F3: Noise in Slog?

[yoclay]

It's interesting but speaking about being in the 1990's.
Why do you continue to think that we need a generator to have light?
LED panels often take only 50watts of power.
I often use up to 5 panels with V-mounts, works great and I have many hours of silent light coming from featherweight heads.
Lighting has changed considerably. All kinds of lighting possibilities are out there.

Not to mention the need for many more flags than before, because the issue is the highlights today, not the shadows.

 

[morgan_moore]

I was specifically commenting on that film and the lights/noise/power crew size used to make it vs the final look achieved. I do think an nex5n would have looked just as good with that light package.. and sounded just as good with a budget to ADR the whole film!

Personally Ive not found LEDs that work for me yet - but using vlock/led is certainly something I would class as 'modern' film making.

As for light modifiers - yep I have a van full of them.

S

 

[bill totolo]

Im telling you I can smell lack of depth in skin a mile off.

It is very simple that at the top of the log curve (but where?) not enough steps are allocated to get a good tonality to important parts of the image.

To me there is no reason why an F3, with HMI/Modifiers package should not look basically like a proper movie - IMO this does not.

S

As I understand it Morgan, there are more code values at the top end of the exposure curve than at the low end. I've tested over exposing middle gray up to 2.5 stops over key (in 10 bit) and have had no problem getting natural looking flesh when graded back to normal (in S-Log).

 

[bill totolo]

Back to the OP:

Here's a tip from Sony's white papers:

3.1.3 Pull Process
Reducing the camera gain will improve the signal to noise ratio performance. This technique is
suited for blue/green screen effects shots where pulling a clean key is of prime importance. It
should be noted that the camera dynamic range will be reduced.

 

[morgan_moore]

there are more code values at the top end of the exposure curve than at the low end

I dont think so. Simply (as I understand it) the flatness of the curve means there are less values .. ie a small change up in scene light level does not lift so much the digital signal - that is what a flat curve is and how it gives DR.

It is well known that sony knee (esp at 8bit) does bad things to skin tones by simply not alocating enough changes in digital tonal value to changes scene tone.

In a 10bit we have 4X more tones than needed to deliver an unbanding 709 8bit final, but one can see that once the curve is running at flatter than 1/4 gradient of the 709 curve scene tonal information is being lost. One must therefore balance exposing up enough to avoid the noise floor without pushing (as is the case in the posted film IMO) important tonality into a part of the curve that is flatter than the 1/4 gradient (actually I would say 1/3 to be safe and allow adding extra punch in post)

I've tested over exposing middle gray up to 2.5 stops over key (in 10 bit) and have had no problem getting natural looking flesh when graded back to normal (in S-Log).

So you are saying that the 1-3 gradient is available up to the 2.5 above grey - any examples? - so where would that translate to placing a skin on a WFM?

S

 

[starcentral]

I posted the code values a few pages back, with 8-bit you more or less have 256 levels and 10-bit gives you 1024.

If you look at the chart I put together, you will see that over 90% white you have (256-159) 100 levels out of 256 allocated towards the highlights. Middle grey to 90% white you have 60 levels. Middle grey down to 2% black you have 57.

I did not make up these numbers, they come from the white paper on slog and can also be calculated using the slog formula. 10-bit will give you more data to work with over 8-bit but the levels are identically and proportionately allocated.

The last column labelled waveform shows you what the final output IRE is based on the input level on the sensor itself.

This is not in stops, but this chart I put together shows exactly how much data bits are allocated throughout the range of slog on the F3 which some of you might find interesting.

Incase it is it difficult to interpret, the 8-bit and 10-bit columns show the number of bit values that are allocated to the various signal brightness ranges. So for example looking at 8-bit storage you can see the amount of data allocated between 18% grey and 90% white by subtracting 159 - 99 = 60 data bits or approximately 23% of the overall data bit container. Look at the data bits allocated over 90% white. A dramatic increase. 97 data bits, or 37% of the overall data bit container.

If you are really looking for rich blacks the only options as I see it are to;

- over-expose your ENTIRE scene (not just your subject) by 1-2 stops and then pull it down in post. This will effectively INCREASE the signal-to-noise ratio further than even the cameras own specification.

- light your shadows and pull only the dark tones in post (this is NOT crushing)

- don't shoot with slog, use a custom picture profile instead that boosts the low and mid-tones (like Cine3) and again in post only reduce the low/mid tones for a cleaner image.

 

[Grug]

On the subject of noise, I'd be very interested to see how the F3's 10-bit 4:2:2 output compares to it's 10-bit 4:4:4 output when it comes to pulling secondary keys in colour grading - this IMO is the the first area in which you start to properly run up against the limitations of the compressed 4:2:2 video signal, with significant noise being introduced and less-than-smooth keys being pulled in all but the most perfect exposures (which means that if your character is in shadow in the frame, or is struck by a hot streak of overexposure - that secondaries on their skin can pose significant challenges.

 

[starcentral]

There's been lots of debate on 444 vs 422 vs 420.

And I'm in the other boat I'm afraid, that its only really the edge resolution improves as you increase "color space" until you get to 444 which doesn't exist on the sensor to begin with.

