I thought the Canon ISO1600 shot seemed better, whereas the Nikon high ISO shots all seem to have heavy noise reduction in them, less detail, a sort of painterly smudginess about them.

Graeme

Graeme,

The consensus view of 22 pairs of eyeballs at my gallery last week was that the D3 ISO 1600 files looked better. But, this just goes to show how subjective these evaluations can be.

Michael

Michael,

Thank you ever so much for this VERY IMPORTANT review. As you may have seen - if you've had the time - I started a thread on the subject of noise, and what you've written is so very topical to what I'm experimenting with just now - but without the luxury of a slew of cameras for comparisons. The comparisons however are important because they put the whole issue in perspective. Over the weekend I hope to be gaining insight into a question on my mind as to whether there is any variance of sensor performance from one camera to another within the same camera model. I would hope not - at this price point (D3, 1DsMK3) one must presume they are individually tested to meet a minimum standard that is also a very high standard.

Another important aspect I believe, based on the preliminary testing I've managed to complete over the past couple of days, is how one exposes and processes the images. For example, if you look at the crop of Jay's eye in the section of your report starting with "ISO 1600", from what I can see on my well-calibrated and profiled LaCie 321, your observations about these two images are correct. But I would go one step further and observe that the Nikon image is generally more contrasty - nothing wrong with that - just a fact which perhaps needs to be thrown into the list of explanatory variables. One can see this insofar as, like you say, there is more shadow clarity in the Canon shot (obversely Jay's eye is darker in the Nikon shot), the skin tone in the Canon shot is darker than in the Nikon and there is more eyebrow detail in the Nikon shot, because there is more aggressive tonal separation between the hairs having different shades of grey. Now - I want you to understand, please Michael - this is NOT *pixel peeping*, but it IS *splitting hairs* (literally) . The serious point about it is - one wonders whether the higher contrast of the Nikon image either suppresses noise or causes less visibility of noise than does the slightly lower contrast Canon image, especially if the causes of these contrast differences relate to any differing placement of identical subject along the histogram. I.E. if more of the skin tone in the Canon shot is further to the left on the histogram than happens in the Nikon shot, could this perhaps explain why there would be slightly more visble noise in the former compared with the latter? I'll have a bit more to say about this matter related to my own images in the thread I started. I had hoped to publish it last night but I fell asleep in front of the TV watching the news. (Yes, for our Canadian readers, the Mulroney-Schreiber hearing was good theatre, but it put me to sleep after a long day of shooting and pixel peeping.)

One other technical point - from what I saw of the images you printed at the gallery during the workshop, it is not only the amount of noise that has changed, but also how it looks. While I have not yet done the intended comparison with my 1Ds 11.1MP, I have a sense that the "look" of the noise isn't what it used to be. Except in the deep dark tones where the green/magenta blotches look familiar, the darkishness of other noise further up the tone curve is - well - less dirty-looking and more granular, which harking back to the film days many would consider more "acceptable".

The three bottom shots in your article are astounding in terms of the clarity, tone quality and resolution achieved at such a high ISO.

Cheers,

Mark

The bottom line from these tests -- 2008 is going to be a very excitimg year. Thanks Michael.

I must say...nice informal comparison. And good real world conclusions. I too belive the D3 has a slight edge now over Canon...(which I am sure Canon will answer) and if such a race ensues...we the users of each brand will only benifit.

I just sold my D2X to get the D300 and I cant be any more pleased. Yes...the D300 has more noise issues than the D3. But it is worlds of improvment over the D2X.

I would be very interested to know what you think of the new lenses also from Nikon (the 14-24 and the 24-70 that just came out) I have heard some good things even from Zeiss users. Would love to hear a more informed and trustworthy opinion.

Thanks Michael for your intersting read...and as always...your timely information.

Roman

Indeed, highly subjective! So let me rephrase my question:

Why do the images look the way they do?

As you know I'm more interested in this from the POV of sensor design and what is state-of-the-art given my current work, but I'm still interested as a photographer.

It's fairly quick and easy to take both the Canon and Nikon images into Photoshop, go into LAB mode and see what A and B, the chroma channels are doing. The Nikon here shows much less detail, which is indicative of noise reduction that's strongly aimed at chroma noise. Given how these sensors works, that makes sense. However, it also means we're not comparing like with like.

Does putting noise reduction algorithm A in camera, help or hinder you using noise reduction algorithm B on post? Is it better to do all NR in post, or is some split more effective?

One of the things that it's very hard to do is actually visualize raw camera data. Without seeing that, you're not seeing what noise is the sensor and what noise is the processing. You're not using like for like demosaicing algorithms (which can do noise reduction as part of their action). All you can do is compare the end result, which is what matters for photographers - it doesn't matter how you get there, as long as you like where you're going, but doesn't always elucidate the optimum route in terms of time, effort and money.

Adrian's comment about contrast is a good one, as shadow noise can be somewhat changed in appearance by cropping off the noise in the near blacks, giving an elevated black level, but having it effectively noise free, and increase contrast, but reduced dynamic range. You can also loose some detail doing this.

So to me, the other question is, not what is the comparison of noise, but what is the dynamic range of each camera, and to measure that, you need a controlled test with a backlit wedge test. I'm sure this information will come out, but I'm keen to know if the D3 sensor is revolutionary or have Nikon just got good NR in there.

