Working at 8 / 16 / 32 bits per channel...
Adam
Will gradients be rendered more smoothly in those situations?
What do printers do with those files? Any problems?
When should one work at 16 or 32 bits per channel?
Thanks.
Johan W. Elzenga
> How many of you work at higher bits per channel than 8?
Nobody can tell you how many of us do that.
> Will gradients be rendered more smoothly in those situations?
Perhaps they do, but as most screens do not support more than 8 bits,
you won't see it. Working in 16 bits matters for *editting* those
gradients (and other things).
> What do printers do with those files? Any problems?
Most printers do not use 16 bits, but sending a 16 bits file to a
printer doesn't cause any problems. The driver will change it to 8 bits.
> When should one work at 16 or 32 bits per channel?
32 bits is for HDR (High Dynamic Range) photography only.
--
Johan W. Elzenga johan<<at>>johanfoto.nl
Editor / Photographer http://www.johanfoto.com
John J
> 32 bits is for HDR (High Dynamic Range) photography only.
Today I read a claim that editing in 16 or 32-bit reduced rounding
errors, thus increasing quality. Rubbish?
And is this just snake-oil?
http://www.epson.com/cgi-bin/Store/jsp/ProImaging/EpsonInnovations.do?invMoreInfo=EpsonInv16BitPrinterDrivers
pupick
The answer is the same as the medieval question about how many angels can
sit on the head of a pin.
Oils vs acrylics.
First of all many photographers would be surprised to learn that there is a
good chance they are judging images on a 6 bit LCD panel, particularly if
their computer is named after a fruit.
The widest photo printer gamuts are in the 8 bit realm, offset printing even
less.
Raw images are most commonly captured in 12 bits, some now 14, truncated to
8 bits when opened, or if opened as 16 bit images most processes will be
applied on bits that contain no information.
The idea that processing a 12 or 14 bit raw image in 16 bits is better than
processing it at 8 bits is technically correct with regard to yielding more
smoothly curved plots of the colors at the extremes of the gamut.
Does this make any practical difference when the final image will be viewed
via a medium that is incapable of accurately reproducing that color gamut
anyway?
As I recall Epson used to be very upfront in their literature on the 8 vs 16
bit issue from the printer standpoint but I don't know if they are
officially still as truthy. They showed graphs and actual color mismatches
generated by working in 16 bits.
Does it make sense to send a 16 bit file, most containing empty data, to a
printer driver that will arbitrarily strip it down to 8 bits and think this
will render a better image --if you pay Apple prices for "new" but already
obsolete Intel hardware the answer is obviously yes, but to those who still
have the capacity to reason the answer may not be yes.
If you think working in 8 bits, 16 bits, LAB color or whatever works for you
then that is all that counts.
John J
> First of all many photographers would be surprised to learn that there
> is a good chance they are judging images on a 6 bit LCD panel,
It would be fair to specify exactly what monitors are/were 6-bit - the
Apple 20" was one, I think. But not all of them are.
Mike Russell
> How many of you work at higher bits per channel than 8?
A significant number of people do
> Will gradients be rendered more smoothly in those situations?
If you mean a gradient from a photograph, no. If you mean a computer
generated gradient, yes, in some situations, particularly if you are
speaking of a mathematical smoothness, and not a visual one.
> What do printers do with those files? Any problems?
Usually no problems at all.
> When should one work at 16 or 32 bits per channel?
8 vs 16 is a personal preference. Some people simply want the extra bits,
periodm and feel that anything less is compromising quality. My main
interest is in producing tools that they can use, rather than convincing
them as to one workflow or another.
16 bits is also required to prevent artificial banding in unusually leggy
color spaces, such as ProPhoto RGB and Wide Gamut RGB. 32 bits is useful
for a number of applications, including, as Johan says, HDR images, and
other high dynamic range images such as combined astronomical images.
--
Mike Russell - http://www.curvemeister.com
John J
> On Fri, 13 Mar 2009 08:44:04 -0400, Adam wrote:
>
>> How many of you work at higher bits per channel than 8?
>
> A significant number of people do
>
>> Will gradients be rendered more smoothly in those situations?
>
> If you mean a gradient from a photograph, no. If you mean a computer
> generated gradient, yes, in some situations, particularly if you are
> speaking of a mathematical smoothness, and not a visual one.
Mike, am I correct that the reason a photograph's graduated tone might
not be as smooth is because it does not follow a regular transition?
Johan W. Elzenga
> Johan W. Elzenga wrote:
> > 32 bits is for HDR (High Dynamic Range) photography only.
>
> Today I read a claim that editing in 16 or 32-bit reduced rounding
> errors, thus increasing quality. Rubbish?
No, and I never said it was. However, because there are no input devices
(cameras, scanners) that have *more than* 16 bits/color depth, there is
no reason to work in *more than* 16 bits for single images. Only HDR
gives you the need for 32 bits. That's what I said and I stand by it.
> And is this just snake-oil?
> http://www.epson.com/cgi-bin/Store/jsp/ProImaging/EpsonInnovations.do?invM
> oreInfo=EpsonInv16BitPrinterDrivers
READ what I said, please. There are indeed *SOME* printers that can
handle 16 bits files. But like I said, *MOST* printers cannot. The
question was: does it create problems? My answer was, no it does not.
I stand by that too.
John J
> John J <no...@droffats.ten> wrote:
>
>> Johan W. Elzenga wrote:
>>> 32 bits is for HDR (High Dynamic Range) photography only.
>> Today I read a claim that editing in 16 or 32-bit reduced rounding
>> errors, thus increasing quality. Rubbish?
>
> No, and I never said it was. However, because there are no input devices
> (cameras, scanners) that have *more than* 16 bits/color depth, there is
> no reason to work in *more than* 16 bits for single images. Only HDR
> gives you the need for 32 bits. That's what I said and I stand by it.
I did not suggest that you mentioned rounding errors, nor did I disagree
with comments about HDR.
I simply asked whether rounding errors are a significant problem. I
think they are not.
Mike Russell
> Mike, am I correct that the reason a photograph's graduated tone might
> not be as smooth is because it does not follow a regular transition?
