High ISO Compared: Sony A7S v. A7R v. Canon EOS 5D III

It's not that there is MORE highlight detail, apart from that tower, which is a methodology issue. As evidenced by most of the windows and dark areas, I said there was no less shadow or highlight detail. Which you would expect with the differences in dynamic range so many bellyache about. The first function of cameras is to shoot photographs. Not test charts.

I am a beginner here but this is what I noticed for the comparison b/w canon 5dm3 and A7S. The yellow colors are somewhere the red colors are more richer in case of Canon 5dm3. Also, some of the places A7S seems overexposed? where as 5dm3 managed it quite well.
please see here
http://www.dpreview.com/reviews/image-comparison/fullscreen?attr65_0=sony_a7s&attr65_1=sony_a7r&attr66_0=25600&attr66_1=25600&attr67_0=raw&attr67_1=raw&attr68=12mp&normalization=full&widget=119&x=0.7929017009570305&y=0.043870286340165855

ijustloves:

Right, neither is the A7R nor the 5DIII is a serious high ISO body.

They're decent, but it's like claiming the Nikon D800 is a serious high ISO body. It's not.

So plenty serious.

Besides the obvious Sony A7S, there is the Canon 6D. Then there are the elephants in the room--all much bigger deals at high ISOs than the Canon 5DIII or the A7R.

I suggest you get raws (with deep shadows) from all and extract those raws.

HFLM:

The A7r does not do particularly good detail at ISO 25600.

That point is real easy to verify.

I've seen comparisons with the D4s RAW elsewhere. The Sony is noticeably better.

Better how?

With the Sony A7s there's less grain, and details blurry in the Canon's raw are clearer in the Sony's raws.

The 5DIII has bit more cyan and magenta banding at that ISO than the Sony.

Why don't you try the above test?

Anyhow the Canon 6D is also a better high ISO body than the 5DIII. And the 5DIII is considered by no one to be Canon's best high ISO camera--that's the 1DX. While the Sony A7s is likely Sony's best high ISO camera.

Preternatural Stuff:

The A99 is a fine camera for many of the reasons you cite, but it would be a lot more interesting if version 2 had this 12MP sensor.

I think he means what he says. 6400 and lower. 12MP @ISO 3200 from A7s vs A7R + NR and then downscalled to 12MP. I also think A7R will be better because it can easily sacrifice some details in NR process which would be lost anyway from downscalling.

Duly noted Rishi!

By lower ISOs I mean the normal range 1600 to 6400 that the noise level is noticeable. I think when you do NR on a higher res image you can do better in terms of the granularity of noise detection that you can't do with a lower resolution. I can sure put it to a test!

I shoot night aerials a few times each year and 6400 iso good for that, I have gone to 12800 just to see what it's like. if I can shoot at night from a helicopter at 6400 iso what is that needs 100,000iso and more?
Never mind the noise where did Sony hide the colour, highlight detail and tonal gradation?

KAllen:

Try shooting available light indoors in a dark restaurant, theater, subway station, and you'll quickly see why beyond ISO 6400 can be of help.

Up in a plane or helicopter do you have to worry about focus and stopping down to get a bit of depth of field? No, you just need shutter speed to freeze movement and you can leave a lens open. It's entirely different when you're only 5-25 feet from the subject.

As evinced by this Seattle night scene, there's not a lot of shadow to peer into from aerial shots at night, another big difference than when shooting indoors, where directed light can cast shadows and you want to be able to shoot what is in the shadows or at least not under direct illumination.

Also as recently as 8 years ago a useable ISO 6400 was unheard of.

So no, 100,000 is unrealistic today, but not in 5 years.

Rishi Sanyal:

Now I have my own Sony A7S raws, and indeed at ISO 25,600 cyan and magenta blotches are beginning to show in the shadows. And the grain is more pronounced than with the D4s or the Df.

Still the Sony 25600 raw is very usable, but the blotches in the shadows start to show up in strength at ISO 30,000.

Sony has a lot to be proud of here and the truly silent shutter is amazing. That shutter option in itself is a really big deal.

I'm not pounding the Canon drum... I'm agnostic when it comes to manufacturers. Cameras are purely tools to be used. Obviously the camera aperture is wide open when not pushing aperture preview. But to say your screen is showing you EXACTLY what you are capturing is utter nonsense. No screen can display what your aperture is capturing. That includes Wide Gamut pro monitors like Eizo. My point is that I like to see what a screen cannot.

Let's hope that monitor technology finally catches up to sensor technology.... because they are worlds apart right now.

It is purely marketing. Lots of non-photographers coming to the market and manufacturer must justify the price/size/weight difference comparing to the P&S.
First, in order to push p&s sales (because of cell-phones), they pushed the P&S ISO up to DSLR's level, although max acceptable ISO is still ~400, now they are pushing DSLR's "sensitivities".

I find it amazing how narrow minded some people are that are supposed to be working in a creative field. Just because high ISO does not give you any advantage to how you photograph does not make it worthless. Event photographers, street photographers and documentary photographers would all probably love good usable high ISO camera. You might find some of these photographers also have a different opinion of what quality is then you. I personally find this camera very interesting and look forward to see how it performs in real world examples.

