

I think this is a bad idea, because (generally) the first node directly from the source will either be a Log image, or it'll at least lack the contrast added by the rest of the grade.

I'm with Walter above: there's always a danger in adding too much noise reduction, where you wind up with a very "plasticky" image lacking in detail, sucking the life out of the image. Remember that there's always a price to pay for noise reduction. Less is more.

I use one of two methods:

1) with Resolve TNR/SNR, I usually will use a node somewhere in the middle of the node tree, so we've already built up a lot more contrast and the levels are reasonable. I generally make different settings on Luma and Chroma for SNR, and I think the Enhanced mode can deliver better results sometimes.

2) when I have the time, I'll render out the entire color-corrected project to a "visually-lossless" format (like ProRes 444XQ) in the delivery resolution, then create a new timeline with one node just for Neat Video. I'll chop up the timeline based on scene content -- day scenes vs. night scenes, interiors vs. exteriors, blown-out shots vs. underexposed shots -- and adjust the Neat Video NR accordingly. I keep at least a half-dozen presets of Neat and usually wind up backing off at least 40% for their versions of SNR/TNR, after using their analysis pass.

In some cases, I'll use a Resolve NR node targeted especially for a single color channel or just the shadows: there are ways to use the Blue-only channel, which (for me anyway) is where a lot of the noise lives in underexposed digital or film. Other cases, rather than use NR per se, I can "flatten" out shadow noise by qualifying a soft key on information below about 20ire, then using SNR Luma only. This at least "softens" the noise in those areas, which can help. Other times, I'll manually desaturate the shadows, which is great if they're contaminated with color noise. Every situation is different and requires some judgement and experimentation.

Very valuable advises, I'll copy and store them.
Thanks a lot Marc.

I second denoising of the blue channel, at least more than the other two. Blue is where most of the noise happens in digital cameras, since silicon sensors are least sensitive for blue and much more for red or infrared. Quite the contrary to optochemical film. And then, or eyes resolve worst in blue, so we are noticing less of the detail loss.

Ya know, I do an awful lot of film (36 titles this year, seriously), and the blue channel is typically the noisiest emulsion layer in 35mm negative scans as well. I use the "Blue-only NR" trick when I get into dark scenes where they underexposed it, and -- for whatever reason -- the lab chemistry reacted in such a way as to make a soft and ugly blue channel.

For those reading in, the two methods you can use for Blue-only NR:

1) create a splitter-combiner node, then just DELETE the first two nodes (Red and Green), leaving only blue. Go into the SNR control panel, and adjust Lum only. You might ask, why adjust Lum when we need to adjust Blue? And the answer is, because this is now a monochrome signal. I try not to do more than 15-20dB of NR, and I'll do some A/B tests to make sure I haven't done too much. I generally then do a node immediately after that to bring the shadows down a "little" bit, usually Log Shadows -1 or -2, something like that. You can put all that in a Compound Node, and it takes up very little space.

2) create a node, right-click it, and select CHANNELS. Turn off 1 and 2 (R & G), leaving only 3, which is blue. Do the same amount of NR, but note that in this mode, Luma and Chroma SNR will both work. Again, after the NR you'll need to apply a very small amount of shadow adjustment to balance them out.

The results of method 1 and method 2 are different, and I'm not exactly sure why, but I use whatever is faster and neater.

The results of method 1 and method 2 are different, and I'm not exactly sure why, but I use whatever is faster and neater.

I'm very curious to why that would be then... the math should be the exact same no? Method 2 sounds/looks a lot cleaner though.

Probably because in one case blue channel is handled as monochrome image, as Marc wrote, and so in NR sense it is the same value in RGB channels, meaning there is no chromaticity. In second case luma-chroma components are calculated as usual but noise reduction is applied only to blue channel, red and green are passed through as-is.

I second denoising of the blue channel, at least more than the other two. Blue is where most of the noise happens in digital cameras, since silicon sensors are least sensitive for blue and much more for red or infrared. Quite the contrary to optochemical film. And then, or eyes resolve worst in blue, so we are noticing less of the detail loss.

I totally agree.

A lot of my footage is shot against blue-sky and once I've done some stabilization (which causes zoom/crop) the noise can be visible. A light amount of temporal noise reduction gets rid of the noise without affecting the other elements of the image to any perceptable amount.

Yeah, cheap (US$1,600) camera with a small sensor but it's the best tool for the job with a 25-600mm zoom (HC-X1500).

I second denoising of the blue channel, at least more than the other two. Blue is where most of the noise happens in digital cameras, since silicon sensors are least sensitive for blue and much more for red or infrared. Quite the contrary to optochemical film. And then, or eyes resolve worst in blue, so we are noticing less of the detail loss.