Graeme

Graeme, the point about contrast was mine, not Adrian's. I was thinking more in terms of the initial exposure, implicitly suggesting that if the same material ends-up on different spots of the histogram between the two cameras this can contribute to different noise levels without differing sensor quality being the determinative cause. Are we saying the same thing in different words, or are you adding another dimension to the concept?

Higher DR suggests to me more space between white and black - but how would this impact noise?

Graeme,

For this to make sense it would have to be done with the raw file, not a JPG off the web. Also, one needs to consider what inherent noise reduction has been performed by the raw processor, which all do some even with settings at zero.

Finally, while looking at the file in LAB is interesting, it does not always correlate well with how the image actually looks in a print, which for me is where the rubber meets the road.

Michael

Thank you for another informative comparison. I am a Canon shooter and have never been a Nikon shooter, but it is exciting to see real competition. The end result can only be better tools for us.

I understand your comments about DPP and Capture NX, but I look forward to seeing (somewhere, if not here) comparisons of these proprietary developers along with Lightroom.

DPP is terrible for workflow, but to my eye often gets cleaner output than LR/ACR.

Jay has great eyebrows.

Michael,

A very informative comparison showing how the cameras stack up in real world shooting conditons. As to noise, the quants will be interested in standard deviations at defined areas in the shadows, perhaps with un-demosaiced data to avoid the variables introduced by the raw converter. Other things being equal, the camera with the larger pixels will have a lower standard deviation. However, the noise spectra will be different for the two cameras. The camera with the higher resolution may have a larger standard deviation, but the noise will be finer grained and perhaps perceived as less bothersome.

Did you actually make 16 by 14 inch prints to compare these subjective differences?

Bill

Well, the D300 only has slightly less read noise than the D200, and about the same image shot noise and greater pixel shot noise than the D200. It is nothing revolutionary. It should not be included in the class of low-noise performance in the D3 or 1DSmk3.

For the D3, the JPEGs look like Nikon is so used to having higher noise that they forgot to use less NR on the D3.

And the good news is, both Canon and Nikon users have reason to be happy! Merry Christmas & Happy Hanukka!

I will have to say both cameras is impressing. I would be happy to have any off them. As a Nikon shooter I am very pleased with Nikons better noise performance. Myself doesnt choice cameras from the noise the most important factor.

I am very impressed with the Canon 1ds MKIII. Even with the almost the double pixel count it holds up very good. I dont think it is night and day. I can understand that noise is important for some people for their style off photography. But it isnt such a "big deal" like some self-chosen forumguru experts made a point off.The evolution off camera sensor technology is nothing about fantastic on both brands.

I thought my puny amateur Nikon D80 got good noise . But I came from another world today some people would claim such noise level render the camera to disposable camera.

If you are pro working hard lighting conditions I understand the obsession with low iso noise. But I almost irritated today as it is hard to read anything about image quality that dont concern high ISO image quality.

I think most people are obsessed with no good reason as most people dont print large, maybe this forum is exception with many landscape guys. ISO noise is often very forgiving in large prints when cram in alot off pixels in alot smaller area than a computer screen. 72 dpi is very low resolution and off course you can even see a dead pixel. I challenge someone to see a dead pixel in print, maybe you can. But then you see alot better than me.

Nothing to write down the effort off the comparison between the cameras. But I want shake some sense into people that both cameras are probaly renders great prints.

I want to read about renderings, low iso, color, AA filter and other stuff that is relevant for my kinda slown down tempo shootings. I am just as egoistic like the other guy.

Quite a long post, but wanted share my view. This forum is also grown up enough so I took my time share my view. Thanks if you read this far

Cheers,
Daniel

While I'm not arguing with the way the "tests" were performed as to the question of image size vs. pixel size, my experience in this has given me some reasons for going with the image size crowd.

I was a partner in a commercial photo lab here in New york for many years, and have seen mucho photo's come through, in addition to my own early work in fashion and advertising. The one thing I've seen time and again, is that the final image size is the thing. Whether using film or digital, it must be sized to the final purpose.

In regards to that, and the medium used, we see the differences in format, and sensor size, and rez. The same for film. ISO, obviously plays a very big part.

An 11 x 17 two page mag spread will remain the same regardless of the original medium used. The same of a gallery print.

Assuming the lenses are capable of resolving close to the limit of the film or sensor, the rez of the medium will be a determining factor. This, I have seen, affects noise as well. What I have found is that, talking about digital (though film is similar), a sensor with significantly higher rez, assuming equal noise at equal final pixel size, shows less apparent noise on the image. The reason is simple, the pixels are much smaller, and the noise is therefore finer, and less observable. Since we purpose our images for a specific image size, that's what really matters in the real world.

Of course, if one is a "cropper" things may come out differently.

But, otherwise, it's like printing a 16 x 20 of an ISO 100 film vs printing it from a 400 ISO film. If the 100 film grain is enlarged to equal that of the 400 film, the "noise" may be about the same, but for the same image size, the 100 film will have the lower noise, and greater sharpness.

I realize there are various ways of looking at this question, but I've never seen anyone print to the grain size for any normal work. The medium chosen is usually that which will give the required quality at the needed image size.

Thanks for this comparison; these cameras are amazing. ISO 25600!!

