Meaning of 16.5 stops of Dynamic Range

[Jan Reiff]

Where did you have to set each camera (t-stop and/or shutter angle) to get them to overexpose to the same degree?

we are still testing - in this case i can´t remember, but to avoid clipping with MX for example was a f11 1/3 and Dragon was f8 1/3

i can tell you more monday, 2morrow is finally sunny day here and this will help bit, better than rain ;-)

 

[Gunleik Groven]

Take something overexposed on both cams with the same lens. Controlled lights is better if you cannot shoot them side by side. Stop down in 0,5 stops increments until they are both within range. That should tell you a lot.

You can do a similar thing at the bottom.

And you probably have the difference info you are curious about.

Do not change exposure with metadata. Only ND, or aperture (or shutter, as long as you are within blackcalibration range)

Anyhow... It is about to get a bit academic...

 

[Brian Boyer]

Brian I don't think you understand the difference between total DR and usable DR. Alexa has around 18 stops total DR but most of that is washed out in highlights so you typically get around 14.5 stops usable. If Dragon claims 16.5 stops, is that total or usable? The designation makes a big difference. Sure it might have a total of 16.5, but 14.5 usable. Therefore to claim 16.5 is inaccurate, otherwise Arri could claim 18.

Sure it would look washed out, but that is all usable info, therefore from that point you bring in contrast and grade rather than having to dig out info and apply noise reduction. I think what were all saying is the 16.5 stops cannot be usable DR but rather total.

I don't really care, just want an honest answer backed up by evidence. Anyone can make a claim. In the end I'm still choosing to shoot Dragon, but industry wide people scoff at RED for hype. It's one of the major reasons a lot of pros stay away from it.

Again, I'm shooting Dragon this week so I'll be able to test it in an exterior situation.

Matt, I understand the difference. I'm not sure how what I wrote makes you think otherwise.

I thought you were saying you wanted to see an image that showed the entire usable dynamic range, without modification.

My point was if your usable dynamic range is more stops than your display can reproduce, those stops are going to have to be compressed to fit the display in order to see them all. That means throwing away stops on the top and/or bottom if you're not going to go in and reclaim that info (what I thought you were referring to as "correction").

I don't consider dialing back non-clipped highlights in post a correction. To me, that's usable DR, provided you actually choose to use it. You're accessing info the sensor captured that can't be displayed without help (via masks, power windows, etc.) due to the limitations of your display device.

I don't see how you can get around that and still have a sufficiently contrasty image with everything the sensor captured, especially on the top end with 7 or so stops above middle gray from both Dragon and Alexa. If you can, I'd sure like to see it when you're done with your tests.

Of course, noise reduction isn't just used for digging out info in the shadows but it can certainly be used for that purpose.

Image-wise, one man's noise is another man's character. Where you draw the line will differ from where someone else draws it, including RED. I don't see how you can get around that, either.

 

[Paris Remillard]

But if you go 16.5 stops above your accepted noise floor on Dragon, you'll have a couple of stops of unrecoverable highlights. That's the point being made. There's nothing to dial back, as you're saying.

What's needed is another test like this one, as I think someone pointed out either in this thread or another:
http://www.cinematography.net/valvula.html

If both the Alexa and Dragon are exposed for the shadows in a similar scenario to that test and Dragon can hold two and a half more stops in the highlights, or conversely if you expose both to be just under clip and the Dragon has 2.5 more stops on the bottom, then we'll stop talking about the discrepancy.

 

[Brian Boyer]

But if you go 16.5 stops above your accepted noise floor on Dragon, you'll have a couple of stops of unrecoverable highlights. That's the point being made. There's nothing to dial back, as you're saying.

Paris, you're confusing my argument for one that's claiming there's 16.5 stops of dynamic range on Dragon. I haven't claimed any number of stops nor am I arguing for any number.

Matt said this (emphasis is mine): "I don't consider digging out noisy blacks a part of the total DR since it requires major post work to be usable. Same for extreme highlights that need to be dialed back."

That has nothing to do with whether or not Dragon has 16.5 stops usable or total. My point is if Dragon only has 13.5, you're not going to see all 13.5 on a display that doesn't show 13.5 without going in and dialing the upper end of that 13.5 back to within range of what your monitor can display.