Ya know, I do an awful lot of film (36 titles this year, seriously), and the blue channel is typically the noisiest emulsion layer in 35mm negative scans as well. I use the "Blue-only NR" trick when I get into dark scenes where they underexposed it, and -- for whatever reason -- the lab chemistry reacted in such a way as to make a soft and ugly blue channel.

For those reading in, the two methods you can use for Blue-only NR:

1) create a splitter-combiner node, then just DELETE the first two nodes (Red and Green), leaving only blue. Go into the SNR control panel, and adjust Lum only. You might ask, why adjust Lum when we need to adjust Blue? And the answer is, because this is now a monochrome signal. I try not to do more than 15-20dB of NR, and I'll do some A/B tests to make sure I haven't done too much. I generally then do a node immediately after that to bring the shadows down a "little" bit, usually Log Shadows -1 or -2, something like that. You can put all that in a Compound Node, and it takes up very little space.

2) create a node, right-click it, and select CHANNELS. Turn off 1 and 2 (R & G), leaving only 3, which is blue. Do the same amount of NR, but note that in this mode, Luma and Chroma SNR will both work. Again, after the NR you'll need to apply a very small amount of shadow adjustment to balance them out.

The results of method 1 and method 2 are different, and I'm not exactly sure why, but I use whatever is faster and neater.

Interesting! The 2 suggestions are a great way to control noise

I'm very curious to why that would be then... the math should be the exact same no? Method 2 sounds/looks a lot cleaner though.

Try it yourself and see. In cases like these, I just assume that there's math going on that goes in one direction and not another, and I just concern myself with the results. In truth, the Splitter node is more effective for me, but it's little clunkier.

When doing noise reduction, I set it to the first node. After, I change the frames to 3, the motion estimation type to better, and the motion range based on the scene. For temporal I set chroma to 25 for less chroma noise and 40 when more chroma noise is present. For luminance, I set up to 50 if needed though anywhere above that leads to the noise looking like water flowing, a weird artifact.

I think this is a bad idea, because (generally) the first node directly from the source will either be a Log image, or it'll at least lack the contrast added by the rest of the grade.

I'm with Walter above: there's always a danger in adding too much noise reduction, where you wind up with a very "plasticky" image lacking in detail, sucking the life out of the image. Remember that there's always a price to pay for noise reduction. Less is more.

I use one of two methods:

1) with Resolve TNR/SNR, I usually will use a node somewhere in the middle of the node tree, so we've already built up a lot more contrast and the levels are reasonable. I generally make different settings on Luma and Chroma for SNR, and I think the Enhanced mode can deliver better results sometimes.

2) when I have the time, I'll render out the entire color-corrected project to a "visually-lossless" format (like ProRes 444XQ) in the delivery resolution, then create a new timeline with one node just for Neat Video. I'll chop up the timeline based on scene content -- day scenes vs. night scenes, interiors vs. exteriors, blown-out shots vs. underexposed shots -- and adjust the Neat Video NR accordingly. I keep at least a half-dozen presets of Neat and usually wind up backing off at least 40% for their versions of SNR/TNR, after using their analysis pass.

In some cases, I'll use a Resolve NR node targeted especially for a single color channel or just the shadows: there are ways to use the Blue-only channel, which (for me anyway) is where a lot of the noise lives in underexposed digital or film. Other cases, rather than use NR per se, I can "flatten" out shadow noise by qualifying a soft key on information below about 20ire, then using SNR Luma only. This at least "softens" the noise in those areas, which can help. Other times, I'll manually desaturate the shadows, which is great if they're contaminated with color noise. Every situation is different and requires some judgement and experimentation.

I will look into those methods, thanks!

Is there a mathematical measurement of noise in an image?

Everyone has opinions on where to put NR is the node tree. I would be interested to see some systematic tests.

PSNR, SIMM?

If we took master clip with no noise, added noise to it. Measure against master.

Then place the NR node in log space, DWG and Rec709, apply same amount of NR in each and test again against master.

See what node position lets you use the least NR to get reasonable restoration against the master?


Do you start with an animation or CG to make sure there is no noise to start with? Or is real footage from a good camera at Low ISO a better test?

Is there a mathematical measurement of noise in an image?

Everyone has opinions on where to put NR is the node tree. I would be interested to see some systematic tests.

PSNR, SIMM?

If we took master clip with no noise, added noise to it. Measure against master.

Then place the NR node in log space, DWG and Rec709, apply same amount of NR in each and test again against master.

See what node position lets you use the least NR to get reasonable restoration against the master?


Do you start with an animation or CG to make sure there is no noise to start with? Or is real footage from a good camera at Low ISO a better test?

I didn't try something completely like this but I did find that on images with extreme noise, the NR worked far better at the absolute beginning (before even the color space transform) rather than after the grade. It preserved far more detail with less noise than at the end of the node tree.