Mysterium-X dynamic range...

[Roberto Lequeux]

Good answer.

Really? I mean, sure, that sounds great, but what about film? I need a 3rd grader 'splanation methinks. Not all of us here are Brook-ish.

Does this mean that the response will be more like film where:

a. highlights have more malleability
b. things look awesome
or
c. all of the above

 

[David Mullen ASC]

What's unique about color negative film is that:
(1) it capture a lot of dynamic range -- some 15-stops perhaps; and
(2) the increase in density (the amount of silver and color dye created during processing due to "latent image" exposed on silver halide grains) is not linear with exposure, but instead falls into a S-shaped Log curve.

What this means practically is that at the extreme ends of brightness and shadow, the film gets lower in contrast because less density is forming with exposure, so the image detail gradually "rolls off" into blackness and white, rather than abruptly cuts off. Abrupt clipping of highlight detail is a particular characteristic of digital photography.

Now there are methods of rolling off the highlight more gradually (knee compression in regular video cameras, for example) but if you are going to do this, it helps to start out with more highlight information than you ultimately need to see, so you have the freedom to "compress" more exposure information into a smaller range. Ultimately though what you want is to record as much information as you possibly can so you have maximum flexibility to adjust it in post.

But I have no idea what tricks the MX sensor and processing are doing with a 12-bit A/D to get closer to this characteristic of film negative.

The F35 / Genesis sensor basically doubles up on the number of photosites needed to get 1080P RGB so that half the photosites per color can be less sensitive and/or darkened somehow in order to capture more overexposure detail. This is why the F35 / Genesis sensor is 12MP for a final 1080P RGB recording, and uses a simpler RGB striped color filter array.

The new ARRI Alexa sensor (Bayer CFA) is using some sort of "dual gain" design approach to capture more overexposure detail:
http://www.arridigital.com/teaser/sensor

Recently Kodak increased the dynamic range of their color negative stocks (the Vision-3 line) by using what I think they call "sub-micron grain" or something -- basically they figured out a way to make really tiny, slow-sensitivity grains. This was necessary in order to make a new finer-grained 50 ASA stock, but they found that the same technology could be put into their high-speed stocks so that at the extreme end of brightness where the negative would normally hit D-Max (maximum density), the tiny slow grains allow even more detail to be recorded rather than burn-out to white.

 

[Roberto Lequeux]

Thank you David for your answer.

Could a double log be implemented for digital? Would anything else ever make data distribution more "film like" than such arrangement?

 

[Rob Ruffo]

Ummm.... Yep.



Actually stops are a level of abstraction. What you're actually saying is that the scale is not linear, which you're right, it isn't. A better way to look at it is that if you have 4096 (12bit) luma values to play with and 12 stops of range, you assign 0 to the lowest value you wish to work with and 4095 to the highest and distribute the other values throughout the range. They can be linear or they can be logarithmic. As of right now, on the current Mysterium, the data is quantified in a linear fashion. As for ISO, that is yet another factor and within this system, ISO rating is mostly a function of where mid-grey is placed within the luma range.

My point was that bits and stops do not correlate directly, the bit depth of the imaging system only determines the level of precision, not the number of stops in its range.

Yes, you are 100% correct. I took "abstract" to mean an arbitray value not tied to measures fo concrete realities, the way a meter or kilometre doesn't really correspond to anything in the real world, whereas a "foot", well, is about the length of an average person's foot.

 

[Rob Ruffo]

Because dynamic range is a marketing tool, so don't believe what anyone tells you, it's nearly useless information in the abstract. If one company or person uses a method of measuring dynamic range that is different than another, then the figures are useless as a point of comparison. Plus there is a difference between recorded range and useable range, and everyone decides what is "useable" differently.

A practical example would be the recent comparisons between the Sony F35 (considered probably the king of dynamic range right now) and the new MX sensor and the new ARRI Alexa sensors -- what seems to be the result of these tests is that the dynamic range is similar but the noise is lower in the new sensors. And less noise gives you more flexibility in color-correction, so more of that range is useable. But the recorded measurable range may be rather similar.

So the only time you are going to get a realistic view of the differences between two cameras is to shoot your own tests and color-correct them, because that can also be a factor in what information is useable and able to be manipulated. In other words, the recording format can have an effect on what the sensor can output.

So Sony releasing "X" figures for dynamic range and Red releasing "Y" and ARRI releasing "Z"... well, it has little value.

An even when an outside party does their own comparison test, even that has to be taken with a grain of salt. Their testing methodology could be potentially flawed or simply represents their own personal way of working with digital images. But at least a side-by-side comparison does make certain differences more immediately visible than specs released by manufacturers, which is just a different form of information to be weighed.

On the other hand, common sense tells you that it is unlikely for an 8-bit prosumer HDV camera to capture the same dynamic range, with the same ability to be manipulated, as better cameras and recording formats working at a higher bit-depth and with better compression schemes.