Yes. This is a good description of what causes banding - an irregular
transition in one or more color channels that creates a visible edge or
ripple, aka "band" in an otherwise smooth gradient. When this happens, the
photographer has lost tonal control of the image.
Johan W. Elzenga
Editting in 16 bits does give an advantage over editting in 8 bits, even
though you often do not see it clearly on your monitor. It depends on
the kind of images.
Robert Barnett
fact it can be used to interesting effect. Back in the days of CS2 Russell
Brown of Adobe created a free script for Photoshop called shake, rattle and
roll. This script made use of the rounding errors in Photoshop to create
some very interesting effects. While the script unfortunately has never been
updated for CS3 or CS4 you can still do it manually.
If you would like to see just how bad and how quickly rounding errors can
affect your images load a low resolution image (high resolution will work
but will take longer) and then rotate the image by 8 degrees using the
transform tool. Now repeat this over and over again. Do it 20 or 30 times
and you will see that the image is slowly destroyed, this is because of
rounding errors in Photoshop. The script I mentioned did other things for
example it would rotate back and forth, spin it around and things like this
to give some very interesting effects. If you would like to get an idea of
the effect keep rotating the image the more you do it the more dramatic the
effect.
I have tried to get Russell to update the script but he says he doesn't have
the time.
The moral of this is don't transform any more than you have to. Each time
you do you damage the image. It may take some doing to notice it, but the
damage is happening from transformation one. This is also just one area of
Photoshop that has rounding errors. Now does this mean that in 16-bit or
32-bit that you have less rounding errors or get less damge I don't know, I
don't care. I never transform an image enough to really worry about it. And,
since just about any output is 8-bit it just isn't worth the trouble in my
opinion. If others thing it is awesome, have fun.
Robert
erpy
That's correct.
16bit and 32bit floating point linear editing yields much more precision
when using many layers with different effects/adjustments.
When using only a few layers it might not be that noticeable...
otherwise it makes much more difference.
As for 32bit editing in Photoshop, it depends whether Adobe is
dedicating part of the precision to HDR data or not.
Anyway, 32bit floating point editing is much more precise than 16bit
(which is integer anyway).
Mike Russell
> As for 32bit editing in Photoshop, it depends whether Adobe is
> dedicating part of the precision to HDR data or not.
> Anyway, 32bit floating point editing is much more precise than 16bit
> (which is integer anyway).
Any examples where any of this matters? Didn't think so.
mike
says...
> Actually John rounding errors are a much bigger problem than you think. In
> fact it can be used to interesting effect. Back in the days of CS2 Russell
> Brown of Adobe created a free script for Photoshop called shake, rattle and
> roll. This script made use of the rounding errors in Photoshop to create
> some very interesting effects. While the script unfortunately has never been
> updated for CS3 or CS4 you can still do it manually.
>
> If you would like to see just how bad and how quickly rounding errors can
> affect your images load a low resolution image (high resolution will work
> but will take longer) and then rotate the image by 8 degrees using the
> transform tool. Now repeat this over and over again. Do it 20 or 30 times
> and you will see that the image is slowly destroyed, this is because of
> rounding errors in Photoshop. The script I mentioned did other things for
> example it would rotate back and forth, spin it around and things like this
> to give some very interesting effects. If you would like to get an idea of
> the effect keep rotating the image the more you do it the more dramatic the
> effect.
>
almost everything to do with the resolution of the image. For example,
create a small 8-bit per channel image (100x100 pixels square is fine)
and draw a vertical black single pixel width line down the centre of the
white background. Now apply an arbitrary rotation like 4.56 degrees
(don't choose an angle that is a factor of 360) and apply iy just once.
Now repeat by rotating in the opposite direction by the same angle.If
you zoom into the image far enough to see the individual pixels as
monotone squares you will note that the line is now a fuzzy shape
approximately 2 pixels wide (with outliers up to 3-4 pixels wide) in
shades of grey. This is not due to any lack of bit depth as the result
is almost identical if you repeat the task with a 16-bit image. Rather
it is due to the fact that, when rotated, the pixels of the line don't
map exactly onto individual new pixels so their colour value is
'smeared' across 4 or more neighbouring pixels. So it is a resolution
issue. If you repeat the experiment with a 10 pixel wide line in a
1000x1000 image, you can see that at the same scale of magnificatiobn
there is less apparent damage to the line.
Mike
John J
> says...
> Actually John rounding errors are a much bigger problem than you think. In
> fact it can be used to interesting effect. Back in the days of CS2 Russell
> Brown of Adobe created a free script for Photoshop called shake, rattle and
> roll. This script made use of the rounding errors in Photoshop to create
> some very interesting effects. While the script unfortunately has never been
> updated for CS3 or CS4 you can still do it manually.
>
> If you would like to see just how bad and how quickly rounding errors can
> affect your images load a low resolution image (high resolution will work
> but will take longer) and then rotate the image by 8 degrees using the
> transform tool. Now repeat this over and over again. Do it 20 or 30 times
> and you will see that the image is slowly destroyed,
I use photoshop for photographs. I don't ever need to manipulate the
image in such bizarre ways.
Robert Barnett
on the image. If you try my suggestion you will see that the round errors
can really have an impact and pretty quickly, surprisingly quickly. I made
and make no statement on wheater 8-bit, 16-bit or 32-bit mode will increase
or reduce rounding errors. I don't know and I really don't care. What I was
trying to show is that round errors are there and they have affect your
images and the rounding errors are quite easy to encounter.
Robert
Robert Barnett
or two transformations (rotations to straighten a horrizon), resize up or
down, correct for keystoning and the like you are definately damaging your
images as you are then encountering rounding errors. All of the
tranformation tools introduce data damage to your image (except for flip
horzontal and flip verticle. All of the others damage. So if you use any of
the tools (and who knows what other tools) you are damaging your images. As
a photographer myself you can avoid having to use these tools at some point,
it is best to keep their use to a minimum and not to keep applying them over
and over to correct a single issue. Straigtening the horizon with a single
application or rotate is fine, having to apply it several times to do it
isn't. Though honest using the ruler with the rotate command I see no reason
why you would have to apply rotate more than once.
I am sure that many of the filters have rounding errors as well.
