Understanding Sony’s SLog3. It isn’t really noisy.

workshops-275 Understanding Sony's SLog3. It isn't really noisy.

It’s been brought to my attention that there is a lot of concern about the apparent noise levels when using Sony’s new Slog3 gamma curve. The problem being that when you view the ungraded Slog3 it appears to have more noise in the shadows than Slog2. Many are concerned that this “extra” noise will end up making the final pictures nosier. The reality is that this is not the case, you won’t get any extra noise using Slog3 over Slog2 and Slog3 is generally easier to grade and work with in post production.

So what’s going on?

Slog3 mimics the Cineon Log curve. As a result the shadow and low key parts of the scene are shown and recorded at a brighter level than Slog2. Because the shadows are brighter, the noise in the shadows appears to be worse. It isn’t. The noise level might be a bit higher but the important thing, the ratio between wanted picture information and un wanted noise is exactly the same whether in Slog2 or Slog3 (or in fact any other of the cameras gamma curves at the native ISO).

Let me explain:

The signal to noise ratio of a camera is determined almost entirely by the sensor. This is NOT changing between gamma curves.

The other thing that effects the signal to noise ratio is the exposure level, or to be more precise the aperture and how much light falls on the sensor. This should be same for Slog2 and Slog3. So again no change there.

As these two key factors do not change when you switch between Slog2 and slog3, there is no change in the signal to noise ratio between Slog2 and Slog3. It is the ratio between wanted picture information and noise that is important. Not the noise level, but the ratio. What people see when they look at ungraded SLog3 is a higher noise level (because the signal levels are also higher), but the ratio between the wanted signal and the noise is actually no different for both Slog2 and Slog3 and it’s that ratio that will determine how noisy your pictures will be after grading.

Gamma is just gain, nothing more, nothing less, just applied by variable amounts at different levels. In the case of log, the amount of gain decreases as you go further up the curve.

Increasing or decreasing gain does NOT significantly change the signal to noise ratio of a digital camera (or any other digital system). It might make noise more visible if you are amplifying the noise more, for example in an under exposure situation where you add gain to make a very dark object brighter. But the ratio between the dark object and the noise does not change, it’s just that as you have made the dark object brighter by adding gain, you have also made the noise brighter by the same amount, so the noise also becomes brighter and thus more obvious. The ratio between the wanted signal and the unwanted noise remains constant, no matter what the gain, it is a ratio and gain does not change ratios. With Slog3 in post production you will need less gain in the shadows than you would with Slog2 and this negates the extra gain that the camera adds to the shadows when shooting SLog3.

Lets take a look at some Math. I’ll keep it very simple, I promise!

Just for a moment to keep things simple, lets say some camera has a signal to noise ratio of 3:1 (SNR is normally measured in db, but I’m going to keep things really simple here).

So, from the sensor if my picture signal is 3 then my noise will be 1, or if my picture signal is 6 then my noise will be 2.

If I apply Gamma Curve “A” which has 2x gain then my picture becomes (6×2) 12 and my noise (2×2) 4. The SNR is 12:4 = 3:1

If I apply Gamma Curve “B” which has 3x gain then my picture becomes (6×3) 18 and my noise becomes (3×2) 6. The SNR is 18:6 = 3:1 so no change to the ratio, but the noise is 6 compared to the 4 of Gamma “A”, as a result Gamma “B” will appear to be noisier when viewed on a monitor.

Now we take those imaginary clips in to post production:

In post we want to grade the shots so that we end up with the same brightness of image, so lets say our target level after grading is 15.

For the gamma “A” signal we need to add 1.25x gain to take 12 to 15. As a result the noise now becomes (1.25 x 4) 5.

For the gamma “B” signal (our noisy looking one) we need to use 0.8333x gain to take 18  to 15. As a result the noise now becomes (0.83333 x 6) 5.

Notice anything? In both cases the noise in the final image is exactly the same.

OK, so that’s the theory, what about in practice?

Take a look at the images below. These are 400% crops from larger frames. Identical exposure, workflow and processing for each. You will see the original Slog2 and SLog3 plus the Slog 2 and Slog 3 after applying the LC-709 LUT to each in Sony’s raw viewer. Nothing else has been done to the clips. You can “see” more noise in the raised shadows in the untouched SLog3, but after applying the LUTs the noise levels are the same. This is because the Signal to Noise ratio of both curves is the same and after adding the LUT’s the total gain applied (camera gain + LUT gain) to get the same output levels is the same.

