The SNR 18% graph for “screen” suggests that “much better” is misleading at best.

3 dB yields about a stop of difference. The D800 on screen is behind the 5D3 by 1.5 dB or less at any given set ISO. That means it’s behind by half a stop or less. And that’s on screen, when viewing at 100%. Since both manufacturers differ in their real ISO versus set ISO values, you have to account for that too. That adds another 0.2 stops at most (based on the measurement at the ISO 25600 settings for each), and typically less than that.

But the resolution difference yields a 0.7 stop advantage for the D800 in print.

The real difference in print, in stops, is as follows (positive numbers indicate an advantage for the 5D3, negative indicate an advantage for the D800. ISO refers to the ISO settings on the cameras):


ISO 100: -0.15
ISO 200: -0.13
ISO 400: -0.15
ISO 800: -0.16
ISO 1600: -0.18
ISO 3200: -0.21
ISO 6400: -0.03
ISO 12800: -0.35
ISO 25600: 0.01


The above was computed by accounting for the measured ISO difference (the 5D3 always shows a higher measured ISO than the D800, so that is given as an advantage to the 5D3), in stops, and adding that to the SNR dB difference converted to stops (3 dB = 1 stop), then subtracting 0.7 stops for the resolution difference. This methodology yields results that will be valid even for 22 megapixel prints, since the resolution difference converted to stops implies downsizing the higher resolution to match the lower resolution (hence, downsizing 36 megapixels to 22 megapixels).

End result: the D800 has the advantage in print, even for the largest prints the 5D3 can muster, for almost all of the range, and the 5D3 only becomes equal at ISO 25600. But because of its higher resolution sensor, the D800 retains more detail.

The biggest problem with the D800 isn’t its image quality, it’s its burst rate. That’s the downside of ultra-high resolutions.

No, sorry, you guys are just wrong about this. The higher megapixel D800 gets a greater dynamic range (14.4 Ev versus 11.7 Ev), better high ISO performance, better color depth (25.3 bits versus 24 bits), and, of course, better detail retention (as long as you use a lens that’s up to the task — this is especially true of the D800E, which lacks the antialiasing filter).

See http://www.dxomark.com/index.php/Cameras/Compare-Camera-Sensors/Compare-cameras-side-by-side/(appareil1)/795%7C0/(brand)/Canon/(appareil2)/792%7C0/(brand2)/Nikon

Now, you may see more noise in the D800′s output if you do the 100% thing that’s been discussed to death here, but the D800 will, based on the above, outperform the 5D3 where it counts: in the final output.

I haven’t heard of a head to head comparison between the two, so I’m looking forward to that. Maybe that will show something drastically different than the numbers above indicate. But I wouldn’t count on it.

And note that none of this is to say that the 5D3 is a bad camera. It looks like a wonderful camera! I’d buy one in a heartbeat if it weren’t so expensive. There’s more to a camera than just final image quality. But the discussion here is about noise and image quality as it relates to pixel pitch, and it appears, thus far, that the D800 handily disproves the notion that smaller pixels get you more noise in most real-world situations.

Look, Crowded sensor = high noise, just plain science. Technology has not been able to reverse that. But we frequently hear naked and shameless lies from producers.

However, 5D Mark III may be a good one. It is learned to have produced noise free images from ISO 100 to 1600. If that is true, even sports shooters may also go for it as all the other fast machines produce high noise (7D and 1D Mk IV).

Not much news available about performance of 1Dx however, it may not be action/wildlife shooters’ choice due to several factors. So, they are left high and dry.

Frankly, if it doesn’t have strange image artifacts (like banding in the deep shadows), then it’ll definitely be the camera for me as well, since I’d like to have the extra depth of field control that full frame brings to the table and I am unwilling to compromise on autofocus or responsiveness to get it. Even if the D800 downsized images wind up being every bit as good as the 5D3′s, I simply don’t have any real use for the kind of resolution the D800 provides.

Things just got really interesting as regards this particular debate.

And if the final image is what’s important, then is per-pixel performance the only thing that matters? The fact that the final image is going to be the same resolution independent of the camera suggests that per-pixel noise is but one part of the overall performance equation.

Hmm…so as your camera gets more capable, your final images get larger? Your target print sizes get larger? How exactly are you using your images?

If your final output size is almost always at the camera’s native resolution, then I can see how you might be disappointed in a camera that has worse per-pixel noise than a previous generation camera of lesser resolution, because it means you won’t be able to use the camera’s native resolution in quite the way you were anticipating.

But I must say I’m at a loss (perhaps because I lack imagination ) to see how exactly you’re using those images. How are you publishing them? If you’re publishing them for a client, how is it that the client’s expectations match your own as regards pixel-level performance and image size? Is the client demanding ever-larger images from you?