My point for my post was that rounding errors are real, they do damage your
images and if you want to see them do it I provided a way to see it. Proof
that they are real in other words. For most they won't be a problem, but one
needs to keep them in mind.
Robert
Mike Russell
> I think this has very little to do with the bit-depth of the image and
> almost everything to do with the resolution of the image.
Good point. I did the experiment. Applying repeated rotations to an 8 bit
image, versus a 16 bit image, results in nearly the same result. Doubling
the resolution preserves image detail much better, will both 8 bit and 16
bit producing nearly identical results.
http://www.curvemeister.com/forum/index.php?topic=2569
Henk de Jong at
This question has been asked many times. Part of it is answered in this
article:
http://www.creativepro.com/article/out-of-gamut-the-high-bit-advantage
With kind regards,
Henk de Jong
http://www.hsdejong.nl
Nepal and Myanmar (Burma) - Photo Galleries
Mike Russell
> This question has been asked many times. Part of it is answered in this
> article:
> http://www.creativepro.com/article/out-of-gamut-the-high-bit-advantage
>
>
> With kind regards,
> Henk de Jong
Have you tried to duplicate what the article describes? I have.
John J
> My point for my post was that rounding errors are real, they do damage
> your images and if you want to see them do it I provided a way to see
> it.
He who needs to do excessive iterative rotations for a photograph, or
has to make so very many non-adjustment layer manipulations to a
photograph can live with the degradation alleged to occur through
rounding errors.
Only someone with more monitor time than lifetime can care about such
unrealistic bench-racing metrics that require a damned spreadsheet to
'see' and still have nothing to do with real-world outcomes as it
regards photography.
erpy
On Tue, 07 Apr 2009 20:58:57 +0200, erpy wrote:As for 32bit editing in Photoshop, it depends whether Adobe is dedicating part of the precision to HDR data or not. Anyway, 32bit floating point editing is much more precise than 16bit (which is integer anyway).Any examples where any of this matters? Didn't think so.
You can do this easy test... sometimes you see this banding artifact on TV adverts too...
- make a new 24bit image, let's say HD 1080 resolution, pick two close colors, say RGB(0,100,100) and RGB(0,150,150)
- make a radial gradient from the center to one corner of the picture.
- Duplicate twice the layer and set the duplicates to Multiply. Duplicate twice again and set these other duplicates to Overlay.
- Notice the banding.(100% zoom)
- Convert the image to 16 bit/component...
- The banding is gone.
mike
grou...@MOVEcurvemeister.com says...
> On Wed, 8 Apr 2009 09:22:12 +1200, mike wrote:
>
> > I think this has very little to do with the bit-depth of the image and
> > almost everything to do with the resolution of the image.
>
> Good point. I did the experiment. Applying repeated rotations to an 8 bit
> image, versus a 16 bit image, results in nearly the same result. Doubling
> the resolution preserves image detail much better, will both 8 bit and 16
> bit producing nearly identical results.
>
before.
It is considered good practice by some to 'upscale' an 8 bit image to 16
bits before applying a whole bunch of colour corrections. If the image
is then set back to 8 bits there are less rounding errors in the final
image than if the whole process is carried out at 8 bits (and yes, I
know there is debate as to whether it actually makes much difference for
many images).
But what does appear a valid arguement to me is to up-scale the
resolution of an image as much as possible before applying distortions
(rotations, skew, scale, pinch, spherise, liquidise etc) then down-scale
to the previous resolution. I tried a simple test where I drew a few 1
pixel wide lines and boxes of various colours on a 100x100 image. Then
created a copy at 1000x1000 resolution. Applied identical rotation,
pinch, and skew to both images then resized the larger image back to
100x100. The first image showed significant 'smearing' of line-shape and
colours bled into neighbouring regions - the upscaled and then
downscaled image showed much less of this (in fact most of the 'damage'
to the second image was due to the downscaling process itself).
In a photographic image at 2-3000 pixel resolution this may not make
much difference unless you went wild with the distortion process - but
if you wanted to minimise 'damage' to fine details it might be worth
while in some cases.
Mike
John J
> John J <no...@droffats.ten> found these unused words:
If one cannot find a difference in these strange abstracted deviations
from 'the the thng itself' (a photographic image) then the result is
FAPP, nill, just usenet typing and meaningless. Let the bit-twi8ddler
have their way; it has nothing to do with compelling imagery.
Mike Russell
> There is a HUGE difference between 8-bit and 16-bit, and depend on the
> photo or condition you may have to zoom in 200-300% or so to learn the
> difference.
I'd like to see an image that demonstrates this.
Mike Russell
Gradient doesn't count - gotta be a photograph for me to count it as an
example.
John Passaneau
>
> Another "Close is good enough"?
>
It is if your producing photographs. Maybe not if your producing pixels.
John Passaneau
Mike Russell
> Aw gee, you meant the 'artistic critics' judging my photos' 'technical
> merit' have been wrong? It's really the image that counts?
They may have been wrong, or they may have been right. The only way to
know would be to see some of your pix.
I am biased in favor of people who bring real names, as well as real
photographs, to the table. The others, I agree with Robert, are just so
much blather.
erpy
I thought the argument was about the "16 bits difference" per-se... not "for photographs". And I was actually right.
If you read the OP, he asks in fact:
"Will gradients be rendered more smoothly in those situations?"
the answer is: "Yes"
erpy
That's because it would show you both why you don't notice much difference between an 8bit and a 16bit photo and at the same time how you're being ripped off for good even with the most expensive camera on the whole market. (well, *maybe* excluding top-of-the-line Hasselblads)
Mike Russell
> Oh, and BTW, the digital photography argument would be too painful for
> users like you to take on.
> That's because it would show you both why you don't notice much
> difference between an 8bit and a 16bit photo and at the same time how
> you're being ripped off for good even with the most expensive camera on
> the whole market. (well, *maybe* excluding top-of-the-line Hasselblads)
I think you, and to be fair many others, are making the simplest of
tautologies: "more bits is more better". Your discussion of an 8 bit
gradient versus a 16 bit gradient can be equally well applied to a 16 bit
gradient versus a 32 bit gradient, 32 bit and 64 bit, and so on. As Buzz
Lightyear says, "to infinity, and beyond!".