I would like to natter about comparing noise characteristics of cameras with different resolutions, and will try to be grown-up and polite:

As you allude to in the article, in general the "problem" with high-resolution cameras as that each photosite/pixel is smaller, so it collects less light, and therefore needs to be amplified more for a given ISO level than a photosite on a lower-resolution camera with the same size sensor. This higher amplification, of course, causes noise. Now, in reality this isn't much of a problem, hence the quotes around "problem," because you can "get it back" by lowering the resolution of the picture. If you use a good algorithm for lowering the resolution, the noise will be reduced, yielding a picture of similar resolution and quality as from the lower resolution camera. I'm assuming PS will use a good algorithm, but I don't own it or use it so I'm not sure.

So - I think the best way to compare the noise characteristics of a full-frame 21mp camera and a full-frame 12mp camera is instead of showing two 100% crops, show a 100% crop of the 12mp camera and a ~75% crop of the 21mp camera. This should result in two crops that are the same size on screen and show the same crop of the picture. Notice the two 100% crops in the article are of different sizes, since the Canon sample has more pixels for the same composition.

I suspect this will eliminate some of the "one stop noise advantage to the new Nikon" that you see, and this is a very meaningful comparison, since the real question is: "For a print of a given size of such-and-such scene, which camera will be less noisy?"

thanks,
Erik Haugen

Sceintific sensors, which are nearly always CCDs rather than CMOS, allow pixel binning to achieve the advantages you mention. As explained in the link, binning reduces read noise, since the binned pixels are read out with about the same noise as the individual pixels. When reading individual pixels, the same read noise in inserted with each read. After the fact binning will reduce noise (shot noise), but will have no effect on read noise.

Bill

I think this is broadly true although the reason is not necessarily that the pixels of the higher rez camera are smaller, when different formats are being compared.

I always remember my surprise when I first read the review of the 1Ds at dpreview comparing its noise characteristics with the earlier Canon D60. Up to about ISO 400, noise in the 1Ds was actually slightly worse than the D60, yet the 1Ds had bigger pixels.

Of course, when dpreview compares noise, they compare it on a pixel-for-pixel basis; grey patches of equal size.

But the point was made by dpreview that this apparent lower noise of the D60 would not be evident in equal size prints and I'm sure those who used both cameras would have been very much aware that the 1Ds had lower noise in practice.

I can appreciate the whole "noise" debate and recognize that it's a concern but what interest me more as predominately a studio shooter at ISO 100, is a direct comparison as to how well these cameras can hold up when enlarging an image to 24x30.

Michael

If you read Michael's postscript, you will see that the shots were in raw and rendered in Lightroom. They are not in camera JPEGs. Of course, Nikon may have done some NR behind Michael's back to the raw file.

Bill

A 24X30 print would show a clear advantage to the Canon. There's no way that a 12MP camera can make a print that size the way a 21MP camera can.

And as I indicated in my article, at print sizes that are within the capability of both cameras the IQ differences at low ISO would be a quibble, and likely subject to the vagaries of processing as much as anything else.

Michael

Congradulations Michael for this very interesting and highly informative quick take!

Nevertheless, to me, the comparison can be better, if we taking in the following factors.

1. The Nikon 24-120mm VR is not really a sharp lens at f/5.6. But this does not affect the noise behavior of course and high ISO noise was Michael's main aim.

2. The one which does affect noise the the RAW converter. ACR/LightRoom in general produce much noisier images from the D3/D300 NEF files, I have found, compared to Capture NX, even when the NR is turned off in the NX. On the other hand the handle the CR2 files much better, to my eyes.

Why? Actually there has been quite interesting discussions on going at Bjørn Rørslett's forum on both NX's (and possibly Nikon's) noise and behind scene CA reduction issues:

http://nikongear.com/smf/index.php?topic=6829.0

, talking about the possible hidden information in the D3/D300 NEFs which I am almost sure the ACR/LightRoom is not current consider, if they do exist.

By the way, I have seen many 1DsIII samples and exactly like Michael I am extremely impressed. Still shooting film, I am tireless to try to squeeze out the last bits of fine details relatively artifacts free from digital captures. Based on what I saw, if I spend serious money on a DSLR today, I would not hesitate to go for a 1DsIII, since I am basically a landscape photographer and the 1DIII/D3's high ISO capabilities does not matter too much. The pixel level fine details and micro-contrast contained in the 1DsIII RAW files are truly astonishing, basically at the same level of 18-22MP MF backs if the right processing tools are used, despite of the presence of the AA filter. Even I label myself a Nikon shooter and lover! In the past I argued with many others on the film vs. digital subject, but even with my more conservative accessment the 1DsIII resolution has passed the 645 Velvia 50 level. For example there is a 1DsIII/MF back comparison here:

http://forums.dpreview.com/forums/readflat...25835108&page=1

Talking about resolution, and again like Michael, most importantly at low ISO values, there has been reports that the D3 images does not contain ultra-fine details, like these from my mentor and friend Lloyd Chambers:

http://www.diglloyd.com/diglloyd/blog.html

The subject is heavily debated here:

http://nikongear.com/smf/index.php?topic=6911.0

For one I tended to agree with Lloyd: without access a D3 I too found my D300 images not to the crispy D2x image level with more soft, maybe more "(negative) film like" feeling which I do not really like (I like chrome and B&W film feelings). I thought the same could be the case for the D3 until yesterday, when Imaging-Resourse posted the RAW files of their preliminary test images. I downloaded the NEFs, and played carefully in different RAW converters, ACR, LightRoom (basically the same to ACR but ACR is just updated), Capture NX, Phase One, and my perferred RAW Developer (RD). At the end, I found the RD does a much surperior job digging out that last ounce of fine details out of the D3 RAW files, although it also outputs more noise (which I do not care), and it lacks the automatic CA reduction feature in the NX.