Slog2-400 Understanding Sony's SLog3. It isn't really noisy.
Slog3-400 Understanding Sony's SLog3. It isn't really noisy.Slog2-to-709-400 Understanding Sony's SLog3. It isn't really noisy.Slog3-to-709-400 Understanding Sony's SLog3. It isn't really noisy.

It’s interesting to note in these frame grabs that you can actually see the improvement in shadow detail that SLog3 brings. The bobbles and the edge of the picture frame look better in the Slog3 in my opinion. A little bit more shadow data has given a more pleasing result with fewer artefacts.

The only way you can alter the SNR of the system (other than through electronic noise reduction) is by changing the exposure, which is why EI is so important and so effective.

Noise is always most problematic in shadows and low key. As we are putting more data into the shadows with SLog3 we are in effect recording the noise in the shadows more precisely, you won’t enhance it or increase it. All that will happen is that it is more accurately reproduced with fewer artefacts, which is a good thing.

In addition Slog3 has a near straight line curve. This means that in post production it’s easier to grade as adjustments to one part of the image will have a similar effect to other parts of the image. It’s also very, very close to Cineon and to Arri Log C and in many cases LUT and grades designed for these gammas will also work pretty well with SLog3.

The down side to Slog3?

Very few really. Fewer data points are recorded for each stop in the brighter parts of the picture and highlights compared to Slog2. As a result Slog3 is slightly less forgiving of overexposure than Slog2. You probably don’t want to push your EI gain quite as hard with Slog3. 1.5 stops over should be OK (so using an EI 1.5 stops down from native) but 2 or more will hurt your pictures.

Want to pick my brain for 10 days, fancy an adventure and a chance to see and shoot some very cool sights? Why not join me for a storm chasing adventure. 

Advertisements

45 thoughts on “Understanding Sony’s SLog3. It isn’t really noisy.”

    1. Raw is linear, so where you place your exposure within the cameras 14 stop range makes little difference (other than noise). S-Log2 and S-Log3 use log encoding so as you go up the exposure range less data is recorded relative to the amount of information in the scene. Over exposing log by more than 1.5 stops (Using an EI 1.5 stops down) can make it harder to grade the image as a result.

      1. For linear Raw, the higher up you go (getting closer to clipping), the more data you’ll have.
        Let’s remember that half a sensor bit depth is dedicated to the highest stop of light before clipping.
        So the more you overexposed without clipping, the richer de linear raw file. (less true for log of course)

        1. The F5/F55/F65 use floating point so there is an abundance of data in the mid range and more data in the shadows than you would have with a true linear system. Floating point puts a lot more data into the shadows and mids at the expense of some highlight accuracy. Bright exposure is of course still beneficial for a better SNR and more data, but the data gains are not as significant as you might expect.

  1. “The only way you can alter the SNR of the system (other than through electronic noise reduction) is by changing the exposure..”

    There’s another way that isn’t practical for users; but, available to camera designers/manufacturers. In fact, it’s rather surprising some higher end OEM’s haven’t used it. Sensor cooling has been used for some specialty sensor applications for decades. SNR can be significantly increased by sensor cooling, adding a whole new dimension to camera operation…:0)

    1. Actually many cameras use passive cooling. The F5/F55 cameras have a heat sink on the sensor and are fan cooled. But to get a significant benefit you really need to cool the sensors with an active cooling system which would really push up the power consumption of the camera and can lead to condensation and moisture problems in humid environments as the sensor would end up below the dew point. This the requires hermetically sealed sensor compartments making the camera bulky and more expensive and even then you may still get condensation on the optical port which will be cold on one side and warm on the other.

    2. The sensors in the Sony pro video cameras are fan cooled. Yes, you can reduce noise with active cooling, but how do you deal with all the condensation issues that you would get if the optical port is colder than the dew point? This would be a huge problem in warm humid climates. In cold climates you could also have problems with frost forming on the port.