In digital audio, you can record an orchestra in 24 bits @ 96 KHz. If you plan to release that recording on CD, you're going to have to reduce the data to fit into 16 bits @ 44.1 KHz because that's the limitation of that medium. It's unavoidable.

When I say dialing back highlights, I mean recovering detail the sensor captured before clipping that you're simply not seeing, preferably without having clipped at all. Maybe to you, dialing back means lowering the brightness of clipped highlights. I don't know. Perhaps we're getting hung up on semantics.

 

[Gavin Greenwalt]

Dynamic range of your monitor is irrelevant. If your sensor has 30 stops of dynamic range you could see all of it on an 8 stop (bit) display. Just like I could take a 32 bit (stop) cg rendering and see it on a 6 bit cheap cell phone display. You don’t even need any terribly complex postprocessing. The default would be what looks like an underexposed image in most shooting scenarios. By moving the gray point back you would end up with a curve with more and more compression in the highlights. It wouldn’t look washed out at all. I regularly have images with more than 20 stops of image data when rendering. Just set your approximate white point and softclip.

 

[Brian Boyer]

Dynamic range of your monitor is irrelevant. If your sensor has 30 stops of dynamic range you could see all of it on an 8 stop (bit) display. Just like I could take a 32 bit (stop) cg rendering and see it on a 6 bit cheap cell phone display. You don’t even need any terribly complex postprocessing. The default would be what looks like an underexposed image in most shooting scenarios. By moving the gray point back you would end up with a curve with more and more compression in the highlights. It wouldn’t look washed out at all. I regularly have images with more than 20 stops of image data when rendering. Just set your approximate white point and softclip.

Dynamic range of the monitor is not irrelevant given the point I thought Matt was making. Maybe I should be talking in terms of "contrast ratio" instead. In Arri's literature, they use contrast ratio and dynamic range interchangeably. I'll take the hit if I'm screwing that up.

In any event, you can only see as much of those 30 stops as your 8 bit display will allow. You say it yourself:

"By moving the gray point back you would end up with a curve with more and more compression in the highlights. It wouldn’t look washed out at all." This throws away data and is exactly how display devices overcome their limitations - they show what they can and inconspicuously omit the rest. You're not seeing everything there is to see, and that was the point of Matt's single, usable DR image, or so I thought.

It's compressing those highlights because the display device can't reproduce their full range. If it could, there'd be no need to compress them. Compression generally implies reduction.

Maybe I'm wrong and the monitor isn't the limiting factor for why you can't SEE the entire DR on the screen without some sort of intervention. One thing I'm convinced of:

If you want to SEE all the brightness values the sensor captures without losing data through progressively compressing the highlights and/or crushing the blacks, which is what I thought Matt was suggesting in his single image request, you currently have to do it through masking and power windows and the like in programs that support it. There'd be no way for him to avoid it.

If what I'm saying is so off base, someone please point me to an unprocessed image that shows an entire 14 stop range, such that all you have to do is throw a contrast curve on it and see everything that was captured by the sensor without having to isolate and separately manipulate a thing, because that's what I thought Matt was asking for and I don't think it exists.

 

[Matt Ryan]

Dynamic range of the monitor is not irrelevant given the point I thought Matt was making. Maybe I should be talking in terms of "contrast ratio" instead. In Arri's literature, they use contrast ratio and dynamic range interchangeably. I'll take the hit if I'm screwing that up.

In any event, you can only see as much of those 30 stops as your 8 bit display will allow. You say it yourself:

"By moving the gray point back you would end up with a curve with more and more compression in the highlights. It wouldn’t look washed out at all." This throws away data and is exactly how display devices overcome their limitations - they show what they can and inconspicuously omit the rest. You're not seeing everything there is to see, and that was the point of Matt's single, usable DR image, or so I thought.

It's compressing those highlights because the display device can't reproduce their full range. If it could, there'd be no need to compress them. Compression generally implies reduction.

Maybe I'm wrong and the monitor isn't the limiting factor for why you can't SEE the entire DR on the screen without some sort of intervention. One thing I'm convinced of:

If you want to SEE all the brightness values the sensor captures without losing data through progressively compressing the highlights and/or crushing the blacks, which is what I thought Matt was suggesting in his single image request, you currently have to do it through masking and power windows and the like in programs that support it. There'd be no way for him to avoid it.