Yes, this is true - workflow and creative approaches to grading, in the end, determine dynamic range. Here in Montreal a friend has developed custom tools for noise reduction. These work extremely well, using a form of AI to perform their magic. Dynamic range for him is completely different from dynamic range for someone not using these tools, as a camera exhibiting lower-IRE noise that is friendly to how his noise reducer works will give him way more lattitude in practice than a camera that is not, even though that second camera may have more "measured" dynamic range.

 

[Rob Ruffo]

Hey David, you or anyone know if the MX is rumored to solve the "half-frame from flashes" artifact?

These are extremely easy to fix using rather simple post algorythms

 

[Rob Ruffo]

What's unique about color negative film is that:
(1) it capture a lot of dynamic range -- some 15-stops perhaps; and
(2) the increase in density (the amount of silver and color dye created during processing due to "latent image" exposed on silver halide grains) is not linear with exposure, but instead falls into a S-shaped Log curve.

What this means practically is that at the extreme ends of brightness and shadow, the film gets lower in contrast because less density is forming with exposure, so the image detail gradually "rolls off" into blackness and white, rather than abruptly cuts off. Abrupt clipping of highlight detail is a particular characteristic of digital photography.

Now there are methods of rolling off the highlight more gradually (knee compression in regular video cameras, for example) but if you are going to do this, it helps to start out with more highlight information than you ultimately need to see, so you have the freedom to "compress" more exposure information into a smaller range. Ultimately though what you want is to record as much information as you possibly can so you have maximum flexibility to adjust it in post.

But I have no idea what tricks the MX sensor and processing are doing with a 12-bit A/D to get closer to this characteristic of film negative.

The F35 / Genesis sensor basically doubles up on the number of photosites needed to get 1080P RGB so that half the photosites per color can be less sensitive and/or darkened somehow in order to capture more overexposure detail. This is why the F35 / Genesis sensor is 12MP for a final 1080P RGB recording, and uses a simpler RGB striped color filter array.

The new ARRI Alexa sensor (Bayer CFA) is using some sort of "dual gain" design approach to capture more overexposure detail:
http://www.arridigital.com/teaser/sensor

Recently Kodak increased the dynamic range of their color negative stocks (the Vision-3 line) by using what I think they call "sub-micron grain" or something -- basically they figured out a way to make really tiny, slow-sensitivity grains. This was necessary in order to make a new finer-grained 50 ASA stock, but they found that the same technology could be put into their high-speed stocks so that at the extreme end of brightness where the negative would normally hit D-Max (maximum density), the tiny slow grains allow even more detail to be recorded rather than burn-out to white.

I was wondering how that worked. Thank you. Fascinating.

 

[David Mullen ASC]

Thank you David for your answer.

Could a double log be implemented for digital? Would anything else ever make data distribution more "film like" than such arrangement?

I think the simple solution (when the technology permits) would be to record more linear dynamic range at higher bit depths -- then you can monkey around all you want in post with the gamma curves and how and when the highlights clip. Someday we'll have a 16-bit camera that captures 16-stops and we can emulate a film negative's characteristic curve and full dynamic range quite successfully if we wanted to.

But the truth is that current dynamic range of the top-end digital cine cameras is quite workable and things will only get better from here on out.

 

[Roberto Lequeux]

I guess higher bit-rates would do the same, but a log curve makes better use of the file size, right? And since there is no need to go back to the recorded codec there is no harm. It would however be heaver processing, but it would be nice to know you are getting the most out of file size regardless of how big drives get... though you can't focus on everything and if bigger media comes along this would be mostly a waste of effort... mhhh... I can't wait to see what Red is holding back on us.

 

[Jeff Kilgroe]

A log curve doesn't have any impact on file size, it's just how you "weight" your data. Increasing your bit rate doesn't increase DR, it can just allow for lower compression to help retain more image detail that would otherwise be thrown away by the compression algorithms. Instead of recording linear information, you could use a curve to give you higher precision in one end of your range vs. another. You could decrease precision in the mid-tones while increasing precision in the highlights to give a smoother roll-off appearance.

I agree with David Mullen about recording linear data, or more of it for more range at higher bit depths, which would give greater precision and control. Back when RED first announced the Monstro sensor and gave us preliminary details, they were talking about 16bit REDCODE. All things subject to change and all that, but this would be a huge jump in the recorded precision. 12bit is 4096 levels on a luma scale. 16bit would be 65536 values or 16 times the precision.

RED is not gong to be the only one pushing this hard and fast on new sensor tech over the coming months / years. Other camera makers will have to step up their game as well. And it will be a good day when they do. Sensor technology still has a long way to progress until we reach the theoretical limits of what is possible.