At some point, I'm sure you will agree, the number of bits becomes "lost in
the noise".For a photograph, the bedrock of whether this reasoning is valid
or not should be in the form of an image that shows the effect. For a
computer generated gradient, there is no noise, and therefore the tautology
holds in Buzz Lightyear.
My challenge, which still stands, is to present a photographic image, in a
normal color space such as sRGB or Adobe RGB, that edits better in 8 bit
than 16 bit. Do you have one, or not? That's all, very simple.
Here are some examples of quality variation that we can agree is real:
1) variation of image quality with jpeg quality - easily demonstrated, and
not a source of discussion.
2) variation of image quality in the face of repeated transforms as a
function of bit depth, and of image resolution. I presented an example
illustrating this earlier in this thread.
3) variation of image quality for indexed versus 24 bit RGB images. Again,
an example of this is easily provided.
4) variation of image quality with ISO setting.
5) variation of image quality with aperture setting.
6) variation of image quality with image stabilization.
7) the difference in image quality between a high quality jpeg and an
uncompressed TIFF or raw image.
Each of the above are easily demonstrated using *photographic images*.
But ah, according to you (and to be fair, others), variation of image
quality versus bit depth is blazingly obvious. But, unlike the above seven
examples, it is not so obvious that you can point to an example photograph,
just a computer generated gradient which d nothing to do with nothing.
John J
> [...]
> They may have been wrong, or they may have been right. The only way to
> know would be to see some of your pix.
>
> I am biased in favor of people who bring real names, as well as real
> photographs, to the table. The others, I agree with Robert, are just so
> much blather.
Mike, what would you like? URLs to digital representations of pictures?
Mike Russell
> Mike, what would you like? URLs to digital representations of pictures?
Sounds fine.
BTW, just to fill in a little bit, Rien and I have had this discussion
before. His name is a play on the words "ca ne fait rien" meaning, "that
doesn't matter at all". He is tickled, a little, when people recognize the
pun.
Mike Russell
> That's pretty much same with 8-bit and 16-bit, or you will get more pixel
> from 16-bit then 8-bits, and if you spread them around then the more pixel
> the smoother.
You've got it backwards. If you keep the file size the same, an 8 bit
image will store exactly twice the number of pixels as a 16 bit image.
IOW, if you are going to double the size of the file, it's better to have
twice as many 8 bit pixels, as it is to have the same number of 16 bit per
channel pixels.
Mike Russell
> May be I have everything in backwards, and that's why I sometime switch to
> 16-mode to work on low-rez image. And I only use low-rez for DVD label.
You may find that you like the results better - sharper lines, etc - if you
change the ppi in Image Size, and work on a higher resolution image. When
you're done, resample back down to the original size using Bi-cubic
Sharpen.
John J
> Mike Russell <grou...@MOVEcurvemeister.com> wrote:
>
> <snip>
>> At some point, I'm sure you will agree, the number of bits becomes "lost in
>> the noise".For a photograph, the bedrock of whether this reasoning is valid
>> or not should be in the form of an image that shows the effect. For a
>> computer generated gradient, there is no noise, and therefore the tautology
>> holds in Buzz Lightyear.
>
>
> - Throwing (100) 100x100" rocks on the ground
>
> - Then also (10,000) 1x1" rocks on the ground
>
> And which one looks smoother (flatter, more evenly)?
>
> That's pretty much same with 8-bit and 16-bit, or you will get more pixel
> from 16-bit then 8-bits, and if you spread them around then the more pixel
> the smoother.
However, you view the image on a low-bit monitor (coarse sieve) and
print to a low-bit printer.
erpy
On Fri, 10 Apr 2009 00:57:50 +0200, erpy wrote:Oh, and BTW, the digital photography argument would be too painful for users like you to take on. That's because it would show you both why you don't notice much difference between an 8bit and a 16bit photo and at the same time how you're being ripped off for good even with the most expensive camera on the whole market. (well, *maybe* excluding top-of-the-line Hasselblads)I think you, and to be fair many others, are making the simplest of tautologies: "more bits is more better". Your discussion of an 8 bit gradient versus a 16 bit gradient can be equally well applied to a 16 bit gradient versus a 32 bit gradient, 32 bit and 64 bit, and so on. As Buzz Lightyear says, "to infinity, and beyond!".
At some point, I'm sure you will agree, the number of bits becomes "lost in the noise".For a photograph, the bedrock of whether this reasoning is valid or not should be in the form of an image that shows the effect. For a computer generated gradient, there is no noise, and therefore the tautology holds in Buzz Lightyear.
My challenge, which still stands, is to present a photographic image, in a normal color space such as sRGB or Adobe RGB, that edits better in 8 bit than 16 bit. Do you have one, or not? That's all, very simple.
Each of the above are easily demonstrated using *photographic images*. But ah, according to you (and to be fair, others), variation of image quality versus bit depth is blazingly obvious. But, unlike the above seven examples, it is not so obvious that you can point to an example photograph, just a computer generated gradient which d nothing to do with nothing.
As a similitude, would you run "precision" tests on a "length measure" taken by hand (i.e. the sensor of a digital camera) or on a measure taken with a laser beam (i.e. a computer-generated image) ?
Although, as I said, 16 bits *editing* superiority is easily shown on digital photos as well. The principle stays the same... have many blending layers on your picture and see the difference. Obviously, the more noise you have, the less the difference... but that's only because the source data is crap from the very beginning (your... anyone's crap, noisy digital sensor).
I wouldn't use such a picture for anything anyway.
Not mentioning, the 16 bits in raw photos have a different meaning than "precision". The bit-depth is used to expand the dynamic range of a picture.
Hence if you take a raw photo and open it straight in Photoshop without touching anything within Camera Raw, you're loosing so much "lighting" data you probably have no clue about.
While, strange enough for you, I imagine, all the image processing taking place within Camera Raw is at 16 bits/pixel - despite the fact you can import at 8bits then.
Why, according to your belief, Adobe would waste so much memory and speed by processing photos at 16bits when 8 bits would be more than enough ?
