I think that 16-bit precision on the GFX 100 is no longer just for specmanship. In extreme circumstances, it is worth turning on.
I stumbled upon this ... found it interesting in the context of how many times folks have said that Jim has indicated 16 bit was of no value, it was only marketing goo ... vs. my point that when pushing the math, that the difference can be realized.
RoamingScott wrote:
lol a 5 stop exposure boost. ok...
Yes, that is one example of "pushing the math". There are other ways (i.e. curves + contrast + exposure + sat, etc.) of pushing things hard (via stacking operations, etc.) If you never push the math, then it isn't meaningful (has always been my point) until you do. But, that's different from folks saying that there is "no difference" ... only marketing goo.
I am willing to believe the claims of the original post by Jim and what you are proposing. Having said this, I cannot help disliking the "non-scientific" experiment by Jim. How black really was the supposedly black area of the image? How meaningful was the WB adjusted on the "grey" cover of another book? I wish Jim used calibrated black and grey targets, either in a light of known temperature, or with a calibrated color profile.
His case was about extreme lifting the shadows, and 16 bit should in theory help recovering more image data. In your examples, I am less sure what 14 vs 12 bit can do. Did you have to lift deep shadows in these images? If not, what happens when 14 bit and 12 bit raw images are converted to jpegs? Wouldn't they become indistinguishable? What kind of existing media can handle (display) a 14 bit image without loss of some data? By looking at your examples, I wonder if these should be more affected by the choice of a color space (wide vs sRGB) more than by the 14 vs 12 bit ADC used to digitize the raw?
Ultimately, is the only advantage of 14 vs 12 bit that one can recover image information from the extremes of deep shadows?
It's all a bit of woo woo to which devotees will non-scientifically swear fealty.
In the end it matters precious little in the mind of anyone but the one behind the screen. Those in front of the screen will never know the details, and in the end, if the art requires them to, it's likely not true art in the first place.
It's been well proven that 14 and 16 bit are effectively identical in the VAST majority of cases. Of course, anyone is welcome to hinder their own workflow in any way they see fit, but unceasing proselytizing of the benefits of rumored ectoplasm is a bit much.
RoamingScott wrote:
It's been well proven that 14 and 16 bit are effectively identical in the VAST majority of cases.
Which is the same as saying that there are some times when 14 bit and 16 bit aren't effectively identical. Thus, it isn't ONLY marketing goo. It may be extreme use case only, etc. ... and something that the vast majority of folks will find it not applicable.
That's always been understood. Thanks for acknowledging that it isn't ONLY marketing goo (albeit, limited cases where it breaches the thresholds of 10 bit or 8 bit outputs).
ruthenium wrote:
Ultimately, is the only advantage of 14 vs 12 bit that one can recover image information from the extremes of deep shadows?
Color banding is of greater significance to me than pulling up shadows from the abyss. So, I'd suggest that recovery from deep shadows isn't the "only advantage". Retaining tonality of transitions, as well.
Again, when the math is being applied, it can create "gaps" in the underlying numbers. Starting with smaller gaps (i.e. 16 / 14 / 12 / 10 / 8 / etc.) of precision, means the gaps remain smaller as the math is being applied. If the final outputs never breach the output limits, no blood, no foul. But, just like trying to ratchet up an 8 bit jpg creates banding sooner than working a 12 bit raw, which bands sooner than 14 bit ... the math continues. Thus, again ... my point being when you push the math hard, is when the 16 bit is (potentially) beneficial over 14 bit, and certainly over 12 or 8 bit (the latter being obvious, the former being lesser realized, yet still real).
ruthenium wrote:
I genuinely don't know whether "Retaining tonality of transitions" depends more on the width of the color gamut rather than on the 14 or 16 bit ADC.
Well, if you're color gamut is wide 10 bit, and you start with 14 bits, you've only got 4 bits of wiggle room. Push things starting at 14 bits by 5 bits and you're into 9 bit territory, relative to a 10 bit output capability (i.e. less refined). If you're starting at 16 bits, and push things 5 bits, you're in 11 bits, relative to a 10 bit output (i.e. more refined). Of course, if you're output is only 8 bit ... starting at 14 bit and you've got 6 bits of wiggle room, etc.
I am interested in getting a better understanding of what we are discussing here.