Unlike most reviewers/testers, I do not look the converted images merely at 100%, but upsample them to a standard size, which is 16x24 at 300dpi, or 4800x7200 pixels, as my comparison base to exam the real fine detail and artifact level. This link brings to the RD processed D3 RAW files to see the level of fine details there:

http://nikongear.com/smf/index.php?topic=7081.0

To my eyes, at the 170% upsample level, the ACR/LightRoom images look plasticky with the fine textures on smooth surfaces totally wiped out. ACR only exports at 25.1MP or 4088 pixels at the short site maximum, so to compare the ACR converted with mine you have to further upres. There is a reason Charlie Cramer and many other used to be film-based fine photographers use RD for RAW conversion. A direct link to the RD output, at 170% resolution, is here:

http://lepingzha.smugmug.com/gallery/50676...229584488/Large

which was from the RD's Hybrid sharpening set to 1/10. Some small crops are below. No Photoshop involved in processing except for saving JPEGs at quality level 10.

Remember you are looking at a 16x24 blow up at 100% on screen. The little mostly demosiacing artifact will disappear on print, or at 50% viewing level, which matches printing results better.

Thank you for reading my long post!

I consider Michael to be the type of guy to get into something and give a very solid opinion or report on whatever he has hi hands on. Even if I don't agree with the outcome, there is enough weight behind his thinking to make his words command attention.

Having said that, there are a few points I would make on some of Michael's statements. One thing that really isn't that important, but also can't be allowed to just fall to the floor is while understanding and appreciating the result of the comparison of the 1Ds MK3 and D3, it is just enough of an annoyance to see that the 1D MK3 and D3 are the real comparable cameras and not the D3 and the 1Ds MK3. This should be kept in mind for the future and just because I mention it it does not mean any disagreement or lack of understanding on my part that the MK3 was not available. It means it is something important enough in my eyes to do in the future.

Two area's where I do have a disagreement is the basic underlining premise that bigger pixels automatically mean lower noise. Canon's entire advancement this last round is based on smaller pixels but at the same time keeping the same size photo receptors as the previous generation of sensor. If the Canon 1D MK3 proves to be fairly close in noise to the Nikon D3, then the Canon is accomplishing that with a smaller pixel - but that test was not able to be done. The 1Ds MK3 however, has a smaller pixel but the same size photo receptor inside the pixel and manages to achieve better results than the 1Ds MK2. The 1D MK3 has a bigger pixel than the 1Ds MK3, so where the chips may fall is still undetermined.

Having the 5D and the MK3 I can tell you that shooting a wedding and in the part of the church service where there is no flash, I much much prefer using the MK3 over the 5D at ISO 3200 and the results are about a stop or more better. Even though I know I am right on this - assuming someone disagrees with me, what this proves is that the next generation of the 5D is poised to become the king of the hill at low noise if Canon uses this same technology on the next 5D and Canon also stays conservative with the amount of megapixels they put into the camera. I would guess the extreme would be something in the 16 megapixel size and a conservative approach would be something in the 14 megapixel size. At the 1Ds MK2's 16.7 megapixels with the new smaller-pixel/larger-photo-receptor technology, that sensor should yield the equal to Nikon's D3. At 14 megapixels it should pass the Nikon's D3 easily. If nikon was able to do the same thing that Canon does with a smaller pixel and larger photo receptor, it would take that camera to a whole new level, Canon is the only one in position to accomplish that now, if they chose to do it.

I also want to add an interesting test in making 12x18 lab prints from both the Nikon D3 and the Canon 1Ds MK3 and having about 10 people choose which was better, the Canon won 9 out of 10, and I also chose the Canon print. The person choosing the Nikon print chose it based on the facial expression - duh. The prints were ISO 3200 images from both cameras.

Peter

Convince me it is fair to make 12x18s out of a 12MP camera to be compared with the one from a 21MP. D3 vs. 1DIII is far more fair since the D3 and 1DsIII are targeted for different applications, just as one guy said you can't compare a Ferrari with a truck.

Don't take me wrong since I will probably get a 1DsIII, but I am a landscape shooter not a sport chaser or photo journalists.

Plus, you should checkout the sensor efficiency test results where the D3 has a clear lead. Maybe the 1DsIII's is even higher, but sorry, no data yet.

http://theory.uchicago.edu/~ejm/pix/20d/po...D300_40D_tests/

The subject is also heavily discussed here:

http://forums.dpreview.com/forums/readflat...25658562&page=1

You see more noise on your prints because the improper RAW converter (ACR) was used. See my post above it does not treat CR2s and NEFs equally well and output heavy noise from the Nikon RAW files. Try out Capture NX and see the difference. We should always compare with the manufacture's own converters, i.e., DPP and NX. ACR or LightRoom are NOT the equal base for all under the same settings, for sure.

I come in here through the back door. When someone starts a thread, referencing an article from the main page, it might not be obvious that it was an article that was being referenced. So, naturally, I thought of default conversions or JPEGs.