Most photographers that I know of either target a specific resolution on the web (which isn’t necessarily 800×600, of course — it could be quite a bit larger — but it’s generally a specific target resolution nonetheless) or print to various predefined sizes. About the only class of photographer I can think of that prints as large as the camera makes possible is landscape photographers, and they seem to be much more concerned about sheer resolution than noise, as long as the noise doesn’t interfere with the detail.

You should also consider what happens when you crop into the higher resolution image. For example, let’s say the high res image was 21MP but you crop into it for composition and then end up with 10MP. How would that compare to another image from an 18MP camera which is not cropped? That’s why I want as little noise as possible at the pixel level.

Oh, I agree, I want as little noise as possible at the pixel level, but that’s only if all else is equal. The problem with different sensor resolutions is that all else is not equal.

It’s fair to say that the cropping potential of a higher resolution sensor is no worse than that of a lower resolution sensor when both sensors are of the same technology level. You may be revealing more per-pixel noise by cropping, but you’ll also be revealing more detail. If the target resolution is capable of showing it, then it will show more detail and stronger, but finer grained, noise, all at the same time. If the target resolution is not capable of showing that larger amount of detail, then the noise signature of the image will also be reduced accordingly.

Everything depends on the characteristics of the final image. And as you agreed, the final image is what really matters.

In your example, since you’re cropping from one camera and not cropping from the other, then how is it an apples to apples comparison between the two cameras? You’re magnifying a smaller area of one sensor and comparing that to the full sensor area of the other. Yes, it would be certainly be nice if the smaller area of one fared as well as the full area of the other, but how does any failure on the part of the smaller (cropped) area tell you anything of real value?

I don’t agree with downsizing to compare noise. Noise should be checked at the pixel level. If the higher resolution image contains more noise than the lower resolution image at the same ISO, then it is useless as far as I’m concerned. If the noise levels are the same, then the higher res sensor is an improvement.

I agree that it’s useful to check the noise at 100% to get a sense of what the camera is capable at its maximum resolution. But what we’re talking about here is comparing two cameras against each other.

You’ve already agreed that the final image is what really matters. So let me ask you this: how does the noise you see at 100% translate to what you’ll see in the final output? More precisely, if you see more noise at 100% with a larger resolution camera than a smaller resolution camera, is it your belief that the larger resolution camera will produce a noisier final image at the final image’s resolution?

Note that I’m not talking about artifacts such as vertical banding. I’m talking strictly about random noise which disappears as an image is scaled down.

Why would it be inappropriate to compare both cameras in terms of noise by having each generate a single final image that has the same resolution for both, since that is exactly how you will be using the cameras in the real world?

That is why downsizing the larger image to match that of the smaller image is appropriate for evaluating noise. Actually, it would be best to size both to match the final image’s resolution, but downsizing to the smaller of the two will at least tell you how each will fare at the point at which the smaller resolution camera generates maximum detail.

Either way, if you downsize the larger image to match the resolution of the smaller image, you can upsize it again and be guaranteed that the amount of detail it will be carrying will be no less than the detail in the smaller image.

Just food for thought…

Agree with that, I hope that’s what we will get. But this is Canon we’re talking about. If they really put the 1D AF in the 5D I’ll be blown away.

I think we think more alike than one might suspect. LOL.

Because what you said here is exactly what I’m thinking. I fully expect Canon to blow it, because I don’t think they’re really in tune with their customer base at all.

I mean, as an example, they just announced the 24-70 f/2.8L II. It does not have IS, and they want $2300 for it! Meanwhile, they released 24mm and 28mm f/2.8 primes with IS. So they release two slow primes (f/2.8? Seriously? Who cares about that when there are plenty of zooms that’ll do the same?) with image stabilization, where it’s not very useful, and neglected it on the lens that everyone wants to see it on.

What morons.

I like resolution too, but I don’t like more resolution plus more noise. If they can keep the noise levels constant at the pixel level, while increasing resolution, then I’m ok with that.

Oh, I’m more than okay with that. If they’re able to achieve that, then I’ll be ecstatic. But what I’m saying is that as long as the noise level when downsized to the previous generation’s resolution is no worse, then I’ll gladly take the additional resolution, assuming there are no other tradeoffs (like burst speed). That’s an improvement on what was there before, because you can always downsize the image and get the same noise signature as before.

Suppose Canon created a 70D with 20 megapixels, and got the per-pixel noise down to 40D levels. You know what that would mean, right? It would mean that noise performance in the frame would be one stop better in the 70D than in the 40D, and that would remain true for any ISO for which that per-pixel noise relationship remained true.