      Active cooling can work with 100% sealed systems with non removable lenses, but it really isn’t practical on this type of camera.

      1. ISO 5800:2001 is the system used to measure the sensitivity of film. ISO 2240:2003 is the standardised way to express the equivalent sensitivity of a digital camera. ISO is short for International Standards Organisation.

        Within the digital camera ISO system there are 5 different standards that a camera manufacturer can use when obtaining the ISO rating of a camera. The most commonly used method is the Recommended Exposure Index (REI) method, which allows the manufacturer to specify a camera model’s EI or base ISO arbitrarily based on what the manufacturer believes produces a satisfactory image. So it’s not really a measure of the cameras sensitivity, but a rating that if used with a light meter will give a satisfactory image.

        As you cannot change the sensor in a digital camera, you cannot change the cameras efficiency at converting light into an electronic signal (which is largely determined by the materials used and the physical construction), so you cannot change the actual sensitivity of the camera.

        Most of the noise in the pictures comes from the sensor and the level of this noise is largely unchanged no matter what you do (some dual ISO cameras use variations in the way the sensor signal is sampled to shift the noise floor up and down a bit). So the biggest influence on the signal to noise ratio is the amount of light you put on the sensor. More light = better signal to noise ratio, up to the point where the sensor overloads.

        To obtain a brighter image when there is less light the signal coming from the sensor is amplified (gain is added). This amplification makes both the desirable signal bigger but also the noise bigger. The signal to noise ratio deteriorates because the added amplification means the recording will clip readily, so the range you can record reduces while the noise gets bigger. However the best exposure is now achieved with less light so the equivalent ISO number is increased. But the camera isn’t more sensitive, it’s just that the optimum amount of light for the “best” or “correct” exposure is reduced due to the added amplification.

        To combat the noise increase as you add gain/amplification most modern cameras use electronic noise reduction which is applied more and more aggressively as you increase the gain. At low levels this goes largely un-noticed. But as you start to add more and more noise reduction you will start to degrade the image. It may become softer, it may become smeary. You may start to see banding ghosting or other artefacts. Arguably the signal to noise ratio is improved by the use of noise reduction, but this typically comes at a price and NR can be very problematic if you later want to grade or adjust the footage as often you won’t see the artefacts until after the corrections or adjustments have been made.

  2. this may seem stupid , forgive me, but does the camera need to be set to slog3 and if so how do you know your footage was shot using that curve from looking at the raw files?

    1. Raw is raw and S-log is S-log. They are two different things. If you are shooting raw, then the camera may be set to S-log3 so that the built in S-log3 LUT’s or 3rd party user LUT’s for S-Log3 can be used. But the recording is still raw. When the raw is de-bayered in post it can be displayed as S-log3 if you wish or almost any other gamma.

        1. When shooting UHD the FS5 is 8 bit. S-log2 is better optimised for recording data from a digital sensor and as a result gives a better end result when you don’t really have enough code values.

  3. I have a possible night shoot coming up, filming a tribe that has evolved to have superior night vision than the rest of us. The director is keen to have the minimum of lighting as otherwise he feels it will diminish the story. The tribe are going to be night fishing from boats. So far, I’ve been shooting in slog3 sgamut3cine and rating the camera at 1250 with good clean results. The camera is recording internally at 2000 in cine EI – am I better sticking to full log, keeping an eye on the low key function and pushing it a stop or so in the grade, or is it better to bake in slog 3 and use the iso more like gain and use it with discretion?

    1. But the brightness to noise ratio, which is what really matters, is identical for both S-log2 and S-log3. There is no difference.

  4. I am one of those who have experienced issues with the FS7 “noise” issues. First of all, lets be clear. The FS7 does not in my opinion suffer from noise issues at all regardless of S-Log2/3 setting. The CINE-EI image at 2000 ISO is completely clear from noise in a conventional sense. There are no multicolored speckles in the shadows. However there is quite a bit of banding in the lower end of the image. This affects skin tones and details in the shadows. I also find color shifts from red to magenta (which should not happen, since those colors are at the opposite wavelengths). I have spent a lot of time trying to figure out what’s going on. To me this looks like aggressive noise suppression and smoothing. It could also be artifacts from the h.264. It’s probably both.

    You mention the SNR is what’s important, I think on the FS7 SNR is 57db. As far as my math can take me, this is the same as 9.5 stops. However in Slog3 mode, the FS7 can record up to 14 stops of information. My question is: what happens in the bottom 4.5 stops of the image? There should be quite a bit of noise, since you are below the Signal to Noise threshold. But there is no noise down there, just weird artifacts, such as banding and color shifts. Try shooting a dark smooth gradient at 2000 ISO, and you’ll see what I mean – it’ll spike on your histogram.

    Coming back to S-Log3 vs S-Log2, I found out that S-Log3 clips at 93 IRE, S-Log2 at 104 IRE. I think this is because S-Log3 is optimized for 16 stops of latitude. This means that you are throwing 14% of the grey scale information away and use it for nothing. Under normal circumstances, that wouldn’t be too much of an issue – you’ll just apply some gain, a gamma curve or a lut and bring the code values up. And that would be fine if you were in an uncompressed 4:2:2 10 bit color space – it would be easy. However as an FS7 owner you are in XAVC, which is compressed by a 30:1 ratio, and may or may not be a completely true 10bit codec (the color palette seems reduced to me, but that could be the XAVC decoding in Davinci Resolve). And when you gain the information up, it seems like there are a lot of quantization errors, due to the limited codec and the smoothing in the darker exposed parts of the image. So personally I get better results by using S-Log2 compared to S-Log3, since I have to move my skin tones less.

    As an FS7 owner/operator, I am aware of the advantages of Slog-3, mainly the lack of color shift that is present in S-Log2, and a more Cineon/LogC like curve. However I rarely send out my footage ungraded, and more often than not, I cut together with other SONY cameras (all using S-Log2), so being able to match LogC in an Alexa/Amira is not a huge concern. I am sure, that once I get a new SONY camera with 16 stops latitude and a more rubust codec, I’ll be using S-Log3 a lot; but for now, it seems like a future technology with limited use for my current purposes.

    1. S-log3 matches Cineon, that’s why it only goes up to 92 IRE, just as the Arri cameras only go up to about 96 IRE. Yes some data is wasted but with 10 bit it’s a marginal effect as above middle grey the difference between S-log2 and S-log3 in terms of code values per stop is only about 20. Below middle grey S-Log3 has more code values per stop than S-log2 so from a noise point of view, even though there is less overall data, S-log3 will normally introduce fewer additional noise artefacts.

      In the Cine EI mode the FS7’s primary noise reduction is disabled (although there is on sensor NR in the A to D’s). However H264 encoding does incorporate noise reduction processes of it’s own. My FS7 certainly has a fair amount of noise in the shadows and I do not see any weird banding or color shift artefacts provided the material is exposed correctly and handled correctly in post.

      Signal to noise ratios do not work as you are suggesting. Signal to noise in a video camera is a linear function, it isn’t measured in stops it’s measured in volts compared to the 0.7 volt video output range. 57db means that the noise amplitude in the video output will be about 1.2mV (1.2mV being 57db below 0.7v). The darkest stop that the FS7 records comes in at around 2mV above black, so even this darkest stop is just about above the noise floor ( but only just and in reality of little use). As we go up from black, as the noise remains constant at around 1.2mV but each stop (for the first 4 or 5 stops) is doubling in recording amplitude the effective amount of noise compared to useable signal is halved for each stop you go up in brightness. This is why deliberate over exposure is so beneficial as simply over exposing by one stop will halve your noise and resulting noise artefacts for a like for like grade.
      In reality SNR isn’t quite as simple as above as gamma gain plays a part (which is why S-log3 appears noisy as there is a lot of extra gamma gain in the shadows). The key thing is it’s measured as a linear value relative to the output amplitude, not the input range. Then it’s typically expressed in db as this allows large ratios to be described using a small number and makes the math easier if doing any multiplication or division.