I don't think that converters that convert many cameras and brands convert all images equally. You stand the risk of upsetting people if their cameras' RAWs convert noisier in you converter, or are more detail-less, so there is motive to try to stick with the MFR's style as closely as possible, unless you can make an improvement without compromising something else (such as ACR's line noise reduction).

I've looked at the RAW data of all three cameras, and they all look pretty honest, with no signs of filtering (noise concentrated near nyquist). The only tampering is with blackpoint, highlight clipping, and rescaling of RAW values on a global or line-by-line basis, etc.

Well, those willing to spend US$8000! The 40D still collects only about 1/4 as many photons as a D3 with the same exposure, and has more than double the individual pixel read noise as the 1Dmk3 at ISO 1600, while having less than half as many pixels!

I totally agree, Michael. I broke my own (16x24) rule and upsampled (in RD) and printed the Imaging-Resource's 1DsIII RAW test image to 24x36 (part of it of course) at 300dpi, and it look great (even I can't stand the most people's 12x18s out of DSLRs).

I hope you would consider going back to China someday with real local photographers, and I will be happy and more than thrilled to be your local logistic manager and interpretor, as I did for Richard Lohmann and George DeWolf last year at this time!

Michael,
Thanks for this practical test. I found it useful, and as always you concentrate on what matters in real world use for those in the market for this type of camera.

A while back I was watching LLVJ #1 and the review of the D30. You held up some large prints and showed just how wonderful the results could be from a camera that might now be called primitive.

The cameras in this test, and others like the 40D and 5D, offer results that are simply spectacular. We can quibble endlessly about technical details, but from the cameras now available there is certainly something to fit the needs and budget of most photographers.




Perhaps we need to ask: is there a camera available that can allow me to print to the size I want and produce excellent results. The actual specification measurements matter not at all. It's the end result we should be measuring.

I'm not talking about some philosophical abstraction; all "end results" originate with signal and noise; it's all recorded as numbers, and some have more numbers, some have less numbers, some have more accurate numbers, some have less accurate numbers.

I admit I'm somewhat of a photographic philistine, having done most of my serious work in photojournalism rather than the various high-quality advertising and art modes, but I just got my D3 and FREAKIN' 25600. You can do whatever pixel-peeping you want, but I promise you, this is gonna be a revolution in pj work -- I took the camera out of the box, stuck in a battery from the D2x and shot my cat walking across the room in the light of a single lamp at ISO 6400 -- and I COULD HAVE FREAKIN' PUT IT IN A NEWSPAPER if it hadn't been a picture of my cat.

I can tell you that an awful lot of PJ work is done in low-light conditions -- late in the afternoon outside of jails, into the interior of cars, press conferences in poorly-lit rooms, or people walking along government-building hallways, or sitting in their living rooms, or night-time crime scenes, or shots taken with long (slow) lenses. Flash works for some of it,, but also makes things look really unnatural, unless you have hours to set up, and there's not much you can do about the long-slow problem. But at 6400 and up you can shoot natural-light shots at fast enough speeds that you don't get blur when people move their lips or turn their heads...And though the photos won't work for art purposes, they'll sure as hell work in a newspaper. I am seriously impressed by this camera, in a way I wasn't with earlier cameras. There was always an argument with earlier cameras about whether they were really as good as film, or just more convenient -- in this case, I haven't seen film do what this camera can. This camera notably pushes beyond film capabilities.

Maybe Nikon shooters just haven't been as privileged as Canon shooters when it comes to high ISO, but I was reading Bjorn Rorslett's site when he got his D3, and he pretty much portrayed it as a jaw-dropper. That's about all I can say: I'm not a pixel-peeper, but I freaked out the first few hours I was shooting this thing. It's like nothing I've ever encountered, and I spent last year working with an M8 and a Noctilux....

Not be be over-excited, or anything. 8-)

JC

I agree with that statement to some degree.

The challenge is there are many times you don't know what you want, how big you want it, or how much you might need to crop it until you see the images. That has been true long before digital came into the picture.

Where we're going doesn't seem to change all that much through the years, but how we get there sure keeps mutating!

John,
I felt pretty much the same way when I got my 5D. I found I could take photos in the street at night without the disturbing flash which brings attention to oneself. I thought it would be ideal for photojournalism.

I wish Michael had also compared the D3 with the 5D which has the same pixel count and sensor size. We could have then found out if the lower noise of the D3 at say ISO 6400 was at the sacrifice of some detail, because of course any camera can have an ISO of 12,800 and 25,600 by simply underexposing the appropriate number of stops at the highest ISO setting.

With this idea in mind, I wondered how my 5D would perform at various stops of underexposure at ISO 3200 (which is already ISO 1600 underexposed one stop).

So I set up my 5D on tripod and directed it to a poorly lit corner of the room. Experimented with a few shots to get a full ETTR, which was 2 seconds at f11 and ISO 3200, then proceeded to reduce the exposure with each subsequent shot, eventually arriving at an ISO equivalent of a quarter of a million (ISO 256,000).

Instead of the figure of ISO 3200, I have used the correct figure of ISO 4000 (according to dpreview) to arrive at an ISO equivalent of a quarter of a million with 6 stops underexposure, ie. from the ETTR of 2 secs exposure; 1 sec, 1/2 sec, 1/4 sec, 1/8th sec, 1/15th sec, 1/30th sec.

