SLOG ISO ratings?

[David C. Williams]

Look at the slide I posted, directly from Sony's engineers S-Log whitepaper. It clearly shows the code values incrementing in a linear fashion.

If the video you linked says the opposite about S-Log, I'd believe the engineers. I'd also believe my own eyes when grading F3 S-Log in Resolve. There is A LOT of information in the highlights you can recover.

In my experience, the Sony marketing departments and Sony engineering departments don't seem to communicate very well. The marketing department tends to make technical errors as they are not engineers and do not appear to run things past people who know often enough.

[Douglas Villalba]

Originally Posted by David C. Williams

I'd believe the engineers. I'd also believe my own eyes when grading F3 S-Log in Resolve. There is A LOT of information in the highlights you can recover.

+1 You can only recover on very few color grading programs like Resolve, Color, etc,.

Edit: I shoot 10 bit 422. I assume 444 is better and 420 can't be as good.

[Alister Chapman]

I think your forgetting some key points David.

When looking at the Sony S-Log white paper, the relationship between the data bits used and f-stops appears linear, because it is. When looking at the graphs in the Sony white paper you need to understand that the horizontal scale is a log scale, not a linear one, so the line on the graph appears straight, because the scale is log. F-Stops are logarithmic. Replace the scale with a linear scale and you would see the familiar log curve. Because the scale is log the line appears straight, but what you are seeing is a non linear relationship between illumination and recording data.

In the video, during my presentation I clearly state that log works by using roughly the same amount of data bits per stop. Exactly as per the graph you are referring to. BUT cameras and monitors are not log devices, they are linear. The scale in the chart is a log scale which is why the line looks straight. If you replaced the log scale with a % scale or lux scale you would see a very highly curved, very non-linear response curve.

S-Log does compress highlights more aggressively than standard gammas and cinegammas. It has to, that's how you squeeze a greater dynamic range into the same sized bit bucket. There is no free lunch, if you want to fit something bigger into the same box, your going to have to squash it more. In order to squeeze more dynamic range into the same bit bucket with S-log the onset of compression is much earlier than standard gammas and cinegammas and the earlier onset of this compression means that accurate exposure is more important and is also part of the reason why middle grey is shifted lower down the exposure range, to keep the important stuff away from the more compressed parts of the curve. Adding and extra stop and a half of dynamic range requires 200% more compression over the cinegammas and hypergammas gammas. Yes there is a lot of information in the highlights of an S-Log image that can be extracted, but if it's anything like skin tones or natural textures, things the human visual system is good a noticing problems with, you will definitely pick up on the compression.

First, camera sensors and monitors are linear devices, but the f-stop scale is a log scale and finally our visual system is essentially a log system.

In a perfect world we would simply take the full linear output from the camera and feed it in to an equally linear monitor and bingo, we would have a very accurate representation of the scene being filmed that to our eyes would look completely true to life.

Sadly though it's not as simple as that as our own visual system is not linear, it is logarithmic, like the f-stop scale. So we don't perceive a significant brightness change in an image unless you DOUBLE the brightness. With linear devices like a video monitor or TV, to make the image look brighter you must double the output, which requires double the data. As we work in a world where there is a finite limit on how much bandwidth or data that is available this is actually very difficult to do, especially if you want a really large brightness range.

IMPORTANT: The numbers given in the examples below are for illustration only, they are not accurate and do not represent actual data or exposure levels, thet are just there to illustrate the principle behind log.

The F-stop scale is logarithmic. Each additional stop of exposure requires twice as much illumination as the previous stop and doubles the tonal range of the image. So if it takes 2 light bulbs to provide the first stop of illumination, then it would take 2 more light bulbs to increase the exposure by 1 stop. Four light bulbs for 2 stops, 8 lightbulbs for 3 stops, 16 light bulbs for 4 stops, 32 for 5 stops and so on. 10 stops would need 1024 bulbs!