But that threads already been started.

 

[morgan_moore]

.. 90% white you have (256-159) 100 levels out of 256 allocated towards the highlights..

Indeed - but how many should you have? Considering each stop in a scene is a doubling of light it should actually be far more than that?

Note this is a log graph..

Image at http://provideocoalition.com/images/uploads/aadams/Log vs. Raw: The Simple Story/Log_vs._lin_log.png

It is easy to see the roll off of an slog curve on many images e.g alaistair chapman site etc.

Of course we dont need full linearity because..

1) 10bits means we have 4X too much data
2) We in general want a film look and film flattens at the top

But log is certainly a compromise (or a set of them as is any camera design) and poor tonality near clip is one of them and failure to crush out noise before post is another (to some)

BTW did you check the windows a 01:00 in the XMas film (highlight compression)?

S

 

[starcentral]

Exactly how much of the data bits we should have for our highlights I simply do not know, but I can say this; full linearity (linear gamma recorded RAW) fundamentally was the ultimate reason that "digital" was able to mimic the look of film which lead to the revolution of DSLR's slowly replacing 35mm film SLR's. The Nikon D1 released in 1999 was the camera that "replaced film for the newspapers" according to Ken Rockwell. Later it was RED who adapted the exact same principle of the DSLR by recording RAW frames at 24fps.

So while linear gamma is extremely inefficient and does not go out of its way to distribute any of the data bits evenly throughout the range the proof is in the pudding - it simply works, and has worked since 1999, and now we're seeing other cameras like BMC, Sony F5/F55 and even the D16 doing the exact same thing. Thankfully disk space is getting bigger and cheaper.

But like you say log does seem like a compromise of sorts at both ends of the scale.


Regarding xmas video I saw some weird yellow highlights out the windows at 0:18 and 0:32.

 

[bill totolo]

I dont think so. Simply (as I understand it) the flatness of the curve means there are less values .. ie a small change up in scene light level does not lift so much the digital signal - that is what a flat curve is and how it gives DR.

It is well known that sony knee (esp at 8bit) does bad things to skin tones by simply not alocating enough changes in digital tonal value to changes scene tone.

In a 10bit we have 4X more tones than needed to deliver an unbanding 709 8bit final, but one can see that once the curve is running at flatter than 1/4 gradient of the 709 curve scene tonal information is being lost. One must therefore balance exposing up enough to avoid the noise floor without pushing (as is the case in the posted film IMO) important tonality into a part of the curve that is flatter than the 1/4 gradient (actually I would say 1/3 to be safe and allow adding extra punch in post)



So you are saying that the 1-3 gradient is available up to the 2.5 above grey - any examples? - so where would that translate to placing a skin on a WFM?

S

Thanks for your feedback Morgan. Your feedback, as a few others, are going a long way in helping me understand LOG exposure. I've been using LOG for a couple years but I think it's important to re-evaluate ones work habits from time to time.

For my test I used a Kodak Gray Card Plus exposed at Sony's recommended rating of 37% reflected, then continued sample exposures in 10 IRE increments up to 90%. Disregarding highlights, I felt middle gray could get back to "normal" pretty reliably up to 75IRE. After that there was degradation, mostly apparent in a side by side test. Your mileage may vary. The purpose was two-fold: to find where I like middle gray, and see how over exposure effected noise in the blacks.

My personal findings put me in the 40-45% range for middle gray at -3db (ISO 570) for the cleanest and most natural image for low light interior settings, and 40-45% at 0db for daylight exteriors.

This test assumes a color grade where middle gray is lowered and blacks restored to 2%.

 

[morgan_moore]

Im not sure what degredation you see in push or pulling a grey card - I think it is important to work on skin (even if it is a mug as ugly as mine). But interesting to hear your findings they seem to make sense.

My test set up has shadows, a hotspot , graduated tones skin and colour. The graduated tones are where one sees the inducement (or not) of banding.

Abel for example use tests that use discreet values and therefore are of minimal real world value as they pay no attention to tonal fidelity. Where the curve is too flat tonal fidelity is lost.



Edit: revisiting this image Cine1 card 50, (not Slog) one can see good tone in my face, clean powerful blacks, little noise, and a horrible fried egg highlight, I think it is quite good!

 

[bill totolo]

Im not sure what degredation you see in push or pulling a grey card - I think it is important to work on skin (even if it is a mug as ugly as mine). But interesting to hear your findings they seem to make sense.

Ha! I should have mentioned I had a person present in the images to evaluate flesh-tones, (I also tilted up to the sky to look for banding, shot with foliage in the b/g to evaluate high frequency detail, and had flowers in the b/g to see when the color red ran away with saturation). I only looked at the 10bit ProResHQ files, since that's my workflow.

At least to my mind, I've finally found the set up that's very clean and maintains all the benefits of S-Log.

 

[starcentral]

Sam, the cines are good and a lot more effort free in terms of nailing it down perfectly with slog. But many of us are convinced slog is somehow superior yet still trying to find evidence of that

 

[morgan_moore]

Indeed - Im sure it is superior too!