It’s also a bit troubling in that we haven’t seen any full resolution shots at high ISO that I know of, so we can’t really evaluate the high ISO capability.

The 5D3 AF had better be more than just the 7D AF stuffed into it. One of the biggest problems with the 5D2 AF is the layout. The rule of thirds locations aren’t covered at all, because Canon basically took the AF sensor from the 30D and stuffed it into the 5D and 5D2. If they do the same thing this time around, we’ll once again not have rule of thirds coverage. We’ll basically have a really good, moderately sized center point. I will be entirely underwhelmed in that case.

No, we need the 1Ds series autofocus system in the 5D3, or the 7D autofocus expanded to cover the same portion of the 5D3 frame that it covers in the 7D frame, for me to consider the autofocus to be anything more than another joke.

I’ll be perfectly happy if they put the 18 or 21 MP sensor in the 5D3! Not because of the improved per-pixel high ISO performance you’re likely to get, but because that’s plenty of resolution for my purposes and the burst rate can be high (7 FPS should be easily achievable). I won’t complain if they put a 36 megapixel sensor (or whatever) into it as long as the noise signature at 18 megapixels is similar to that of the 1DX.

Not to me. I want the image to be as clean as possible.

I want the image to be as clean as possible, too, but if I can retain the same amount of noise in the frame while getting more resolution, I’ll go for the higher resolution and downsize when I need to. I’d rather have more flexibility than less when it doesn’t actually cost me anything in the final image.

But let’s say you crop out a 4MP part of the 18MP image and then resize it to 640×480, then all of a sudden you see more noise.

Of course.

Now let’s say you crop out a 2.2 MP part of a 10MP image (the same portion of the frame as you cropped from the 18MP image) and then resize it to 640×480. You might think you’d see less noise, but you won’t. The reason you won’t is that you’re downsizing that 2.2 MP part less than you are the 4 MP part. And you know from your experience with making web images that downsizing an image reduces its noise signature.

Now, if you crop 4 MP of that 10 MP image and downsize it, then of course you’ll see less noise than you would the 4 MP part of the 18 MP image, but you’ll also be seeing a larger portion of that original 10 MP image.

When you take a photo, do you frame it so that your subject occupies a certain portion of the frame, or do you frame it so that your subject will get a certain number of pixels in the resulting image? I’m betting heavily that you do the former. In fact, you probably try to get as much of the frame filled with your subject as you can, as long as you get the parts of the background you want as well, no matter what the resolution of your camera is. Which is to say, you probably try to take your shots in such a way that you minimize the amount of cropping that you do.

If you shoot based on the frame, why wouldn’t you evaluate the resulting final image as a whole, at whatever target size you intend to use it at?

The marketing hype might say that you can crop more heavily with a higher resolution image, and that still has some truth to it in that even if the noise is higher, the detail captured is also higher (up to a point, but if you’re talking ISO 100 then it’s certainly the case). But I don’t listen to or care about marketing hype. I care about what the camera can really do. And if I can crop more heavily while retaining a given level of detail, that’s useful even if the resulting crop has more noise at the pixel level. Which is to say, I haven’t lost anything as a result, and I’ve gained a potentially useful advantage. After all, I can always clean up the noise, and may even be able to clean it up in such a way as to retain most of the detail (more than would be had by a similarly proportioned crop from a lower resolution image).

By the way, take a D800 image, apply a gaussian blur pass with a radius of 0.7, and then downsize the result to half the resolution (0.7x on a side). That will give you an idea of what the noise signature from the D800 would be if it were an 18 megapixel camera.

Yeah and even with the gapless microlenses, the 50D still majorly sucked with regards to noise.

But I’m not talking about random noise, I’m talking about artifacts, like pattern noise. In other words, I’m talking about banding, mazing, etc. — artifacts that will tend to show up independently of the size of the image. Those appear to be independent of pixel pitch.

The 50D was pretty bad at higher ISOs in terms of horizontal banding but managed to remain free of that up to ISO 3200 (I know this because I have a 50D as well as a 7D).

The 40D also has pattern noise but you have to really boost the shadows to see it. Perhaps they all do have it, more or less.

If you’re talking about a visible crosshatch pattern, yeah, I think they all have that, but either the 5D2 has it a lot worse, or it exhibits an additional mazing pattern that shows up a couple of stops above the floor. See http://photography-on-the.net/forum/showthread.php?t=750731.

Basically, the difference is that with the 5D2 and 7D, you’ll see pattern noise even at levels where real shadow detail is still visible. With the other cameras (e.g., the 40D and 50D), you may see the mazing pattern, but you won’t see it until you’ve gone beyond the point where any shadow detail is left.

But the problem with the 5D2 is that you can see the pattern noise and banding even if you don’t brighten the shadows too much.