The shot which was underexposed by 6 stops was hoplessly noisy but I wondered if the 'mean' stacking mode in CS3E would make it usable. I think it has.

So here it is, with a 'before and after' stacking comparison. Perhaps the first ISO quarter of a million shot ever displayed on LL .

Click to view attachment Click to view attachment

I hope it is not simple exposure. Checkout the measures on the page I posted:

http://theory.uchicago.edu/~ejm/pix/20d/po...D300_40D_tests/

the 5D's quantum efficiency (4.1) is way behind the D3's (7.7). 5D collects 0.061 electrons per unit area while D3 captrues 0.108. These are not surprising since the 5D is two year old technology and the D3 use the new double micro lens design.

Sorry for getting into numbers again. But I am trained as a scientist as well. From looking samples I have a feeling that the 1DsIII's QE is on the par of the D3's, but there has been no measures yet.

From the samples I saw (checkout the Imaging-Resources, for example) the D3's exposure parameters are consistent with the ISO ratings, and to my eyes the D3 RAW images contains more details than my 5Ds when examed at not 100% but 170% to 200% at high ISOs, and properly (and painfully) sharpen. As Lloyd going to point out D3's initial soft images actually makes them exceptionally well to be enlarged, just like my very soft Imacon 949 scans.

I suppose the fact that the D3's base ISO is 200 (rather than the 100 of the Canons) is explained by it's greater quantum efficiency. I'm not doubting that the D3 has lower noise than the 5D at high ISO's, it's just not clear by how much and I simply don't know enough to make assumptions that what might appear to be as much as, say a 1 stop improvement at ISO 1600, is carried forward to ISO 25,800 and beyond. Although it would be reasonable to assume that it would be the case.

As I understand it, ISO 3200 on the 5D (actually ISO 4000 and this should be taken into account when making comparisons) is really ISO 1600 underexposed 1 stop (ie. no pre A/D analog boost) and likewise, ISO 12,800 and 25,600 on the D3 is really ISO 6,400 underexposed 1 stop and 2 stops respectively.

If this is the case, it seems a legitimate comparison to me to compare the 5D at ISO 4000, underexposed by 2/3rds of a stop to make it effectively ISO 6,400, with the D3 at ISO 6400, just to see the difference .

Also, I see no reason why one should not extend this principle and compare further underexposure in both cameras. Perhaps at ISO 256,000 the D3 could produce a result with a single image as good as that 5D stacked image I posted above . (7 images of the same exposure stacked in mean mode and auto-aligned).

Thanks Ray! With the stacked images in hand you should download the trial version and try this "super resolutin" (and noise reduction) stuff out:

http://www.outbackphoto.com/CONTENT_2007_0...cute/index.html

I tried it out with a set of my infrared converted 5D shots and blow is a crop of the results at 200% (you know my habit ). Do I need to say the smoother looking one (right) is after the "super resolution" treatment?

Canon 5D-IR, 24-105mm/f4 L lens at 105mm, upsampled to 200% in PhotoAcute from 4 identical exposures on the tripod ("2x", the PhotoAcute "super-resolution" default) to DNG RAW file, with "Oversharpenng protection" at "Normal" and no other options (such as color defringing, geometry correction, or dynamic expansion), then converted to TIFF in Raw Developer (ACR or Light Room much worse) with R-L deconvolution sharpening 1.2/50, opened in Photoshop with little "Film Noise" added to the 5,836x8,768 pixel (almost 20x30 at 300dpi) file, JPEG saved from cropping the ~150MB image at 100%.

Comparison image: one of the four source RAW files, processed in Raw Developer with the same parameters and output to the same pixel dimensions, then went through the same operations in Photoshop.

To me, the "super-resolution" does not really matter that much, but the much less artifacts at the 200% level is impressive, as well as the much more natural looking and the improved dynamic range. Remember we are looking images through the demosic process, and at 200% level. Much better than Q-Image or Genuine Fractals, to my eyes.

Here is the Wiki entry of the process with an impossible-looking example:

http://en.wikipedia.org/wiki/Super-resolution

Cheers,
Leping

Yes. This is pretty much what I'm saying. The greater rez of the original 1Ds over the D60 would tend to make the noise less apparent at the same image size.

Of course, being able to shoot at 12,800, or 25,600, would tend to result in less noisy pictures than a higher rez camera that can't shoot at those speeds at all.

We seem to be talking about similar technology here but perhaps used in a different way. I posted a comparison of shots of my hotel room in Krabi on another thread with a tongue-in-cheek title about reducing D3 low-noise envy.

The problem is essentially this. You want to photograph a poorly lit scene with a fairly extensive depth of field where flash will not do because distant object will be severely underexposed, and/or, near objects will be severly overexposed.

Fortunately, the scene is pretty static so no problem. You use a tripod, a slow shutter speed and f11 for good DoF.

But what happens if you are not carrying a tripod. It's my experience that most people do not carry a tripod most of the time. I don't, although I have one in my room. With a camera slung around my neck and a couple of lenses in my photographers' vest, I find it very inconvenient and cumbersome to also carry a tripod.

It seems that stacking in 'mean mode' with CS3E has now provided another option in such situations. If, for a hand-held shot of the scene at the sort of aperture necessary for good DoF, it becomes necessary to use ISO 256,000 (ISO 3200 underexposed 6 stops), then a burst of 6 or 7 shots with the 5D at the same shutter speed could provide a usable image, provided the dynamic range of the scene was not great.