Answer: for all the good reasons I said. :))
(oh well, and why most of the raw-dedicated image processors around are featuring 32bit floating point precision, spending money and effort on this ? Just for the sake of a magic marketing word ? Not this time around.)
Mike Russell
(Note: For those who may not want to read everything that appears below, I
will mention erpy provides no examples to substantiate the claim that
photographs edit better in 16 bits than 8 bits. - Mike Russell)
....
> That's where your "problem" is... you take noise for granted. That's the
Thanks for your concern. I do think your comments are substantial enough
to deserve individual replies.
...
> The challenge is for a photo that *edits* better ? Like anything goes
> within Photoshop without plugins except Camera Raw ? That'd be very
> easy... don't put any money on the table for your challenge! ;))
...
I have, in the past, put money on the table, and paid up. The results were
less than conclusive, due to poor design of the challenge on my part. It's
actually fairly difficult to set the challenge up fairly and clearly.
I paid out the reward, but did not end up with an example of an image that
edited better in 8 bits than 16 bits. This is the main reason that I am
confident that you cannot provide such an example.
...
> Do you actually know how digital sensors in a camera work ? Sensors
> actually *capture* and *store* *1/3rd* of the data needed for an RGB
> image. The rest is "interpolated".
Yes, I do. You're talking about the Bayer pattern. It is inaccurate to
say that the sensors capture 1/3 of the data. Luminance data, which is
what the eye is most sensitive to, is not interpolated. Chroma data is, to
a certain extent. Our eyes do the same thing.
> As a similitude, would you run "precision" tests on a "length measure"
> taken by hand (i.e. the sensor of a digital camera) or on a measure
> taken with a laser beam (i.e. a computer-generated image) ?
This doesn't fly from a practical standpoint. My concern is with
photographs, and not with manipulating arrays of numbers. This is one
reason I reject histograms as a meaningful measure of image quality.
> Although, as I said, 16 bits *editing* superiority is easily shown on
> digital photos as well.
Yackity yackity - if it's "easy", why not do so? I suggest it is because
it is not easy to do so, and may well be impossible, even with extreme
editing after the fact.
> The principle stays the same...
You can demonstrate a principle using numbers and conclude that "more bits
is more better". Whether that principle translates to effective practice
is another question, and it's the one that
> have many blending layers on your picture and see the difference.
Fine. This isn't global warming or world hunger we're talking about, it's
a psd file with some layers in it. So point us to an example? Until you
do so, I suggest you cannot do so, and that you are blowing smoke and
mirrors, with Toto about to pull the curtain away.
> Obviously, the
> more noise you have, the less the difference... but that's only because
> the source data is crap from the very beginning (your... anyone's crap,
> noisy digital sensor).
Reality has a way of messing up theoretical principles, doesn't it? LOL.
> I wouldn't use such a picture for anything anyway.
Not all of us have that luxury.
> Not mentioning, the 16 bits in raw photos have a different meaning than
> "precision". The bit-depth is used to expand the dynamic range of a picture.
It's probably more accurate to say that the additional bits (typically 12)
from the camera sensor can be manipulated in Camera Raw. I would agree that
access to this data can be beneficial, but that's not the same as saying
that working in 16 bits (or 32) will give a result that is better or
different than working in 8 bits. The only thing that will demonstrate
that is an example of an actual photograph.
> Hence if you take a raw photo and open it straight in Photoshop without
> touching anything within Camera Raw, you're loosing so much "lighting"
> data you probably have no clue about.
Right - again no example though. See a pattern here?
> While, strange enough for you, I imagine, all the image processing
> taking place within Camera Raw is at 16 bits/pixel - despite the fact
> you can import at 8bits then.
I have no problem with any of this. What I have a problem with is the
religious belief that working in 16 bit instead of 8 bit in Photoshop
confers any demonstrable advantage.
> Why, according to your belief, Adobe would waste so much memory and
> speed by processing photos at 16bits when 8 bits would be more than enough ?
> Answer: for all the good reasons I said. :))
I've been in computer graphics for quite a while - 25 years or more, and am
well aware of the trade-offs of performance and accuracy. Curvemeister
uses floating point internally everywhere, and in some cases I believe I
get a more accurate, smoother result than Photoshop does.
With today's processors, 8, 16, and even 32 bit ints are processed at about
the same speed. In the early days of computer graphics, the idea of doing
a floating divide per pixel was a show stopper, and that was when a
megapixel image drew oohs and ahhs at SIGGRAPH.
> (oh well, and why most of the raw-dedicated image processors around are
> featuring 32bit floating point precision, spending money and effort on
> this? Just for the sake of a magic marketing word ? Not this time around.)
This sums up your argument, I believe. If 8 bits were enough, why would
the big companies even bother with 16 and 32 bits. One reason, I believe,
is because they can. Another is dynamic range, as you said. But none of
this addresses my request that you provide an example of a photographic
image that edits better in 16 bits than 8 bits.
Getting back to my original point about examples. Great piles of words are
meaningless if there are no examples to back them up. Yet the words grow
and grow, with no photograph to back them up. Just a feeling that 16 bits
must be better than 8 bits.
If it were as easy to find such an example as you claim it is, it should
take you less time to find one than to read, and perhaps, this rather long
post.
All the best to you,
John J
How in the world can we bring real photographs to the digital table?
This is an entirely impoverished format.
John J
negatives, you could not possibly understand the message.
That's how it is here with images at this time.
Digital sucks.
Mike Russell
> How in the world can we bring real photographs to the digital table?
> This is an entirely impoverished format.
Links to a web page, I guess. Or do you mean "real" as in cow hoofs and
silver?
erpy
> This sums up your argument, I believe. If 8 bits were enough, why would
> the big companies even bother with 16 and 32 bits. One reason, I believe,
> is because they can. Another is dynamic range, as you said. But none of
> this addresses my request that you provide an example of a photographic
> image that edits better in 16 bits than 8 bits.
>
> Getting back to my original point about examples. Great piles of words are
> meaningless if there are no examples to back them up. Yet the words grow
> and grow, with no photograph to back them up. Just a feeling that 16 bits
> must be better than 8 bits.
>
> If it were as easy to find such an example as you claim it is, it should
> take you less time to find one than to read, and perhaps, this rather long
> post.