First, I believe that the width of a color gamut is different from the bit depth of an ADC. The latter tells how a signal (light, in case of photography) is digitized, how many numbers(levels) of light intensity are in the range from the black point to the top of the highlights. For example, the output of a 16 bit ADC contains 2^16 numbers corresponding to different amounts of light (numbers of photons) hitting the photodiodes of a sensor. This has little (if anything) to do with the range of colors observed after demosaicing.
Is having more numbers in the range from black to highlights good? To a point.
I see this like different rulers. For example, a ruler with ticks separated by 1/16 of an inch is arguably better (more useful in practice) than a ruler that has 1/4 of an inch between the ticks. One can argue that a ruler with the ticks at 1/32 of an inch is even better. However, from here we run into a situation when, at some point, e.g., at 1/128 or 1/256, the human eye of an average person would no longer see the individual ticks. In terms of light, there is also a limit of what the human eye can see as what you refer to as "tonality transitions." I am prepared to claim that there must be a point from which "when pushing the math" the human eye would no longer see any more smoother "tonality transitions." This would not surprize me, although I don't have the relevant knowledge to make any claims, if equivalent images obtained with a 14 bit and a 16 bit camera would look exactly the same when viewed on a high-quality modern display or when printed. Again, as non-expert, I wonder if there are displays or printers that can even render a 16 bit image?
Thus, when 16 bit might be better than 14 bit? The same as a ruler with very fine divisions between the ticks - when one wants to amplify, measure, or reveal small things. In terms of light, this is about lifting shadows to reveal the image information that is in the small numbers/low levels of light. The more levels (numbers) are there at the bottom of the digitized signal, the more detailed is going to be the manipulated image after applying a selective gain toward the shadows.
Your examples, if I understand your intent, are more about producing dramatic colors with smooth color transitions. Images 1 and 2 have dark areas transitioning into apparent black (that may or may not be true black) but they don't seem to have strong highlights (unlike image 3). Thus, if I am correct about images 1 and 2 as NOT displaying a high-dynamic range of light, then a high-dynamic range ADC would not provide any visible benefits there.
ruthenium wrote:
I am willing to believe the claims of the original post by Jim and what you are proposing. Having said this, I cannot help disliking the "non-scientific" experiment by Jim.
ruthenium wrote:
Thus, when 16 bit might be better than 14 bit? The same as a ruler with very fine divisions between the ticks - when one wants to amplify, measure, or reveal small things. In terms of light, this is about lifting shadows to reveal the image information that is in the small numbers/low levels of light. The more levels (numbers) are there at the bottom of the digitized signal, the more detailed is going to be the manipulated image after applying a selective gain toward the shadows.
Precision finer than that required for proper dithering is not useful. So it depends on the read noise.
Noise matters. Colorspace matters. Contrast matters. Exposure matters. And so does bit-depth matter.
In all cases, the tipping point of the discussion is when, where and how much do they each matter? In most case an exposure that’s ⅓ stop under won’t matter much. One that’s ⅓ stop over might matter more. ISO 1600 noise may be acceptable where ISO 3200 isn’t, but 800 and 400 are virtually guaranteed to be fine if 1600 was. The sRGB colorspace is fine for displaying images on most computer displays, but wholly inadequate for a quality print.
So when do the extra 2 bits of color data over 14 matter? When your colors are near the limits of a relatively large colorspace; where for example shadow gradients start going blue or yellow instead of holding gray…. Conversely, anything over 8 won’t help at all for an sRGB web image; but 14 or even 12 is likely a whole lot better when you’re processing an image out before converting it to sRGB. Hence, it all matters, at least to some extent in certain parts of the imaging chain.
Re "So when do the extra 2 bits of color data over 14 matter?" (in the above post): this is one of my questions, whether the 14 or 16 bit has anything to do with color? Correct me if I am wrong, my understanding is that references to 14 or 16 bit are references to the bit depth of an ADC. In a digital camera, the ADC converts photons to raw data - the numbers from low to 2^14 or 2^16 corresponding to the levels of light from low to highlights (if I am not mistaken, the black point is set arbitrarily to a low number, e.g. 256). Raw files contain numbers of light intensity recorded by the photodiodes under red, blue, and green microlenses. The color is "synthesized" from the data, in different ways, in post-processing.