If what I'm saying is so off base, someone please point me to an unprocessed image that shows an entire 14 stop range, such that all you have to do is throw a contrast curve on it and see everything that was captured by the sensor without having to isolate and separately manipulate a thing, because that's what I thought Matt was asking for and I don't think it exists.

Have you shot Alexa? Just look at a RAW LogC image and you'll see 14 stops (given that the scene utilizes that range).

 

[Hrvoje Simic]

I don't consider digging out noisy blacks a part of the total DR since it requires major post work to be usable. Same for extreme highlights that need to be dialed back.

It's likely a semantics thing but total DR is exactly that.
Its usability depends not only on the camera but the people using it to capture and manipulate the imagery.

My thought is if the claim is 16.5 stops, then it should be available in a single shot without any corrections in post. That means above the noise floor and below the clipping point.

If the range is measured it is available in a single shot, just may not be directly visible and needs a proper gamma transform.
Noise has a roll-off and DR judgement depends on subjective criteria and individual noise tolerance.

 

[Hrvoje Simic]

1) As far as I know, there isn't a monitor capable of showing 12 stops of dynamic range, let alone 16.5, without post manipulation of the image. Even our eyes "cheat" by irising up and down as needed and can't display the entire range we're capable of seeing in one go.

Any typical monitor can show higher DR captured information re-mapped to the DR and bit depth of the display.
DR information can be displayed but not necessarily in the same DR.

Post manipulation of unprocessed image is required anyway.
OLEDs can also display the image in roughly the same DR just at lower luminance.

2) Because of number 1, wouldn't an image showing 16.5 stops be extremely washed out, making it difficult to judge the separation between values by eye? Since you're concerned about usability, who would realistically use that image in such a state? In the end, without manipulation, you're going to end up throwing away DR in order to fit the final image inside the limitations of current display technology.

Any gamma uncorrected image would look washed out.
16,5 stop scene is a high contrast scene and once typically gamma corrected should appear that way.

 

[Jan Reiff]

i post a new R3D in the other thread ...

http://www.reduser.net/forum/showthread.php?116216-More-Dragon-Samples-with-R3D-stills

 

[Brian Boyer]

Any typical monitor can show higher DR captured information re-mapped to the DR and bit depth of the display.
DR information can be displayed but not necessarily in the same DR.

Post manipulation of unprocessed image is required anyway.

This is what I thought I was saying, just in 100 times more words. ;-)

I like your terms better.

Any gamma uncorrected image would look washed out.
16,5 stop scene is a high contrast scene and once typically gamma corrected should appear that way.

This is what Matt originally said and why I responded the way I did (emphasis is mine):

My thought is if the claim is 16.5 stops, then it should be available in a single shot without any corrections in post. That means above the noise floor and below the clipping point.

Initially, he specified uncorrected. I took that to mean no curves, no LUTs, no nothing. I couldn't understand what good an uncorrected, washed out image alone would do in determining DR. It's not as if you can tell how many stops an image has just by looking at it, especially one without much contrast. If you look at it that way, most of my rambling kinda makes sense.


He later amended that statement by saying (emphasis is mine):

Sure it would look washed out, but that is all usable info, therefore from that point you bring in contrast and grade rather than having to dig out info and apply noise reduction.

As it turns out, he didn't actually mean uncorrected in the way that I thought. Which makes my line of reasoning seem rather off-the-mark now, but also answers my question of how a washed out, uncorrected image alone can aid in conclusively determining DR - it can't.

I should've picked up on the change in wording sooner but it was late and I hadn't eaten in like 12 hours.

Have you shot Alexa? Just look at a RAW LogC image and you'll see 14 stops (given that the scene utilizes that range).

I've never shot Alexa but I've played with the raw files somewhat.

I'm having a difficult time finding a .ari or Log C rendered DPX that shows the type of situation one might typically encounter where some masking or power windowing would be called for to preserve highlights - say, shooting indoors with daylight illuminated detail outside a window you want to keep.

Maybe you know of an image that illustrates this you can point me to? If Log C renders masking and power windows to bring out details in highlights unnecessary, it's news to me and I'll take the hit for being wrong on that one, too.

 

[luigivaltulini]

I've never shot Alexa but I've played with the raw files somewhat.

I'm having a difficult time finding a .ari or Log C rendered DPX that shows the type of situation one might typically encounter where some masking or power windowing would be called for to preserve highlights - say, shooting indoors with daylight illuminated detail outside a window you want to keep.