>
> All the best to you,
>
I've done my homework long ago.... I don't see why I should do yours too.
Take a photograph with some out of focus area, import at 16bits from
Camera Raw, duplicate the photo a couple of times and set those layers
to Overlay.
View one channel at a time, R, G or B and switch to 8bit.... banding all
around in the out of focus area (despite the noise).
Switch back history to 16bits...it's all gone.
Are you afraid of doing it by yourself and discover that actually you
destroyed a lot of data in your pictures by editing at 8bits ? I don't
have to convince you of anything, I know my stuff - plus, I don't do
others' homework. :)
You're more than free to keep editing your 14bit+ photos using 8bits
mode. The pictures aren't mine in the end. :))
John J
> I've done my homework long ago.... I don't see why I should do yours too.
> Take a photograph with some out of focus area, import at 16bits from
> Camera Raw, duplicate the photo a couple of times and set those layers
> to Overlay.
> View one channel at a time, R, G or B and switch to 8bit.... banding all
> around in the out of focus area (despite the noise).
> Switch back history to 16bits...it's all gone.
Under what circumstances would a person have to do that to a photograph?
Dave
>
>I've done my homework long ago.... I don't see why I should do yours too.
>Take a photograph with some out of focus area, import at 16bits from
>Camera Raw, duplicate the photo a couple of times and set those layers
>to Overlay.
>View one channel at a time, R, G or B and switch to 8bit.... banding all
>around in the out of focus area (despite the noise).
>Switch back history to 16bits...it's all gone.
>
>Are you afraid of doing it by yourself and discover that actually you
>destroyed a lot of data in your pictures by editing at 8bits ? I don't
>have to convince you of anything, I know my stuff - plus, I don't do
>others' homework. :)
>
>You're more than free to keep editing your 14bit+ photos using 8bits
>mode. The pictures aren't mine in the end. :))
talk talk talk ... erpy, how on earth can anyone keep on making a fool
of himself the way you do? Without even realizing it?
This started being a very interesting
thread; only until it became obvious how desperately you try
to win by talk talk talk without the simple prove requested
by Mike. How does it feel to a mannerless fool like you
losing an argument to an gentleman like Mike?
No class at all:-(
erpy
Under what circumstances one could not ?
If you give me your favourite editing test, I'd apply yours.... until
then... it's a test.
erpy
Dave, eat your tongue next time you call someone a fool without knowing him... who's the mannerless here ? Did I insulted anyone, or you or Mike ?
You just did actually. Proof of you being an uneducated punk.
Mike Russell
> Well, now this "mannerless" fool requires *proof* from Mike, and made by
> himself.
I would not say you are mannerless at all, quite the contrary. It's not
possible to prove a negative, as you surely know.
I would agree with Dave and others that, given that 16 bit files are double
the size of 8 bit, the burden of proof should be on those who believe there
is a visible advantage when it comes to manipulating an ordinary
photograph.
It's not enough simply to prove that an image can represent more than 8
bits of data. That's what's called a tautology. Some sort of example of
its practical usefulness for a normal sRGB or Adobe RGB (or similar space)
image ought to be forthcoming, before we all decide to double our file
size.
I provided seven examples of image quality issues that are easily proven.
Why is 16 bit so different that there are no examples easily at hand? I
suggest it is because there are no such examples, or at least, it seems,
none from you.
Hand waving, proof by vigorous assertion, and personal criticism of my
abilities as a photographer, are not substitutes for an example.
erpy
I don't mean to force anyone to leave 8bits and go 16. If one can't notice the difference between the two it most probably means he doesn't need 16bits.
My point is that, it is mathematical truth that editing at 16bits is superior - regardless if a user will perceive *the inevitable loss* using 8bits in his specific environment/application, or not.
So, in the end, whether this loss at 8bits is "good enough" or "not perceivable" it's entirely subjective, although I think I provided a simple editing scenario where rounding errors are even visible with only 2 layers on top of the original photo.
A small trivia for you: do you know why even a 50-bucks graphic card renders pixels at 128bit floating point precision, instead of 24 ?
erpy
On Sat, 11 Apr 2009 20:12:07 +0200, erpy wrote:Well, now this "mannerless" fool requires *proof* from Mike, and made by himself.I would not say you are mannerless at all, quite the contrary. It's not possible to prove a negative, as you surely know.
I would agree with Dave and others that, given that 16 bit files are double the size of 8 bit, the burden of proof should be on those who believe there is a visible advantage when it comes to manipulating an ordinary photograph.
If you edit at 8bits, the error is all there.
I provided seven examples of image quality issues that are easily proven. Why is 16 bit so different that there are no examples easily at hand?
Mike Russell
> Mike, what do you think about the image I attached to my previous post
> then ?
Thanks for trying. This is a non-binary group, so attachments get
discarded. If you email it to me, I can give it a home on my web site so
others can check it out.
> My point is that, it is mathematical truth that editing at 16bits is
> superior
I don't question the theory. It amounts to 16 is greater than 8, and I
accept that. I seek an example photograph that shows that more than 8 bits
is of any practical use, no matter how perceptive one may be. Again, I'm
talking about a color photograph in a standard color space such as sRGB or
Adobe RGB.
I'll bet Rubles to raisins that your image, come what may, will be a
contrived image in some way
> So, in the end, whether this loss at 8bits is "good enough" or "not
> perceivable" it's entirely subjective, although I think I provided a
> simple editing scenario where rounding errors are even visible with only
> 2 layers on top of the original photo.
The layer example is of interest. I did try your earlier example, and if
the photograph is similar I will have some comments about it.
> A small trivia for you: do you know why even a 50-bucks graphic card
> renders pixels at 128bit floating point precision, instead of 24 ?
Because, being part of objective reality, and not an abstraction, the
signal contains analog noise.
Mike Russell
> It surely is possible to prove "equality" though. Take a 16bit source
> photo, make some extreme editing at 16bits, convert, without flattening,
> at 8bits - compare if they're equal or not.
It would be possible to perform an edit that left an image in the bottom 8
bits of a 16 bit image. Converting to 8 bits would result in a black
image. Likewise, a 32 bit image could contain a 16 bit image in the bottom
two bytes of each channel, and so on for 64, 128. Does it follow, then,
that each increase in precision makes for a better image, or an easier to
edit image? Of course not.