Basically a cine curve goes too flat at the top, breaking the 1/3 slope rule IMO (check that highlight in my test - rather than blowing the hotspot it looks like a fried egg - horrid)

Also it does retain the shadows and if you want to dig in to them (and get some noise) you can

It takes all from the chip that it can - I dont think sony are clowns.

.. but one needs
1) to know where to expose - I think higher than 38 if possible
2) a good method of delogging

S

 

[starcentral]

Its honestly really too bad that Sony never provided their own color / gamma science like RED did. It leaves much to be guessed and experimented in my opinion, and also does not help improve the marketability of their cams.

In fact right now people on Sony forums are complain why LUTS have only been released for F55 and F65 leaving the F5 behind like a fart in the wind.

 

[morgan_moore]

So there is no official lut!

It is very similar to 2005 and my Sinar stills camera, whose software was so bad we used an open source debayer, written by a German war games nut, with separate black ref files per image, basically one would wade around in green magenta hell, and then suddenly 'pop' it all came in to place after a few months fiddling and waffling on the Lula board! At least that Sinar set me up to take on any digital camera

S

 

[Barons]

Your exposure in slog will depend on what you are shooting - there is no absolute way to decide this. Your best bet is to use a waveform monitor (such as the one in the Alphatron) and use common sense - when you have a lot of sky in a scene pay a lot of attention to the highlights and keep them under 70 units. I keep most faces around 35 - 40 units. There is an amazing amount of range at the bottom end and you can raise levels easily in most situations, but you can't ever get back detail from overexposed highlights. The monitoring LUT in the Gemini or a box from BMD that allows you to monitor through a LUT - ideally the Sony LUT, is a very simple method of assuring you're very near the correct point in exposure. The Gemini LUT, which I believe CD said was a Sony LUT - presumably slog 1 - gives excellent results. It is my experience that the camera performs better in any situation in slog than with Cine gammas. Since we grade everything, anyway having a few shots in Cine gamma, even if they look fine, really doesn't save much time since our workflow is set up for log. In grading sometimes leaving the LUT in, as you can with Media Composer or Baselight, helps the grade, but in high contrast scenes it works much better to start with the straight log footage. Fyi, you can download the Sony LUTs for both slogs from the Sony website and then import them into Avid and Baselight, to name two platforms. Presumably in Resolve and Premiere as well, I've just never tried. Just make sure you have the correct type for the software you're using - Avid and Baselight take the .cube version.

 

[paulcurtis]

I'd like at add a couple of points.

One is that the curves that you record through (Be them log, cinegamma or whatever) need to have a corresponding way to get the data back to linear for grading. Number one issue. You really cannot grade properly unless you are dealing with scene linear values. (obviously you can technically grade anything but for the light to react in a real world manner you need to be basing your math in scene linear space).

There are various LUTS to convert from Slog to Linear, however there are little to no LUTs to convert from cinegammas and, more importantly, tweaked cinegamma curves into Linear. This, IMHO, is one of the reasons why we tend to say 'video' at pieces shot in cinegamma/hypergamma curves. Not because the data isn't intrinsically in there but because when that data is taken into whatever post pipeline it looses it's relationship to scene linear. Classic example is shoot a light source then bring the levels down and the levels around that light source do not decrease in a real world way because those actual brightness values are clamped to REC709 or whatever interpretation the software has made about the image. The grading software does not know how bright these values are supposed to be.

Second point is that what we do not know about is what's happening internally in the camera before we see the data out. So if there is a 14 bit ADC, then depending on how sony process the signal, we may see compromised shadows precisely *because* the overall range of the sensors are so much better - we have to squeeze more data into a bucket that isn't big enough. So in fact the F3 perhaps has an advantage because the overall dynamic range is less than the F5 but the ADC and signal processing isn't orders of magnitude better. So the shadows are better. It's impossible to know what goes on inside the camera - is it floating point, is it integer, what kind of signal processing. All i'm saying is that there is a black box involved before we see any data and we don't know how compromised that is. 14 bit range is a lot more than 12 bit but all that extra data is in parts of the image which we don't care so much about.

cheers
Paul

 

[yoclay]

I'd like at add a couple of points.

One is that the curves that you record through (Be them log, cinegamma or whatever) need to have a corresponding way to get the data back to linear for grading. Number one issue. You really cannot grade properly unless you are dealing with scene linear values. (obviously you can technically grade anything but for the light to react in a real world manner you need to be basing your math in scene linear space).

Paul

Totally disagree. Said it before, will say it again.
LUT's are shortcuts. Nothing more.
The notion that somehow mathematically it is going to give you something more ("light behaving in a real world manner") than the actual image as exposed is a falsehood.
You CAN grade properly without having delogged.
You can take any image that has not been delogged and arrive at the same image with the best LUT applied to it.
Find a black point and a white point and then go to it.
In real life you are going to tweak an image after a LUT anyhow and that can be just as time consuming. The blacks a little too crushed, a magenta cast, etc.
I find the time it takes to rebalance an image shot in LOG is so little anyhow that often I don't even bother with LUT's most of the time.