Check out my thread at http://luminous-landscape.com/forum/index....showtopic=21343

One would certainly expect this to be the case, but often things are not what they seem to be. For example, the D3's ISO 12,800 and 25,600 settings are apparently bogus. The purpose of having them is (presumably) to make assessing correct exposure in the histogram and the camera's LCD much easier. The jpeg image has been adjusted, in-camera, despite real and actual underexposure. Without such settings, one would be squinting at apparently underexposed shots in the camera's LCD and trying to judge if the shot was really too much underexposed or not.

I think also the purpose of these ultra high ISO settings relates to ISO bracketing. I'm not sure about the features on Nikon cameras, but I believe they've had ISO bracketing for a while. Of course you can't bracket an ISO which doesn't exist.

For example, if I take a shot with my 5D at ISO 3200 and the histogram and highlight flashing tell me I've overexposed, I've ruined the shot. If I need a particular combination of aperture (for DoF) and shutter speed (to freeze the action or camera shake), then autobracketing of exposure is not a good option.

Without ISO bracketing there can be a disadvantage in using an 'ersatz' ISO setting.

Since Nikon is now offering these 2 ultra-high settings which are really underexposures, it's quite likely they have built in a level of noise reduction to make them more acceptable. Once detail has been lost through in-camera noise reduction, it cannot be regained.

However, there is now a way of reducing noise, not only without destroying some degree of resolution, but with actual enhancement of resolution.

I'm suggesting here that using such stacking processes with ultra-high ISO shots, the D3 might not even equal the 5D. It's something which I think should be tested.

Ray,

All manufacture's streched ISO's are, just you mentioned, streched from the real ISOs, and the D3's is no difference. What I don't understand is why do you think what your 5D can do (handhold multiple exposure stacking) the D3 would not.

If you exam available example carefully you see D3's high ISO NRs kept most of fine details, when processed in Capture NX.

Ray,

Pls. explain your use and the justification for using therms ' apparently bogus' and 'underexposures' when refering to the D3's implementation of the high sensitivity settings (High1 and High2 not calibrated 12800 and 25600).

If what you're saying is that the camera at these settings just digitally amplifies the signal rather than amplifying in in the analogue domain, you might be right or partly right, but your use of the term 'underexposure' is not appropriate I believe.

Furthermore, this issue is not new with these cameras as all cameras that I know off behave similarly in their respective uncalibrated high sensitivity settings.

I believe it is very appropriate in the sense that there is usually a qualitative improvement when using a higher ISO. The image has less noise than it would if one used the same exposure at a lower ISO. One is to some extent taking a risk in using a higher ISO. If one miscalculates and overexposes, the shot is ruined. However, if one gets it right, a good ETTR, one gets a better image.

But when using these false ISO's, the risk of overexposure is the same but the benefit regarding image quality and less shadow noise is zero. A correct exposure to the right, with the 5D at ISO 3200, produces the same result as using ISO 1600 at the same shutter speed and aperture. The only difference is in the appearance of the histogram and image on the camera's LCD.

Using ISO 1600 with the confidence one can underexpose a bit and preserve highlight detail without compromising image quality compared to using ISO 3200 is a real advantage. I'd describe it as being better that the latest feature of 'highlight tone priority' in Canon cameras, which is causing some confusion apparently



That's true. It's not new. But Nikon have taken it a step further and given us 2 false ISO settings, and yours truly has taken it a few steps even further and produced an ISO 256,000 image with his 5D.

Hi Ray,

I'm not sure what you mean by "causing some confusion", because the description in the manual and the outcomes are quite consistent. To quote from the 1DsIII manual:

"Improves the highklight detail. The dynamic range is expanded from the standard 18% grey to bright highlights. The gradation between the greys and the highlights becomes smoother."

Then immediately under that cryptic description (yes, Canon still struggles with how to write a good manual), they do warn us: "With setting 1, noise in the shadow areas may be slightly more than usual."

My limited experience with it before I turned it off indicates that the manual is correct: it generates noise. Another Torontonian whose work I've seen a couple of times and respect has a 1DIII and he told me he turned it off permanently, so my experience has been replicated as the manual leads one to expect.

Perhaps there's some confusion about exactly how the "feature" operates, but not much confusion about its impact. People with more technical knowledge about how sensors work than I possess can probably reverse engineer an explanation in their own minds, because we certainly don't get it from Canon. That said, it seems to me there is only so much light hitting those receptors, hence perhaps the firmware is somehow allocating more of the available light to the part of the tonal range between 18% grey and white in order to improve tonal gradation in that part of the range, leaving less for the 3/4 tones and below, which are inherently relatively "starved" for light anyhow and therefore produce less S/N. Does this make any sense? Or am I just displaying the confusion you mention?

Mark

As far as I understand the highlight mode just re-biases lattitude trading a stop or two below key, for a stop or two above key. To make the resulting image look "right" a tonal curve is used to brighten the image without crushing the highlights. This is something you can "do by hand" quite easily, but Canon have automated it for you.

Graeme

Graeme, I'd appreciate if you could restate the first sentence using what Ben Willmore calls "a techno-babble decoder" so I can understand what you are saying . If you would kindly explain the meaning of the terms "re-biases lattitude" and "key" that would do it for me. Despite my difficulty with the first sentence I agree with the second sentence, because it occured to me as well that whatever Canon is doing in HTP one can most likely do by hand in Camera Raw. Then the final missing element is how all this may generate more noise.