This is the Buzz Lightyear "to infinity and beyond" argument. It is
deficient because it lacks an example of a photograph that edits better in
16 bits. It's not just you. Everyone who promotes 16 bit editing does so
based on a warm fuzzy feeling about more bits is better. I have a lot of
respect for many of these people - the ones who take excellent photographs,
and *especially* the ones who are customers of mine.
Every once in a while, I like to rock the boat and ask one of the people
who promotes this point of view for an example of an image that does
better.
erpy
On Sat, 11 Apr 2009 21:01:02 +0200, erpy wrote:Mike, what do you think about the image I attached to my previous post then ?Thanks for trying. This is a non-binary group, so attachments get discarded. If you email it to me, I can give it a home on my web site so others can check it out.
A small trivia for you: do you know why even a 50-bucks graphic card renders pixels at 128bit floating point precision, instead of 24 ?Because, being part of objective reality, and not an abstraction, the signal contains analog noise.
The question, and the right answer, is clearly related to our debate.
At the beginning 3d cards used 32bit RGBA processing, i.e. 8bits/component. At that times whenever many polygons were layered on screen at the same time banding was inevitable and very noticeable. (imagine, for example, 10+ explosion effects one on top of each other...or many semi-transparent objects and so on)
That is exactly the difference between image processing at 32bits and at 128bits... today you don't see banding anymore on games because rendering is taking place at 32bit/component floating point precision, thus resulting in a correct blending.
It doesn't matter if the output is still 24bits/pixel (although DVI dual link and HDMI reach a maximum of 48bit/pixel digital output). It's the processing precision that matters. If the numerical rounding takes place in-between operations, it accumulates rapidly and the net final effect is that more and more adjacent pixels are set to the same color/value - hence the banding.
This is true for any layered image processing where pixels are blended in some way... and thus it holds true in Photoshop as well.
erpy
On Sat, 11 Apr 2009 21:22:17 +0200, erpy wrote:It surely is possible to prove "equality" though. Take a 16bit source photo, make some extreme editing at 16bits, convert, without flattening, at 8bits - compare if they're equal or not.It would be possible to perform an edit that left an image in the bottom 8 bits of a 16 bit image. Converting to 8 bits would result in a black image. Likewise, a 32 bit image could contain a 16 bit image in the bottom two bytes of each channel, and so on for 64, 128.
You simply need to exhibit a "difference image" of the result of both edits, 16 and 8 bits.
Once you have both images done and layers flattened and all effects rasterized if any, convert the 8bit result to 16bits - nothing changes obviously, it just brings it to the same depth for comparison.
Open both results in HDRShop, which performs 96bit/pixel floating point calculations, and subtract one image from the other. You now have a difference image...which, according to your point, should be black - i.e. all pixels = 0,0,0
If it's not, then you have lost something. The more pixels you can actually see from the difference image, the more data you lost at 8bits, comparing to 16.
HDRShop also has the ability to make the sum of all pixel values in the entire image. If you sum up all the pixels in that difference image and divide that number by the number of pixels, you get the mean error - and that's it for the objective "equality" test.
Mike Russell
> This is true for any layered image processing where pixels are blended
> in some way... and thus it holds true in Photoshop as well.
You are conflating two different universes. In the world of
Photo-realistic rendering, with checkerboards, Rob Cook's rendered metallic
teapots, and crystal spheres going off to infinity (and beyond), yes,
floating point precision is a necessity, not an option. You may have
gathered, surely by now, that this is not the world I'm talking about. As
complex as computer graphics is, it remains an idealized approximation to
the real world. IOW, it's not real.
The world of photography is real. It is in the trenches, using physical
cameras to capture light from physical objects, and real printers with
actual human beings who view the images. It is constrained by the
realities of physics of light, electronics, pigment, cognition, and
physiology. With ambient light reflecting the camera, is more like a VGA
video card than a pure digital one. Real, physical lenses that diffract,
diffuse, and flare, and real printers, monitors, and most important of all,
real eyeballs that are looking at the image.
Taken in its totality, reality is a much more complex world than the world
of CGI. CGI can be thought of as simply a way to verify the accuracy of
our assumptions about how light and perception really work. We have a long
way to go, as I'm sure you would agree.
In the real world, "this" world, I maintain, an 8 bit color photo using a
normal color space provides more than enough headroom for severe edits, and
16 bits provides no advantage. If you disagree, and maintain that it is
trivial to prove otherwise with an example image, I, and the world, await
you.
--
erpy
As soon as you "touch" the source data, you enter the realm of image processing math, which includes numerical precision and error.
I'm sending you the photo editing test. The photo isn't mine, I found it on a website that "donates" raw samples... thus you might probably have seen this picture before, and I disclaim any use of such a photo. If I remember well the source was http://www.rawsamples.ch/ ... but I might be wrong.
In the archive you will find:
- whole view of the original photo
- PSD of an edited crop at 16bit, with layers
- BMP of the 8bits result of the edited crop
erpy
Banding occurs when numerical rounding (error) takes place on adjacent pixels, with low-spaced values (smooth gradient/signal), thus setting those adjacent pixels to the same value. (it does include "fine noise" as well)
Practical example. Say you have source pixels in one of the RGB channels set to the following values, at 8bit (0,255):
(I divide "bands" with "|")
127,127|126,126|125,125|124,124
(4 groups of pixels at 4 different values.. i.e. 4 "bands" 2 pixels "wide")
Say you have a layer that multiplies by 0.5 (50%) - i.e. it halves the brightness. The pixels become:
63,63,63,63|62,62,62,62
Result: 2 bands, 4 pixels wide... more visible. Put another multiplying layer, again half brightness, on top of the resulting pixels...you'd have:
31,31,31,31,31,31,31,31
Result: 1 band, 8 pixels wide.... all the source pixels now have the same value. Pretty bad. You've lost exactly 7/8 (87%) of the original data.
(btw, you get the same rounded values if you multiply directly by 0.25 == 0.5*0.5)
There's little you can do about that... except editing at a higher bit-depth of course.