Maybe you know of an image that illustrates this you can point me to? If Log C renders masking and power windows to bring out details in highlights unnecessary, it's news to me and I'll take the hit for being wrong on that one, too.

Arri have Ftp with file demo.

http://www.arri.com/camera/alexa/learn/alexa_sample_footage/

 

[Tommaso Alvisi]

Arri have Ftp with file demo.

http://www.arri.com/camera/alexa/learn/alexa_sample_footage/

Thanks Luigi, much appreciated! I'll have a look.

 

[Brian Boyer]

Arri have Ftp with file demo.

http://www.arri.com/camera/alexa/learn/alexa_sample_footage/

Thanks Luigi. I've been to that page before but forgot all about it.

 

[Gavin Greenwalt]

It won't look washed out though if you compress the range. In fact the easiest way to get 16 stops of dynamic range are a number of stops in the shadow detail. When you render an image in a 3D package you have as many stops of detail as you have bits of rendering (and then some) since the precision is so high. If I showed you a "normal" rendering it could have immense detail in the shadows (with no noise) but look like a normal image essentially without any manipulation.

A really easy example of this is the simple gradient. If you created a linear gradient from black to white in a 32 bit canvas you would have 32 "stops" worth of gradation. If you applied some dithering and reduced the *precision* you would still have "32 stops" worth of dynamic range, it would just be a lower accuracy with some dithering and posterization (this is how Nuke's viewer works, it's a 32 bit frame buffer with an 8 bit real-time dithering). But even after converting a 32 bit gradient to an 8 bit gradient, with sufficient dithering it will look subjectively the same (maybe a little more grainy) it's *not* going to wash it out or change the image.

The reason we tend to associate localized tone mapping or log gamma curves with high dynamic range imagery is because the real world generally has a split. There is tons of information in "0 - 10" stops of detail from the average reflectivity of objects, and then usually there is just a few light object which are stops like 16-20. It's that large gap between a light source and a diffuse reflector that makes grading a high dynamic range image tricky. The metaphor for that situation would be like a gradient that goes from 0 - 75% and then just 75% until 100%. It's not the display's fault, it's just a bad gradient (subjectively speaking) However, there are plenty of other examples of imagery where you can have 32 stops of dynamic range like the initial example without a gap.

 

[Brian Boyer]

Dynamic range of your monitor is irrelevant. If your sensor has 30 stops of dynamic range you could see all of it on an 8 stop (bit) display. Just like I could take a 32 bit (stop) cg rendering and see it on a 6 bit cheap cell phone display. You don’t even need any terribly complex postprocessing. The default would be what looks like an underexposed image in most shooting scenarios. By moving the gray point back you would end up with a curve with more and more compression in the highlights. It wouldn’t look washed out at all. I regularly have images with more than 20 stops of image data when rendering. Just set your approximate white point and softclip.

It won't look washed out though if you compress the range. In fact the easiest way to get 16 stops of dynamic range are a number of stops in the shadow detail. When you render an image in a 3D package you have as many stops of detail as you have bits of rendering (and then some) since the precision is so high. If I showed you a "normal" rendering it could have immense detail in the shadows (with no noise) but look like a normal image essentially without any manipulation.

A really easy example of this is the simple gradient. If you created a linear gradient from black to white in a 32 bit canvas you would have 32 "stops" worth of gradation. If you applied some dithering and reduced the *precision* you would still have "32 stops" worth of dynamic range, it would just be a lower accuracy with some dithering and posterization (this is how Nuke's viewer works, it's a 32 bit frame buffer with an 8 bit real-time dithering). But even after converting a 32 bit gradient to an 8 bit gradient, with sufficient dithering it will look subjectively the same (maybe a little more grainy) it's *not* going to wash it out or change the image.

The reason we tend to associate localized tone mapping or log gamma curves with high dynamic range imagery is because the real world generally has a split. There is tons of information in "0 - 10" stops of detail from the average reflectivity of objects, and then usually there is just a few light object which are stops like 16-20. It's that large gap between a light source and a diffuse reflector that makes grading a high dynamic range image tricky. The metaphor for that situation would be like a gradient that goes from 0 - 75% and then just 75% until 100%. It's not the display's fault, it's just a bad gradient (subjectively speaking) However, there are plenty of other examples of imagery where you can have 32 stops of dynamic range like the initial example without a gap.