Remember a camera is a linear device. To directly record the output from the sensor, if each lightbulb takes one bit of data to record, you would only need 512 bits of data to record 9 stops, but would need twice that, 1024 bits to record 10 stops and a whopping 4096 bits to record 12 stops. So as you can see for a camera, a linear device, to record ever greater exposure ranges requires massive amounts of extra data.

Lightbulbs/sensor data 2-----4-----8-----16-----32-----64-----128-----256-----512-----1024-----2048
F-stop-------------------1-----2-----3------4------5-------6-------7--------8-------9-------10---------11

You can see from this crude example that if you are using an 8 bit recording system with only 240 bits of data to play with, that recording anything beyond about 9 stops requires something different (the knee).

What you must remember is that as each additional stop doubles the tonal range of the image, if you only allocate the same amount of data to each stop, then as the image gets brighter then progressively less and less data is being used to record the linear output from the camera.

Lightbulbs/sensor data 2-----4-----8-----16-----32-----64-----128-----256-----512-----1024-----2048
F-stop-------------------1-----2-----3------4------5-------6-------7--------8-------9-------10---------11
S-Log data bits--------10----20----30----40-----50-----60------70------80------90------100------110

Taking the first stop, the sensors output uses 2 bits of data. With S-log we have 10 bits available to record it, no problem.
Taking the 10th stop, the sensors linear output uses up bits 512 to 1024, 512 bits of data, yet with log recording, allocating roughly the same amount of data to each stop, we squeeze that into only 10 bits of recorded data. So, while the first stop had an excess of data, the brighter stop can only be recorded by discarding sensor data. This is the key to the way log works, as you go up the exposure range, more and more sensor data is discarded, thus more and more tonal information is discarded in the brighter parts of the image.

Conventional gammas like 709 are nearly linear, so almost twice as much data is used to record subsequent stops. It must be remembered that video camera sensors, TV's and monitors are linear devices. For them to accurately portray the real world then you would ideally just take the full linear feed from the camera and simply send that to the monitor. This would however require a massive amount of bandwidth or data for anything beyond about 6 stops (because each extra stop of brightness would need DOUBLE the amount of data). That's why Rec-709 is only actually designed for 6 stops, because it is a linear gamma curve (although most camera manufacturers cheat and provide a Rec-709 compatible gamma that gives around 8 or 9 stops) Because us humans have a visual system that is essentially log we can cheat the system and record using highlight compression (reduced information in the highlights) and the viewer will not in most cases notice this and this is where cinegammas, hypergammas and log come in to play, discarding data in the highlights.

If your still not convinced take a look at the unprocessed linear raw output from a DSLR or camera like the BMCC or Epic. What you will see even with a correctly exposed image is a very, very dark image with perhaps a few highlights here and there. That's because the vast majority of the image data is getting used to record the high lights and very brightest parts of the scene so the bulk of the image (the mid range etc) ends up way way down the data range making the image look very very dark. With the new F5/F55 with 14 stops of DR and 16 bit linear raw your mid tones will have around 2,000 bits of data per stop while your highlights may use a whopping 30,000 bits.

422, 444, 420 make absolutely no difference to dynamic range. With a RGB 444 recording it will be slightly easier to make large post production colour shifts than 422 or 420. The difference between 422 and 420 in progressive is extremely small and makes very little difference to grading, it only really effects chroma resolution. The colour space is the same for 444, 422 and 420.

[Mark Kenfield]

Who needs film school when you've got Alister?! Thanks again for clear, detailed and incredibly informative post mate.

[David C. Williams]

Originally Posted by Alister Chapman

I think your forgetting some key points David.

S-Log does not compress highlights more than regular gammas. Easily testable in Resolve. I use it to my advantage everyday.

Of course the scale is logarithmic, that's why it can allocate the bits evenly. Seriously, that is the only reason log gammas exist, to allow even bit distribution rather than reducing bit distribution in the highlights.

[Douglas Villalba]

Originally Posted by David C. Williams

S-Log does not compress highlights more than regular gammas. Easily testable in Resolve. I use it to my advantage everyday.

Of course the scale is logarithmic, that's why it can allocate the bits evenly. Seriously, that is the only reason log gammas exist, to allow even bit distribution rather than reducing bit distribution in the highlights.

Regardless of what the numbers may tell you, I get more out of the highlights than I can get out the shadows.

Miami Sea Port 001 on Vimeo

[Alister Chapman]

Originally Posted by David C. Williams

S-Log does not compress highlights more than regular gammas.

Then please explain how exactly do you fit an additional 3 stops of dynamic range into the same size file?

You only have to look at an ungraded S-Log scene to see the compression. S-Log looks flat because something that might be twice as bright as something else in the real world, only looks a little brighter after recording with S-Log because fewer bits are being used to differentiate one stop from the next compared to a standard gamma as a result the brighter stops are brought closer together, the contrast is reduced and the image looks flat. That is an effect of log compression.

The file size for an S-Log is exactly the same as it is with standard gammas (assuming the same codec/bit depth is used). Standard gammas give 9 stops, cinegammas 11.5 stops and S-Log 13 stops. In all cases there is no surplus or spare data, the file size does not change, so how do you think the engineers managed to squeeze 13 stops instead of the original 9 stops of a standard gamma into the same sized file? Answer: More compression. Do you compress the important mid range and skin tones or do you compress the highlights where people tend not to notice issues? Answer: Highlights. It turns out that if you use log based compression this works well as it follows the log response of the human visual system. But you MUST remember that the camera and the TV/Monitor are linear devices, it is only the bit in between that is log. So you record using log, but then have to convert that log back to linear (or at least a near linear standard power law gamma) in post to make it look right. So now what's going on with your brighter stops? Your taking your log sized 70ish bits per stop and expanding that to the required 200ish bits for the brightest stops. There's a shortfall of 130ish bits of data, data discarded during capture, lost picture information, but provided it's in the highlights it doesn't really matter as our own visual system won't notice highlight issues. Over expose while shooting however and then you can run into issues such as plastic looking skin etc.

If as you assert, log is not adding compression then there would be no need for linear raw. But if you have ever worked with linear data you will know exactly how much more picture information there is in the highlights in linear data than there is in log. You will know that with linear you can expose a face anywhere in the exposure range and provided it isn't clipped it will grade perfectly. This is definitely not the case with log and we have seen plenty of examples on this board of exactly this. Over expose a face and it can be very difficult to make it look nice. While you can grade the face to a normal levels, the skin looks odd if it's been over exposed, why? Because the extra compression in the higher stops means that you loose the subtle textures that make a face look real and not like plastic.

If you still think I'm wrong then you might want to read through Adam Wilts article on log http://provideocoalition.com/aadams/...amma_curves/P3

If you look at the plots in the Abel cine videos http://blog.abelcine.com/2011/08/04/...on-the-charts/ you will see that between 40ire and 104ire with Cinegamma one there are 4 stops, with S-Log there are 5.5 stops, so an extra 200% of dynamic range has been squeezed into the same space, the highlights must be more compressed to do this. Consider that compared to Rec-709, Cinegamma one already uses a lot of additional highlight compression and it should be plain to see that the S-log does use a lot of highlight compression. If you compare the gaps between the steps for a standard gamma plot (with the knee off) and S-log you will see the gaps between the brighter exposure steps are much much bigger with the standard gammas than S-log, because the S-log highlights are compressed, the compression bringing the steps closer together. It's only by bringing the steps closer together that you can make space for the additional stops of dynamic range.

[Bruce Schultz]

Just a small aside, I recently bought one of these middle grey/ white flex cards and now it is easy to meter both grey and white by flipping the thing over. Folds up to about the size of an iPad mini.

Lastolite EzyBalance Gray Card LL LR1250 B&H Photo Video