      When correctly exposed skin tones in S-log2 will be lower than S-log3 (both have similar white points, but S-log2 middle grey is much lower, as skin tones are approx half way between mid grey and white S-Log2 skin tones will be lower), so if you are finding that you are having to move S-log3 skin tones more than S-log2 in post you are incorrectly exposed somewhere. If you deliberately over expose S-log2 or s-Log3 by 1.5 stops both should be recording skin tones at approx the same level of 60% which is the point where the two curves cross-over, so between correctly exposed and 1.5 stops over S-log3 skin tone will need less movement than S-log2, in addition S-log3 will be recording the darker tones at a higher level so there will be more data per darker stop than S-log2 and thus fewer compression artefacts. In practice this is what I see, S-log3 giving much improved darker zones with little difference in the high range.

      I think before you stop using S-log3 you need to look at the exposures you are using as something appears to be very wrong.

  5. Great article. I read every one of yours I find! I learn something each time, or am validated in my own findings in some areas. Coming from the C100, this FS7 just has SO many more options (and thus confusion), but I love the potential power over the image. Keep the articles coming.

  6. Great article. Explains a LOT. So what are the best settings all throughout the large in camera menu, and most importantly, I’m confused on what the El means, and I am trying to figure out the best total way to expose in slog3, and the best way to set a custom lut to help me expose with the best results, especially as far as dynamic range?

    1. Best settings for what? There is no such thing as “best settings” because it depends on exactly what it is you desire. Take a look at my FS7 cine EI guide.

  7. Hi there,

    I recently filmed with both the FS7 and A7sII in S-log 3 mode and also edited the footage and I have been puzzled because it does look noisy, but I know it’s not supposed to. Especially the A7sII is hailed as the noise-free king, so what am I doing wrong or how am I supposed to treat it in post? In Premiere I usually just adjust color balance and add more contrast back in. Sometimes I play with some of the Premiere LUTs such as “Neutral Start.” I know in the first Lumetri tab I could specify Slog-3, but it’s already shot in S-log3 so I never apply that pre-lut… or do I have to?

    Thanks for helping
    Elias

    1. No video camera is noise free. Even the A7sII has a fair amount of noise. The secret to reducing the noise is to expose correctly. This means deliberately over exposing by about 1 stop with the FS7 and 1.5 to 2 stops with the A7. Then in post applying an S-curve to de-log the footage prior to any further grading operations or the application of the correct LUT for the curve used. S-log3 is not a good curve to use with a 8 bit codec as in the A7s, you may get better results using S-log2.

      1. Hi Alister, thank you so much for your patience and good advice.

        I’m planing to shoot with the a7s mark 2 soon and thought I could give it a try to use SLOG3 recording in ProRes422 to the Ninja Assassin (still 8bit, I know…).

        As I read the comment below, you thin it isn’t a good idea right? So I rather should shoot in SLOG2? What exactly is your appreciated advice?

        Best regards, Julian

        1. S-Log2 makes better use of the recording range available when shooting with 8 bit. So that’s one reason for choosing S-Log2, but the difference is very small. The main reason is that as the A7 and FS5 cameras don’t have proper LUT’s I am usually monitoring looking at the log image. S-Log3 is very flat and very hard to gauge exposure as a result. S-Log2 has more contrast so it’s easier to estimate exposure and also easier to focus.

  8. Hello, Sir.
    I bought an FS7 for documentary filming and i`ve noticed some noise in some scenes at ISO 2000. So this is the native ISO of the camera? can i film with lower ISO in outdoor scenes, or is it better to film always wits ISO 2000 to use the full dynamic rage of the camera?
    And another question is about color space. What is the difference between S-Gamut3.Cine/SLog3 and S-Gamut3/SLog3? They seem the same to me. What do you recommend for my work?
    Thank you!
    Florian

  9. Thanks Alister, for these very informative comments.
    I still wonder if it is worth using SLOG3 versus Cine2 or Cine4, when the objective is to preserve details in the shadows.
    Indeed, the Cine4 footage is so much cleaner than Slog3, than it seems to me that after grading it will preserve more details than Slog3, where we have to crush more the blacks to remove the noise (unless using a denoiser type of software).
    So using Slog2 to preserve details in the highlight is understandable, but Slog3 looks to me more indicated for shadow details, but the noise reduction in post production seems to annihilate most of the benefits compared to using a Cine gamma. I may be doing something wrong though…
    What’s your view on this ?