Photoshop might compensate banding artifacts at 8bits with some dithering, but I haven't tested it for that - it looks like it doesn't at a first glance.
I hope you get the importance of editing at 16bits now.
John J
With respect I must ask if you think any photograph would be susceptible
to editing to the extreme suggested in the post above. If you have
such a candidate, then post ther URL
Thank you,
j
erpy
I have made a reasonable test and sent to Mike. He might want to show it
somewhere.
It doesn't matter how "hard" the single layer is... it's the number of
operations that counts in this case. The more operations you do the more
banding will occur.
(read my latest post about banding)
The banding worsen even quicker if you "flatten" or save to disk at
24bits and then re-open and keep editing over and over.
That is (also) why professional photographers shoot and edit raw
images... because you can re-edit again and again without perceivable
loss of data and precision.
John J
monitors and printers, but until then, all this discourse is bullshit.
erpy
> Someday in the future high resolution will be significant with new
> monitors and printers, but until then, all this discourse is bullshit.
What's resolution got to do with bit-depth ? Banding is somehow related
to resolution, but only because you have more pixels to cover with a
smooth gradient, using the same number of "shades" per-channel.
Otherwise, you can see banding even with small images...it depends by
the content of the image itself.
Talker
<grou...@MOVEcurvemeister.com> wrote:
>On Fri, 10 Apr 2009 18:50:11 -0500, John J wrote:
>
>> How in the world can we bring real photographs to the digital table?
>> This is an entirely impoverished format.
>
>Links to a web page, I guess. Or do you mean "real" as in cow hoofs and
>silver?
When someone wants to post a picture for the group to see, why
not use the group that was created for that...alt.binaries.photoshop?
It's practically empty, and there is no size limit to the post.
Anyway, it was just a thought....
Talker
Mike Russell
> That is right, till when you begin *editing* the digital "capture". At
> that point, blending precision holds true.
> As soon as you "touch" the source data, you enter the realm of image
> processing math, which includes numerical precision and error.
I would like to invite others to join the digital realm, and participate in
the experiment using the images erpy has so kindly provided:
http://mike.russell-home.net/tmp/erpy/
John J
> That is (also) why professional photographers shoot and edit raw
> images... because you can re-edit again and again without perceivable
> loss of data and precision.
A competent professional doesn't have to edit in such a manner. He/she
knows how to do it right in-camera.
John J
> I would like to invite others to join the digital realm, and participate in
> the experiment using the images erpy has so kindly provided:
I see it is time for me to resign from this discussion. In an hour I am
leaving the house to shoot some 8x10" color film. Enough of the digital.
Mike Russell
> I see it is time for me to resign from this discussion. In an hour I am
> leaving the house to shoot some 8x10" color film. Enough of the digital.
An 8x10 chrome, or any piece of film that size is no small potatoes these
days. Hat's off to you!
erpy
By reading many "pro" forums I'd say they want to be able to get as many
"good" versions as they can...so their clients can choose their fav.
Artistically, I can see re-editing as a good occasion to experiment with
different "moods" of a photo. (it's not just "getting it right")
erpy
The fact he wasn't noticing the banding at first was because Photoshop conversion engine was applying dithering to it - as he reports in his webpage.
As I said to Mike, dithering does mitigate banding - although you can see it on each RGB channel, individually - but the real data you're left with for subsequent editing is the one shown in the BMP found on the page... which suffers from severe banding (i.e. loss of data).
This is a delicate issue we must not underestimate, the same loss is in the dithered version as well since the reduced range of colors stays the same. Dithering only introduces algorithmic noise in order to trick human perception. (again, if that's good for someone, fair enough...)
Now, it's up to you. I gave evidence of the fact that editing a source image/photo at 8bits produces loss of data and its severity is proportional to the number of layers modifying the original data. If you can live with it, keep editing at 8bits... if you want to be sure you're not losing too much data and details, edit at 16bits. (at least)
John J
rounding errors? I don't find many perfect gradients in reality, not
even in a well graduated sky.
How likely is a photograph to require such severe edits that banding
will occur? Who in the world has to iteratively rotate an image several
times?
John J
> John J ha scritto:
>> erpy wrote:
>>
>>> That is (also) why professional photographers shoot and edit raw
>>> images... because you can re-edit again and again without perceivable
>>> loss of data and precision.
>>
>> A competent professional doesn't have to edit in such a manner. He/she
>> knows how to do it right in-camera.
>
> By reading many "pro" forums I'd say they want to be able to get as many
> "good" versions as they can...so their clients can choose their fav.
Shit, then they don't have to do the photo until the client makes up
some image specifications. It's all made up.
> Artistically, I can see re-editing as a good occasion to experiment with
> different "moods" of a photo. (it's not just "getting it right")
Usually, after a hundred traumatic edits one realizes he could have done
it with far fewer and starts over for smoother results.
John J
> On Sun, 12 Apr 2009 10:33:46 -0500, John J wrote:
>
>> I see it is time for me to resign from this discussion. In an hour I am
>> leaving the house to shoot some 8x10" color film. Enough of the digital.
>
> An 8x10 chrome, or any piece of film that size is no small potatoes these
> days. Hat's off to you!
It did not go too well. It was quite windy. I should have brought the
heavy steel monster rather than the Deardorff. At least the bridge
didn't sway.
(I'm sure someone will tell me there is an anti-wind plug-in, or spout
some brainstorm using motion-blur (anti-motion?). :) )
erpy
John, this is not "an accident". We're not measuring "how many times it
happens".... it just happens all the times.
The point is, how much you accept the result. If you do accept to have a
dithered result of your editing (all the times), you're done.... you can
perfectly live with it.
As I said, even a single multply layer will double the banding... but
when Photoshop dithers the result you can't notice it too much.
Although, colors aren't there anymore. (loss of data)
So, consequences to this fact are entirely subjective.
People who decide they always want to keep their original "colors", all
of them, and keep smooth gradients, and re-edit without severe losses,
will edit at 16bit+... people who can set for "visually acceptable" will
continue editing at 8bit.
Mike Russell
when I read some of what he's saying.
I'm bowing out of the conversation for the next day or two, to give others
a chance to chime in and perhaps try out the example image that erpy
provided.