Possibly true at some level, but "bogus" comes in different shades. There isn't necessarily a visible difference in IQ between ISO 1600 pushed to 25,600 and actually having amplification 16x as strong as ISO 1600, if the RAW data clips below black, or exactly at black. The only problem is if the data is clipped above black, in which case whatever is clipped is 4 stops bigger in ISO 1600 pushed to 25,600 than what "true" 25,600 might give if it clipped at the same ADU level.

As far as I can tell, I am the first person in the world of usenet and web photo forums to note that ISO 3200 and sometimes 1600 were achieved in a different manner than the lower ISOs on my 10D (and many other cameras too), and popularized the term "fake ISO" (some of the astrophotography folks may have noticed, too, but kept it to their astroworld). That terminology was based on the context of the 10D, where, clearly, read noise in electrons dropped quickly from ISO 100 to 200, less so from 200 to 400, but still a good amount from 400 to 800. The "real" ISOs, then were 100 to 800, which were achieved apparently through different amplifications *at* the photosites. Upon examining other cameras, however, like the Nikons using Sony CCD chips, it became apparent that another level of fakeness existed, as all ISOs on these cameras have the same amplification at the initial read, and the differences of ISO were achieved only through amplifying the signal yet again before hitting the ADC. There is no opportunity, however, for this second analog amplification to improve on the photosite read noise in electrons, as that read already occurs earlier in the chain, so the amplification basically only puts the signal at an advantage going into the ADC, with less resulting ADC noise in electrons at higher ISOs. In the realm of high ISOs, ADC noise is so small, relatively, that there need be no major difference between amplifying and multiplying. The differences are the kind you need to be stacking a dozen or more images to take advantage of.

You're then asking Camera Raw to pull down 1 stop more highlights into the image, which it may or may not handle gracefully.

Noise breaks down into two basic sources; shot noise, which comes from the actual exposure of the sensor (and has nothing to do with ISOs or modes, except as how they affect metering, and therefore, exposure), and read noise, which is a fixed blanket of electronic noise generated by the camera while reading the sensor and converting its signal to numbers for the RAW data. The read noise is the only one that varies directly with ISO, and the HTP setting. The read noise is usually about the same for ISO 100 and 200, being slightly higher for ISO 200 with modern Canon DSLRs. This means that relative to your signal, using ISO 200 and HTP (which uses -1 EC and ISO 100 internally) has almost 2x the read noise of straight ISO 200, for the same absolute exposure. The significance of read noise depends on what you are shooting, and how it is exposed. Read noise is significantly visible only in the shadow areas, so if you shoot in controlled lighting, without deep shadows, then you won't have any tones swamped with read noise, and the loss is not significant. If you are going to pull up highlights or have areas of slow gradients in near-blacks, however, the extra read noise from HTP can be significant.

Mark,
I'm relying on people like John Sheehy for the explanations of these matters. But I have verified for myself, as probably you have also, that an underexposure at a low ISO produces a noisier image than the same exposure at a higher ISO. The effect was fairly marginal with early Canon DSLRs like the D60, so an ISO 100 shot, 2 stops underexposed, would have hardly more noise than the same exposure at ISO 400 (which in relation to the ISO 400 setting would be a correct exposure).

It's this principle at work which I believe is responsible for the noiser shadows with HTP enabled. The reverse of HTP would be to have a suitable shutter speed and aperture at say ISO 100, take an experimental shot and discover it was 1 stop underexposed, then take the shot again at the same aperture and shutter speed (on manual) but at ISO 200. The second shot would then be correctly exposed, along the lines of ETTR, and shadow noise would be less.

This probably sounds a bit complicated and convoluted, but it makes sense to me .

Just a word here for M. Reichmann with regard to Jay Maisel's close-up enlargements.
You say: "Possibly there's a bit more shadow detail in the Canon frame though (look at the iris). "
I think it's not the iris but just the reflection in his glasses which is positioned very slightly more to the left in the Canon shot thereby "drawing" what might seem to be detail from the eye rather than off the glasses themselves.
In any case, both wonderful machines and as always a refreshing article, thank you.

John, thanks this is getting clearer as it continues. Where I noticed most of the noise on December 4th when I had HTP enabled was indeed in the dark areas of tall charcoal-coloured buildings in the TD Centre of Toronto. When I turned HTP off and returned there on December 5th to re-photograph those same buildings (intrusiveness from *security* notwithstanding) I obtained less noisy results. The interesting thing, though, is that these noise differences are much more apparent on a high resolution display (LaCie 321 at 1600*1200) than they are in a Premium Luster 13*19 inch print from an Epson 3800, where it hardly shows. Once this discussion moves from display to print (at least to 13*19 inches) I'm discovering that the issue's existence may be less significant than its impact.

One thing I would like to see discussed when judging the Nikon D3 high ISO quality or compare it to other (Canon) cameras is not only concetrating on noise but discuss the colour integrity of the results.

Colour integrity does not involve only saturation but fidelity and apparent 'depth' as well, things that cannot be easily restored through postprocessing.

From the examples I have seen it looks like the D3 performs much better than other cameras in this respect, managing to produce acceptable colour integrity at sensitivities where other cameras 'break down' so to speak.