  10. Hi,

    First let me thank you for this great article. However there is one thing that keeps bugging me. I mostly shoot Canon (C100MK2 and 5D4). I just sold my a6300 which I liked for the specs but not too much for the ergonomics and usability in general. I also own a Kinefinity 4K which has an EI mode. Now one thing that I noticed with the Sony is that indeed it seems to have more dynamic range than my Canons. However as Sony’s “need to be” over exposed by a stop or two don’t you kind of lose the DR advantage in way. Let’s say you expose a backlit person. Now to expose the “correct” 1-2 stops over you might have to blow out the sky. I find this really puzzling as the A7R II and D810 are really great when underexposed. I know this is not really apples to apples but still.

    Generally speaking all cameras seem to preform better when over exposed.

    Thank you for your time.

    -George

    1. If using the EI mode there is no change to the DR when using a high or low EI. What you do change is the dynamic range mid point and signal to noise ratio which can help reduce noise in the final image by exposing brighter.

  11. Last year I shot a doc for BBC4 on the FS7 using SLog2 in Custom and it worked really well and the post house was very happy with it.

    In the summer I shot a sequence for Countryfile in their preferred STD 5.

    I am about to shoot another series for an indie for the BBC in S.Gamut Cine/SLog3. They have specified using P1 (709) as the MLut. I’ve had a play and I can expose in the usual way and the pictures grade nicely.

    What I have noticed though, when comparing colour bars with these different camera settings is that black is recorded at 0.38v, peak white at 0.91v on the bars. I can see why a different gamma might be applied to the bars, as bars are meant to reflect the recorded images. However, the levels do not change between camera settings/modes. In which case, why aren’t bars recorded at the SMPTE levels on the FS7? Seems weird.

    The bars (at 0.38v) do reflect the black level of the images in STD5 & S.Cinegamut/SLog3. However, whites appear correct in the STD 5 images on a proper waveform monitor, unlike bars where maximum white is at 0.91v. A similar shot with the same lighting conditions/exposure in ungraded SLog3 never gets above 0.8v, although I would expect that of SLog3.

    So, I guess what I’m saying is that the FS7 doesn’t record standard bars AND in every gamma mode I’ve tried, including Rec 709, black is recorded almost 10% higher than standard black. Is this right? I should add that I had my camera alongside a hired FS7 last week for a multi-camera studio shoot and the output was identical to mine (apart from some magenta cast which was probably the lens).

    1. OK. Having done extensive testing today, it appears that this is a trait of my editing software and not the camera, which is why I had the same problem with the footage from both cameras. You live and learn!

      1. Alister, I’m looking at the waveform monitor in software and an external waveform monitor. Both correspond. However, as mentioned above, it appears to be a trait of how the software is interpreting the footage, not the footage itself. The way my system is usually set up, I can’t look at the material other than in software. It appears to be effectively doing a double conversion of the colourspace. Now I know, I can sort it.

  12. I get what you are saying, but since Sony has locked the s-log gamma curve to an ISO of 2000, they have locked in and effectively multiplied the underlying snr to borderline unacceptable levels. So whether it is the sensor or them locking the curve to a higher ISO, the bottom line is the image can be unacceptably noisy unless you take corrective steps. The sensor, the ISO, and and curve are all interlocked, so you really can’t separate them out and conclude it is not that noisy. It is noisy or at least noisier than it should probably be.

    1. You obviously don’t understand the article. The cumulative gain from camera to post with a like for like grade will be the same whether you use S-log2 or S-log3, so the end result is the same noise level in the final images with either curve. One is not noisier than the other, the signal to noise ratio is the same. The SNR is determined by the sensor design. ISO is a gain function not related to the sensor and gamma is added on top. The SNR is not unacceptable, not at all. Most people that shoot wisely with these cameras get great low noise results.I know I do and I don’t use any post production NR. Just expose well.

  13. My last project was shot entirely SLog3 on the FS7 in all sorts of conditions. Interiors and exteriors, natural, artificial and mixed light. Panel lights were used sparingly. There were no issues in the grade and the post house spoke glowingly of what they had to play with. I made good use of the high and low key function and took Alistair’s advice, exposing as normal with the Rec709 LUT. What I did do was shoot mostly at 1250EI and I set my zebras at 67%. There was no noise to speak of.

Leave a Reply

Your email address will not be published. Required fields are marked *

*