Yeah, exactly. That kind of noise at those levels is simply inexcusable in a camera that’s supposed to be all about image quality (it’s not like the 5D2 has any other redeeming qualities. The autofocus is a joke compared to its contemporaries, the camera is slow, the shutter lag and viewfinder blackout are both high, and the viewfinder doesn’t even give you 100% coverage). I find it deeply annoying that the 7D exhibits something similar (vertical banding) in the deep shadows as well.

Lightroom 3 seems to do a very good job at minimizing the effects of that vertical banding in the 7D (so that when it’s there, you can barely see it when you push your shadows, and you can’t see it at all otherwise), but such things shouldn’t be there at all. The 60D, T2i, and T3i don’t exhibit this problem.

Like on the D800. That’s just terrible.

Aside from the color “banding” (which seems to be a problem with color gradients or something, and may indicate a tone mapping issue and is not really pattern noise as I normally think of it), have you seen other artifacts in D800 images? If so, can you point at some examples?

I haven’t seen pattern noise on the 1DX sensor though. Try it on their samples. Even a +5EV boost doesn’t show any

5EV seems to be a bit much to attempt to boost a JPEG image (you don’t have access to any 1DX RAW images, do you? If you do, please share!). I’m not sure that the lack of pattern noise when boosting a JPEG image by that much really means a whole lot, especially if the blacks were crushed during postprocessing.

I very much hope the 1DX doesn’t exhibit any sort of pattern noise, even when the exposure is boosted by that much. For the kind of money they want for that camera, the images had better be flawless.

I can’t wait until we can get our hands on RAW images from the 1DX, D800, and D4 so we can see just what sort of images they really produce. I have high hopes for the 1DX, but will not buy one simply because it’s priced way beyond reason for me.

I’m more interested in the 5D2′s successor, actually. I certainly won’t buy it if it exhibits pattern noise or other strange artifacts (like the 5D2′s black fringe problem that showed up around bright lights and such).

If the D800 actually has pattern noise issues, then that is a major problem.

But what about the more normal random noise? Isn’t what matters the final image, at the size that’s going to be used? If not, why not?

The tech in the 5D2 sensor is very old. It doesn’t even have gapless microlenses which the 7D and 60D have. I’m sure the electronics have also been improved in the later sensors. So it’s no surprise that they perform better than the 5D2. If the 5D2 sensor were developed with the same 7D/60D tech, it would outperform them. You have to keep this in mind when you compare sensors. So although the 5D2 has larger pixels than the 40D, the 40D also has gapless microlenses which means it can capture more light.

Gapless microlenses were introduced on the 50D, not the 40D (see http://www.dpreview.com/reviews/canoneos50d/page3.asp).

The 20D, which has pixels the same size as the 5D2 and which certainly doesn’t have gapless microlenses, also doesn’t exhibit pattern noise in the deep shadows. That seems to be something that is “unique” to the 5D2 and 7D. I know of no other Canon cameras which have that problem.

So artifacts like that don’t seem to be related to pixel pitch, but to something else (perhaps the electronics used to read the data from the sensor, or errors in the way the data is saved, or something).

When I menion more/better dynamic range I mean USABLE dynamic range.

Oh, well, in that case, why didn’t you say so in the first place? We could have saved ourselves a bit of sparring (not that it’s not enjoyable, mind you).

“Usable” dynamic range is a subjective term (and is acknowledged as such in the source I cited earlier). It depends on your perceptions and the uses to which you’re going to put the images, as well as your measurement preferences. Since you prefer to measure everything at the pixel level, then yes, you will generally see lower dynamic range out of smaller pixel sensors than out of larger pixel ones. That might not always be the case, but it’s likely to be in the general case.

If you always print the same size, you may find that the dynamic range of one camera with small pixels manages to exceed the dynamic range of another with large pixels, for that particular purpose.

I should note that banding and other such artifacts are not a function of the pixel size. For instance, the 5D2 has larger pixels than the 40D and is newer technology, and yet exhibits pattern noise in the deep shadows at low ISOs while the 40D does not. That makes the 5D2′s “usable” dynamic range less than what it otherwise would be, and quite possibly less than the 40D’s. Similarly, the 5D2 exhibits horizontal banding at very high ISO while the 7D, despite its much smaller pixels, does not.

I, too, prefer more usable dynamic range, so I’m quite critical of cameras that introduce artificial artifacts into the image that reduce its usable dynamic range. The 5D2 and 7D are both guilty of this (the 7D exhibits vertical banding in the deep shadows at low ISOs), but the 60D (with the same pixel pitch as the 7D) is not. So it may be that the 60D has more in the way of usable dynamic range than the 5D2, despite its much smaller pixels.