To the best of my understanding, a wide color gamut simply has more colors (a wider color palette) than a narrow/small color gamut. I have never come across requirements like "the use of a wide color gamut requires data from a 16 bit (or 14 bit) camera; while it wouldn't work on a 12 bit raw" or "demosaicing a raw file from a 14 bit camera results in a wider color gamut compared to that of a demosaiced 12 bit raw." This connecting the color gamut with the bit depth of a camera doesn't ring true to me.
ruthenium wrote:
Re "So when do the extra 2 bits of color data over 14 matter?" (in the above post): this is one of my questions, whether the 14 or 16 bit has anything to do with color? Correct me if I am wrong, my understanding is that references to 14 or 16 bit are references to the bit depth of an ADC. In a digital camera, the ADC converts photons to raw data - the numbers from low to 2^14 or 2^16 corresponding to the levels of light from low to highlights (if I am not mistaken, the black point is set arbitrarily to a low number, e.g. 256). Raw files contain numbers of light intensity recorded by the photodiodes under red, blue, and green microlenses. The color is "synthesized" from the data, in different ways, in post-processing.
To the best of my understanding, a wide color gamut simply has more colors (a wider color palette) than a narrow/small color gamut. I have never come across requirements like "the use of a wide color gamut requires data from a 16 bit (or 14 bit) camera; while it wouldn't work on a 12 bit raw" or "demosaicing a raw file from a 14 bit camera results in a wider color gamut compared to that of a demosaiced 12 bit raw." This connecting the color gamut with the bit depth of a camera doesn't ring true to me....Show more →
You need to do some more homework on digital color science. Simply stated, to maintain smooth hue gradations between colors in larger gamuts, you need more colors; and greater bit depth provides precisely that. What you're suggesting is akin to to claiming you can stretch 100 pennies and create a 100 dollar bill.
ruthenium wrote:
Re "So when do the extra 2 bits of color data over 14 matter?" (in the above post): this is one of my questions, whether the 14 or 16 bit has anything to do with color? Correct me if I am wrong, my understanding is that references to 14 or 16 bit are references to the bit depth of an ADC. In a digital camera, the ADC converts photons to raw data - the numbers from low to 2^14 or 2^16 corresponding to the levels of light from low to highlights (if I am not mistaken, the black point is set arbitrarily to a low number, e.g. 256). Raw files contain numbers of light intensity recorded by the photodiodes under red, blue, and green microlenses. The color is "synthesized" from the data, in different ways, in post-processing.
To the best of my understanding, a wide color gamut simply has more colors (a wider color palette) than a narrow/small color gamut. I have never come across requirements like "the use of a wide color gamut requires data from a 16 bit (or 14 bit) camera; while it wouldn't work on a 12 bit raw" or "demosaicing a raw file from a 14 bit camera results in a wider color gamut compared to that of a demosaiced 12 bit raw." This connecting the color gamut with the bit depth of a camera doesn't ring true to me....Show more →
Most importantly, I look at the work of people who swear by the need for 16bit.
- the difference between 14-bit and 16-bit is immaterial unless you are doing some fairly significant pushing in post. As a starting point it is obviously meaningless if you are shooting jpg on your camera, and essentially meaningless if you are just doing basic, straightforward conversions of the raw files and little more than tht.
- When _really_ pushing the raw files — as in a LOT — it might make an extraordinarily subtle difference in a few edge cases. Maybe. If you looked really close. Or measured rather than looked.
- in these cases it _could_ be one factor affecting color. As we know, that last bit is always an approximation. The actual luminosity in the area represented by a pixel lies somewhere between the recorded value of that last bit and the second-closest bit. So things always get rounded slightly. It is conceivable that if, for example, the source was precisely gray, that this rounding of the R, G, and B values could introduce slight (VERY slight) color to that gray. You would never notice this in an image that isn’t pushed in post. However, if you took, say, a near black tone (from the part of the luminosity range that is represented by a smaller range of values) and then pushed the crap out of it — like 5 or 6 stops or more — then your “black” color might start to shift a very tiny bit. Maybe.
- Here, though, we get to the difference between what the numbers tell us (that due to that last bit being an approximation, inaccuracies are necessarily introduced, that the representation of luminosity in color channels is not perfectly accurate, and that extraordinary pushing of those values can amplified differences) and what we see in actual photographs (not much at all, if anything, actually).
No digital system _perfectly_ represents anything, but they can reduce the magnitude of the inaccuracy to the point at which the error introduced by approximation is not perceptible.
I guess that if you have 16 bits available and there’s no cost (in speed, memory, money) that bothers you, why not? It doesn’t hurt, right? But does it make a meaningful difference in the end to go from 14 bit to 16 bit? Not likely.
(I just imagined us back here in 10 years debating whether going from 48 bits to 50 bits is the ne plus ultra of photography…)