From both of these posts, it seems as if you're equating bit depth with f-stops because of the binary and exponential nature they both share. I'm not so sure that's an accurate correlation.

From what I've been reading, it's a misconception to equate the two. At the very least, it seems to be open to debate and I'm sure you'll be able to point to sources that support the position.

Here's one such link from Raul Moreno's website: http://www.qvolabs.com/Digital_Images_ColorSpace_Log_vs_Linear.html

Maybe you can refer me to a more reputable source than this or what I've provided below in support of this assertion?

At any rate, I offer the following links as counterpoint in the spirit of broadening all of our collective knowledge.


cambridgeincolour.com - http://www.cambridgeincolour.com/tutorials/dynamic-range.htm
According to this article, the bits-to-stops correlation is only valid for the A/D converter.

From the section "Bit Depth & Measuring Dynamic Range" (emphasis is theirs):

"Even if one's digital camera could capture a vast dynamic range, the precision at which light measurements are translated into digital values may limit usable dynamic range. The workhorse which translates these continuous measurements into discrete numerical values is called the analog to digital (A/D) converter. The accuracy of an A/D converter can be described in terms of bits of precision, similar to bit depth in digital images, although care should be taken that these concepts are not used interchangeably. The A/D converter is what creates values for the digital camera's RAW file format."

And, on the chart that accompanies that section:

"Note: Above values are for A/D converter precision only, and should not be used to interpret results for 8 and 16-bit image files. Furthermore, values shown are a theoretical maximum, assuming noise is not limiting. Additionally, this applies only to linear A/D converters."


rags-int-inc.com - http://www.rags-int-inc.com/PhotoTechStuff/TonesnZones & http://www.rags-int-inc.com/PhotoTechStuff/ETTR
This guy wages full out war on the bits-to-stops idea. He's a maverick and contradicts the info from the guy in the next link; he doesn't buy into the theory behind why people ETTR. He does seem to agree with the previous link in stating the bit depth refers to the precision of the conversion, not the dynamic range.


normankoren.com - http://www.normankoren.com/digital_tonality.html
Although he doesn't explicitly state the bits-to-stops relationship is only for A/D converters, he only uses it in connection with describing the A/D converter in his chart.


photo.net - http://photo.net/filters-bags-tripods-accessories-forum/00Ke2M
Here's a discussion that resulted from a person asking about the bits-to-stops relationship.


All of these pages are from guys on the internet, not an established company or entity, so take them all with a grain of salt.

 

[Gavin Greenwalt]

Bits to stop aren't directly related. That was precisely my point. You can have a 32 bit gradient or an 8 bit gradient and they will look subjectively similar if they go from black -> white. However when you are talking about image synthesis whether that's a simple photoshop gradient or else a 3D rendering, since there is no noise there is a direct relationship. 0.002 is one stop more than 0.001 therefore your floating point precision directly translates into how many steps, or stops there are. 2^10 bits = 2^10 steps = 10 stops. So the takeway isn't that 16 stops is 16 bits, it's that if you have 32 bits of image data in a 3D rendering, it'll look perfectly normal, even though it mathematically by definition has 32 stops of precision. You don't need a 32 stop display to appreciate an image with 32 stops just like you don't need a 32 bit display to appreciate an image with 32 bits.

 

[Hrvoje Simic]

cambridgeincolour.com - http://www.cambridgeincolour.com/tutorials/dynamic-range.htm
According to this article, the bits-to-stops correlation is only valid for the A/D converter.

From the section "Bit Depth & Measuring Dynamic Range" (emphasis is theirs):

"Even if one's digital camera could capture a vast dynamic range, the precision at which light measurements are translated into digital values may limit usable dynamic range. The workhorse which translates these continuous measurements into discrete numerical values is called the analog to digital (A/D) converter. The accuracy of an A/D converter can be described in terms of bits of precision, similar to bit depth in digital images, although care should be taken that these concepts are not used interchangeably. The A/D converter is what creates values for the digital camera's RAW file format."

In short:

Bit depth at the point of capture affects the ADC's ability to quantize the available DR.
After that point it serves only for precision.

 

[scott william]

what excited me about the dragon sensor was that 2000 ISO will be as clean as 800 ISO on MX, that doesn't seem to be the case. I would happily trade resolution for a cleaner image.