I think it's pretty remarkable what is being done with approximately 12~13 stops of DR recorded with 12bit precision. Just imagine having yet another stop of DR and and all of it recorded as 16bit RAW. Oh, my...

And obviously higher bit-rates and more internal bandwidth is necessary to handle 16bit vs. 12bit. 16bit is 33% more data than 12bit.

 

[David Mullen ASC]

Log curves make sense for recording a lot of dynamic range on a 10-bit recording format, but once you start talking about 12-bit, 14-bit, 16-bit recordings, you might as well just record the linear RAW output of the sensor, which will contain everything the sensor can capture, and just change the gamma in post as desired.

 

[Roberto Lequeux]

Doesn't linear give WAY more emphasis to the blacks? As a matter of fact I thought that was the actual reason for digital's highlight issue...
Or is it just dynamic range limitations and the need for thick mids?

 

[Eki Halkka]

Doesn't linear give WAY more emphasis to the blacks? As a matter of fact I thought that was the actual reason for the highlight issue.

It's the opposite - linear gives much more precision in the highlights than in the shadows. Half of the available brightness values are spent on the one brightest stop, half of the remaining values to the next stop and so on.

There's 4096 brightness values total in 12 bit linear, assigned as follows if i've understood correctly - one stop each:

White

2048
1024
512
256
128
64
32
16
8
4
2
1

Black

 

[Roberto Lequeux]

Ok, I see. So more emphasis doesn't sound like a smart thing to do, and linear sounds very suitable for digital. Still, we are always pulling out of the shadows. I guess this has to do with the dynamic range of the human eye in relation to the range of digital and film. How many stops do we perceive at once? Nevermind. I just realized, late, that I have once again rudely pestered the community for my personal gain. My sincere apologies. Thank you to everyone. Especially to Jim and Co. who stir me up and get me sidetracked with such fancy tools. Can't wait to use this beast of a time machine!

 

[Jeff Kilgroe]

Most of it has to do with digital sensor technology and how sensors work. Current sensor are more sensitive to overload or blowing out as they receive more light. So there is such a fine line between something bright and something that overloads. We are digging out of the shadows all the time because that is how we protect our twitchy, irritable sensors from clipping. The only way to really fix this is to improve sensor response and DR while recording at higher bit depths to allow for more accuracy and control.

Eki is right about values in relation to stops, but that's only because stops are a logarithmic scale of their own, so they do not relate evenly to a linear luma gradient, which is how the sensor and the digital data interpret the light. Which is fine, stops are only one way of measuring light transmission. It's one that camera people are comfortable with and understand. Or at least many think they do. The new FLUT control is directly related to stops, so increasing or decreasing the FLUT value in REDCINE-X by one whole number seems to be like adjusting exposure or ISO up or down by a whole stop.

As digital sensors progress and DR increases, we will find that adding "more stops" of DR will become an increasingly slower process. Each time a stop of DR is added, the effective range of the sensor is doubled. So if the upcoming Monstro sensor gains half a stop over M-X, it has effectivly 1.5X as much DR.

I've attached a quick and dirty graphic that shows a luma ramp for 12bit and how it would relate to 12 stops of DR. It's not 100% accurate, but shows how luminance and digital representation of it are linear, but thinking in stops is not.

 

[Stephen Williams]

So if the upcoming Monstro sensor gains half a stop over M-X, it has effectivly 1.5X as much DR.
.

Hi,

I would describe the difference between 13 stops & 14.5 stops as less than 12% increase in DR.

Stephen

 

[Daniel Browning]

I would describe the difference between 13 stops & 14.5 stops as less than 12% increase in DR.

Would you describe the difference between 1/48 and 1/108 as less than 12% increase in shutter speed? How about ISO 500 to ISO 1125? T/2 to T/1.2? Those are all only 1.5 stop increases like 13 -> 14.5 stops. It is a 12% increase in the number of stops, but the increase in Dynamic Range is far greater than that.

 

[Michael Lindsay]

These are extremely easy to fix using rather simple post algorythms

How?

If a strobe fires and splits a frame you will often have a exposure where there is none in the normally exposed image part frame. Matching the over exposure is easy(ish) in the split frame but seeing the unseen seems impossible??

regards

Michael L

 

[Stephen Williams]

Would you describe the difference between 1/48 and 1/108 as less than 12% increase in shutter speed? How about ISO 500 to ISO 1125? T/2 to T/1.2? Those are all only 1.5 stop increases like 13 -> 14.5 stops. It is a 12% increase in the number of stops, but the increase in Dynamic Range is far greater than that.

I would describe the difference as 1.5 stops, for somebody that shoots a lot of film it's not a big deal, comfortably within the latitude of modern negative stocks.

 

[Todd M.]

I would describe the difference as 1.5 stops, for somebody that shoots a lot of film it's not a big deal, comfortably within the latitude of modern negative stocks.

Does that mean 12%? :devil: