Nikon is NOT on crack - Initial D3x image quality is AMAZING!

[Dan Wells]

I just got out for about 4 hours of serious (landscape and macro) shooting with the D3x this morning, and have been looking at the files for most of the afternoon. The easiest comparison I can make is to the 1Ds mk II (note: NOT mk III), as that is the highest-resolution camera I have a lot of experience with (other than the D3x). I am comparing at ISO 100, converting from 14-bit NEFs (in Nikon Capture NX2) and viewing at 100% on screen (unfortunately, my printer is 180 miles away right now, and I'm on my laptop monitor, NOT my calibrated work monitor, because I am visiting my parents for the holidays).  There are two apparent differences - first is the incredible sharpness of the D3x. When I've nailed the focus, the D3x looks very darned sharp at 100% without applying any sharpening - because of the AA filter, the Canon never did that. It would PRINT very sharp, but 100% on screen always revealed a slight blur. There is absolutely no noise in an ISO 100 D3x file, even at 100%, which adds to the impression of sharpness - very slight shadow noise in the Canon files adds a slight haze to dark ones - that is simply not there in a D3x file. The second difference is the dynamic range - the D3x has about a stop more range in the highlights, plus at least an extra stop in the shadows, maybe even 1.5 stops extra in the shadows, all of it very clean. This camera, properly handled, should print 24x36 inches with ease from its base ISO of 100 (I get 16x24 out of the Canon, but don't like to go larger than that).
     ISO 400 on the D3x is very usable - it looks roughly like an ISO 100 file from a 1Ds mk II in terms of noise - it may have extra dynamic range, which I wouldn't have seen because my ISO 400 tests were on a very dreary, grey day and would have fit easily within the DR of the 1Ds mk II). This is comparing on a per-pixel basis, so the D3x file still has 3/2 the detail in it, due to the increased resolution... I have even fooled around a bit with ISO 3200 (HI 1.0), which looks awful on screen (although quite good considering that it's ISO 3200 - much better than any ISO 3200 film ever looked), but will make a pretty decent 8x10 print with no trouble, and should even print 11x17 with some careful handling. I didn't buy the camera to shoot at ISO 3200, but it's nice to know the capability is there should it be needed. Those unbelievable ISO 100 files are what I bought the camera for, and it is certainly worth its price for its low-ISO performance! There is something highly unusual in the imaging chain of the D3x to get these results - the sensor may NOT be stock Alpha 900 issue (I suspect it isn't - I don't have a lot of Alpha experience, but the test files I've seen are not anywhere near as clean, even at low ISOs) , and if it is, the AA filter in the D3x is extremely unusual, probably made out of pure unobtainium.
     Add that performance to a superb rugged and ergonomic camera body with class-leading AF and metering, and the result is a remarkable machine. Yes, it's expensive, but the only way to get better files is three times as expensive and not nearly as rugged.

                                                                          -Dan

   

[Tony Beach]

Yes, it's expensive, but the only way to get better files is three times as expensive...

I suspect that depends on the photographer and other variables:  http://forums.dpreview.com/forums/read.asp...essage=30468356

[Panopeeper]

The second difference is the dynamic range - the D3x has about a stop more range in the highlights, plus at least an extra stop in the shadows, maybe even 1.5 stops extra in the shadows, all of it very clean

The DR of the D3X @ ISO 100 is about ONE stop higher than that of the 1DsMkII; this measured, not estimated. I don't have suitable raw files for conclusive comparisons with different ISOs; if you create some, I can do the measurements.

The D3X appears to be an excellent camera; no need to make phantasy claims.

Btw, digital cameras do not have separate dynamic ranges in the highlights and in the shadows. There is only one dynamic range, everything else is the question of metering.

[douglasf13]

I don't think anyone doubts that the D3x is a superb camera. It looks to be the top of the 35mm heap right now  From the tests I've seen, the D3x and A900 are very similar until ISO 800, and past that the Sony lags. The question is whether the D3x is worth 2.5 times the alpha, and that's a decision unique to each shooter. I'm curious to see what an Alpha firmware update brings.

[Ray]

It's becoming increasingly difficult to extract more DR and S/N from 35mm sensors. With the D3 and D700, Nikon caught up with Canon and actually surpassed Canon on the noise front. This surpassing was subject to much exaggeration, but after the hype and excitement had settled down it was discovered that these Nikon cameras had about a 1/2 stop DR and noise advantage over the ancient 5D. Furthermore, on the basis of comparison of equal size images, the D3 high-ISO noise is on a par with (or as close as matters) to the 1Ds3.

The initial impression that Canon created with the introduction of the 50D was that individual pixel performance might be close to that of the 40D. In fact, we now know that at the pixel level the 50D is noisier than the 40D. However, at the total image level, it's about the same.

The big question with D3X noise and DR performance is this. Can we expect a higher DR and S/N than the D3 produces, when D3X files are downsized to 12mp?

[Panopeeper]

The big question with D3X noise and DR performance is this. Can we expect a higher DR and S/N than the D3 produces, when D3X files are downsized to 12mp?

Ray, I got used to your equalization fanantics (=fanatic antics), but this one surprized me. Your faith in the myth of eliminating noise by downsizing is one thing; but how on earth do you imagine, that the DR would be increased by downsizing? Honestly, do you have an idea, how downsizing is supposed to effect the noise? That is the recipe for lowering the DR

I suggest you to make a test finally, not only talk about it. I did, with one of your 5D shots for the DR measurement (with Jonathan's pattern) and with your #9199, Ratana's Kitchen, in 100%, 75%, 66% and 50%. I do not post the result, a layered TIFF, for it is over 144 MB.

[Ray]

I suggest you to make a test finally, not only talk about it. .

I make frequent tests, Gabor. Haven't you noticed? However, my instruments for measuring the results are my eyeballs. I rely upon people like you to provide the mathematical confirmation of what I'm seeing and to raise issues that might not be readily apparent with normal viewing of images and prints.

Dynamic range issues are also subject to exaggeration. The A900's DR advantage in relation to the 5D2, for example, exists only at ISO 100 and, from memory, is only about 1/3rd of a stop, according to DXOMark.

Nevertheless, I confess I'm not totally clear in my mind as to the distinction between DR and S/N. All else being equal, if one were to increase S/N by reducing noise but keeping the signal the same, one would increase DR. It would be theoretically possible to produce a P&S camera with the DR of a Phase DB, if someone were to develop a new way of recording the signal which was essentially noiseless, say a material which was totally insensitive to any radiation other than the frequencies of light, and which was at the same time extremely sensitive to the frequencies of light, more sensitive than current sensors.

[Ray]

Gabor,
Perhaps we could shed some light on this issue by stating a principle that a small sensel with an excellent S/N can have the same (or greater) DR as a big sensel with a relatively poor S/N.

Would you go along with that?

[ejmartin]

Ray, I got used to your equalization fanantics (=fanatic antics), but this one surprized me. Your faith in the myth of eliminating noise by downsizing is one thing; but how on earth do you imagine, that the DR would be increased by downsizing? Honestly, do you have an idea, how downsizing is supposed to affect the noise? That is the recipe for lowering the DR

Noise and DR are scale dependent quantities.  Any quotation of figures for either that does not take this fact into account are figures without context and meaning.  Downsizing eliminates high frequency noise, because downsizing eliminates high frequency everything (noise, detail, etc).  The result is that, while noise at a given spatial frequency remains unaffected, the noise as measured by standard deviation of uniform tonality patches is lowered, because of the elimination of high frequency components of noise.  Similarly, DR at a given spatial scale is unaffected by downsizing, insofar as the given spatial frequency is still present after downsizing.  DR as measured by the std dev noise floor is again increased by downsampling, because the std deviation of noise is an amalgamation of noise at a variety of spatial frequencies, some of which are eliminated by downsampling.

[Ray]

Noise and DR are scale dependent quantities.  Any quotation of figures for either that does not take this fact into account are figures without context and meaning.  Downsizing eliminates high frequency noise, because downsizing eliminates high frequency everything (noise, detail, etc).  The result is that, while noise at a given spatial frequency remains unaffected, the noise as measured by standard deviation of uniform tonality patches is lowered, because of the elimination of high frequency components of noise.  Similarly, DR at a given spatial scale is unaffected by downsizing, insofar as the given spatial frequency is still present after downsizing.  DR as measured by the std dev noise floor is again increased by downsampling, because the std deviation of noise is an amalgamation of noise at a variety of spatial frequencies, some of which are eliminated by downsampling.

Thanks, Emil, for weighing in here. I'm totally mystified as to why Panopeeper is so resistant to this idea that downsampling reduces the perception of noise. It's a phenomenon which is easily visible on one's monitor and on print.

Thanks also for pointing out that DR is also scale dependent. I wasn't sure about that.

[Roy]

There is an interesting article in the editorial blog at DP Review on downsampling and noise.

[Dan Wells]

Where did anyone see a measured DR on the D3x - I'd be really interested to see what it is (and I suspect that the lower the acceptable noise, the greater the difference from an older camera like the 1Ds mk II)? Imatest reports four separate DR figures, depending on acceptable noise level (which I like better than DxO's method of reporting only a single figure which corresponds roughly with Imatest's widest and highest-noise figure). Depending on which noise figure you use and how you make the measurement, the 1DsII is between half a stop and 1.5 stops below the Alpha 900 in measured DR, and I would be surprised if the D3x wasn't wider than the Alpha 900 at low noise levels, because the D3x files are so clean (I've never shot an Alpha, but the sample files I've seen are not extraordinarily clean - if anything, they are noisier than I might expect - even the RAWs). As I said earlier, I'm presently viewing files without a calibrated monitor or a printer, but the D3x results are really blowing me away. It will be interesting to compare with better tools right after the first of the year.

                                                         -Dan

[Panopeeper]

However, my instruments for measuring the results are my eyeballs

There is nothing wrong deciding what you like based on visual inspection. In fact, that is the perfect method.

The usual problem with such comparisons is, that important aspects are not involved. I see very often demonstrations (of incompetence), how low or how high the noise is of a certain camera at a certain ISO, with such specifications like "in a dimly lit kitchen" (and usually noise reduced, downresed, the blackpoint cuts off everything in the really noisy region, etc.).

The A900's DR advantage in relation to the 5D2, for example, exists only at ISO 100 and, from memory, is only about 1/3rd of a stop, according to DXOMark

This is exactly the result of my measurements. However, I am cautios with the A700 and A900 raw files; I can not interpret them properly (nor can ACR).

Anyway, I suggest you to convert for example the #9199 without NR in ACR, pass it to PS and create different downsized versions of it. There are some good spots to measure the noise, like on the side of the huge clay-like planters at the left side. Make a selection on it and look at the statistics under Histograms, Mean and Std Dev.

Please post the result (change of StdDev) here.

[Panopeeper]

The result is that, while noise at a given spatial frequency remains unaffected, the noise as measured by standard deviation of uniform tonality patches is lowered, because of the elimination of high frequency components of noise

Like in this example, right? Canon 40D @ ISO 1600; different sizes are in layers. I made it easy for you, you need only to make a selection in PS and look at the standard deviation.

4.5 MB demo TIFF

[ejmartin]

There is an interesting article in the editorial blog at DP Review on downsampling and noise.

That blog post was thoroughly debunked in a couple of threads in the DPR forums:

http://forums.dpreview.com/forums/read.asp...essage=30176643

http://forums.dpreview.com/forums/read.asp...essage=30190836
http://forums.dpreview.com/forums/read.asp...essage=30211624

Gabor, you will find my demonstration of the effect of downsampling in the latter two links.

[Panopeeper]

Gabor, you will find my demonstration of the effect of downsampling in the latter two links.

Emil, you will find my demonstration on the effect of downsampling in the TIF I linked to above.

[grepmat]

Reducing (or increasing) the size of an image can be done in many ways. They will affect noise differently. All this talk about frequencies, etc., is useless without discussing the particular sampling technique used. For example, many of the most common point-sampling techniques will not affect "per pixel" noise at all!

Dynamic range, as it is most commonly defined (or at least as it should be defined) is just the maximum signal (the point of saturation) divided by the root-mean-squared pixel noise in darkness. Only if the noise is changed will dynamic range be changed, since virtually none of these scaling methods affect the maximum signal level. The noise "of an image," i.e., of anything other than an unclipped image taken in pure darkness, is also a relatively meaningless, irrelevant distraction when discussing dynamic range.

If we are talking about camera-level pixel averaging ("binning"), something that few cameras do in a true sense, then if four pixels are binned, the noise will be reduced by a factor of two (the square root of four), and the dynamic range will double. What photoshop, etc., does is quite different and may affect the noise little or not at all. This is most likely why people do not see dramatic reductions in noise (and concomitant increases in dynamic range) when scaling.

[ejmartin]

Emil, you will find my demonstration on the effect of downsampling in the TIF I linked to above.

Your linked file shows several things:

1.  The chosen test image has relatively little noise power at high spatial frequency.  This is why the "grain size" is fairly large.  Downsampling will reduce noise by less than the ratio of linear dimensions, since the downsampling removes only the high frequencies (assuming it is done properly).
2.  You downsampled in PS using Bicubic sharper, as near as I can tell by looking at the 50% reduced image, and overlaying the original downsampled in PS CS3 by that method.  The sharpening of course enhanced the noise of the reduced image, as would sharpening the unreduced image.  As a result, the standard deviation of noise didn't change much at all in the downsampled versions.
3.  A straight bicubic 50% downsample reduces the noise standard deviation from about 22-23 in the patches I examined to about 18-19.  PS bicubic has the propensity to alias noise into the downsampled result -- that is, noise in the original which is beyond the Nyquist frequency of the target is sampled into the target image.  A gaussian blur of radius about 1 or so prior to the 50% downsample does not affect any spatial frequencies of the target image, but dumps the noise power beyond the target Nyquist frequency.  By this route the noise standard deviation was reduced to about 13-14 without affecting detail of the downsampled image.  Uncorrelated noise would have reduced to about 11 by this method, the fact that it wasn't reduced quite as much is due to the relative lack of high frequency noise, as I mentioned above.

In contrast, the test image I examined had noise power uniform almost out to Nyquist.  Resampling by a Lanczos filter using ImageMagick fairly closely reproduced the expected reduction of noise upon downsampling.   Details may be found in the links in my initial post in this thread.

[ejmartin]

Reducing (or increasing) the size of an image can be done in many ways. They will affect noise differently. All this talk about frequencies, etc., is useless without discussing the particular sampling technique used. For example, many of the most common point-sampling techniques will not affect "per pixel" noise at all!

Indeed.  Nearest neighbor for instance does nothing for noise.  Bicubic sharper will reduce the noise via bicubic and then enhance it by sharpening.

Dynamic range, as it is most commonly defined (or at least as it should be defined) is just the maximum signal (the point of saturation) divided by the root-mean-squared pixel noise in darkness. Only if the noise is changed will dynamic range be changed, since virtually none of these scaling methods affect the maximum signal level. The noise "of an image," i.e., of anything other than an unclipped image taken in pure darkness, is also a relatively meaningless, irrelevant distraction when discussing dynamic range.

Noise is not a number, it is a function of spatial frequency.  Therefore a S/N ratio can be defined for each spatial frequency, and dynamic range computed for each spatial frequency.  In this sense, the noise of an image is quite well defined.  When comparing cameras of different pixel counts for noise and dynamic range, the proper and well-defined comparison is to examine the noise, S/N, and DR at a reference spatial frequency.  This is roughly what DxO does when they scale results for "print output", under the assumption that the noise is uncorrelated.  For a sufficiently accurate raw conversion, which does no or very little noise reduction (unlike Gabor's example), the noise is little correlated until quite close to Nyquist.  For instance, the DPP conversions I used in my examples have a rather uncorrelated noise spectrum.

If we are talking about camera-level pixel averaging ("binning"), something that few cameras do in a true sense, then if four pixels are binned, the noise will be reduced by a factor of two (the square root of four), and the dynamic range will double. What photoshop, etc., does is quite different and may affect the noise little or not at all. This is most likely why people do not see dramatic reductions in noise (and concomitant increases in dynamic range) when scaling.

Photoshop bicubic is a rather poor filter for downsampling.  It can alias noise into the target image.  One way to deal with this is to blur the image slightly before downsampling.  But I disagree that the dynamic range, noise, S/N are increased in a meaningful way by downsampling.  Pixel peepers may see a reduction in the noise std dev, but that is because the std dev is measuring different aspects of the noise spectrum for the original and the downsampled image.  In a proper downsample, the noise at fixed spatial frequency is unaffected.  This is shown in the links in my initial post in this thread.

[Panopeeper]

1.  The chosen test image has relatively little noise power at high spatial frequency.  This is why the "grain size" is fairly large

This is the way the Canon 40D works, converted by ACR, NR set to zero. The subject is a grey strip wedge, very uniform.

The noise is lower if converted by DPP (NR set to zero). One could speculate about the reason: if DPP is doing some NR despite NR set to zero (ACR is doing so), or if ACR mistreats the 40D raw data. However, I am using ACR as are most photographers.

2.  You downsampled in PS using Bicubic sharper, as near as I can tell by looking at the 50% reduced image, and overlaying the original downsampled in PS CS3 by that method

Of course I did. That is the suggested method for downsampling. Perhaps you choose a method, which is suitable to create a "proof", but I choose the best method for the purpose, namely downsampling with the least loss of details.

Anyway, the difference in noise between "bicubic sharper" and "bicubic" is nothing to be excited about.

As a result, the standard deviation of noise didn't change much at all in the downsampled versions

Oh well... At least I did not select the methods of creating the samples with the goal in mind to prove anything. I did this just like I am doing it normally.

I disagree that the dynamic range, noise, S/N are increased in a meaningful way by downsampling

Well, then where does the myth come from?

Anyway, did someone claim that the noise increases by downsampling? Ray claimed, that the DR increases by downsampling. I find that nonsense, even though the noise can be decreased by downsampling.