At ISO 51000 and higher Canon has so much noise than A7S, that no amount of NR would help.
Actually at this amount of noise, NR is totally useless.

What the previous two said.

You can download the RAW files yourself & apply NR to see if you can get any real image detail out of the highest ISO shots. See if you can match the A7S in terms of shadow detail retention at ISO 204k & 409k.

DxOMark's Full SNR curves show that the pixel-level SNR of the A7S at ISO 409k is higher than that for the 5D Mark III in shadows. At some point (brighter tones), the SNR curves cross, & the 5DIII ISO 102k takes over with higher SNR. Naturally, as these brighter tones are photon/shot noise limited & therefore benefit from the longer exposure at ISO 102k.

Now, those are pixel-level SNR curves. Normalization will shift the ISO 102k 5DIII curve to the left, which'll move its crossing point (with the A7S 409k curve) to the left. In other words, the 5DIII ISO 102k file will be better at slightly darker tones than what the crossover point currently suggests. But I doubt it'll move that much, & so A7S 402k shadows will continue to show a 2 EV advantage.

That's a pretty silly argument! Canon has the clear advantage up to ISO 1600, it is about a draw from 1600 to 25600 with Sony having a VERY slight advantage strictly defined by noise but Canon having clearly superior all around image quality due to highlight performance, with Sony finally pulling ahead of Canon over 25600. 25600 is pretty well up there! So Canon's superiority goes well beyond "base ISO" and it's only in the realm where both look terrible that Sony has an advantage.

So you're saying when they're both great, one is better than the other but when they're both bad, neither is useful?

"Good Enough" more than describes low ISO performance on all of these cameras. It's getting to the point where people want more quality than their eyes can see. I don't need good sound reproduction at 40,000 hz, I can't hear it. I don't need more detail in a shot than I can see on a 4k screen.

k:

And I didn't say it twice. Take it up with the webmaster.

Downsampling does NOT remove noise, it's a near total myth.

If I had money, I might buy one, but I'd be more inclined to buy a D4s or Df--they're both better high ISO cameras. I'm still glad to see Sony releasing this camera. Hope they do an A99 II with this sensor.

HAR, I never said downsampling reduces noise. It actually reduces detail. to my eyes, at iso 25600, comparing the 7s to downside 7R is see no difference in noise. now, would be very interesting to print both at 300 dpi at 12x18. I bet you the A7R image would have more detail for same noise, or less noise for same detail, once processed from raw.

kevindar:

The problem remains that the A7r makes for a noisy blotchy mess at ISO 25600, and downsampling won't fix that.

This noise and magenta and cyan blotching will most certainly mess up detail. Better to stick with the better high ISO sensor to begin with.

Rishi Sanyal:

Nope.

It's a myth, try it with any setting in Photoshop. It may be what the math claims but when the results are that bad, and they are, then the math is suspect.

S/N ratio enhancement claims: My guess would be that the program counting signal is confused. This is often the case with this kind of engineering estimation.

The problem with the downsampling idea is that somehow downsampling would have to be able to readily identify noisy pixels in the process.

If removing noise were that easy, then NR would be real easy in the camera. And as you know it's not.

Now what sort of works, tossing about 90 percent of the data from a noisy image shot with many pixels may be a bit better than shooting with fewer pixels, but then you're left with a tiny thumbnail--that's pretty blurred.

Really, please try downsampling in detail, it doesn't work.

Perhaps something will radically improve in the understanding of this engineering in say 10 years--then maybe this will work.

" number of pixel has very nominal effect on the final output noise"

Assuming same sensor size, more pixels = smaller pixel. Each pixel generates read noise, thus more pixels per area = more noise per area.

Read noise is not bothersome when the signal is strong enough for shot noise to overwhelm read noise. But for a weak signal, read noise can be very significant.

High ISO means that the signal is weak. It is simple physics that sensors with smaller pixels will generate more noise at high ISO (with other variables being more-or-less equal).

Plus there are a few other potential sensitivity advantages of larger pixels. For example, larger pixels are less obstructed by circuit elements and may thus produce slightly superior QE than smaller pix.

AstroStan:

Yeah, spot on. Except that sometimes sensor designers will be able to decrease read noise per pixel in the higher resolution camera, in which case read noise per area does not necessarily have to be higher. All you have to do is decrease per-pixel read noise by sqrt(n) for a n times higher resolution sensor to make read noise per area the same. In which case normalized comparisons will show little to no noise differences between the lower res & higher res camera.

Therefore, in practice, lower resolution cameras don't always offer the advantages they're meant to. Although in this case, the A7S does - but only at very very high ISOs. You'd be hard pressed to see normalized differences between the A7S and A7R up to ISO 12,800 - have a look yourself in the widget.

Above that ISO - shadows see a significant benefit with the A7S' lower aggregate read noise (3x less # of read events) and potentially lower resulting quantization error due to the higher conversion gain.

Also, re: your point about potentially higher QE - this advantage is sometimes not as fully realized as it might otherwise be had gapless microlenses not been invented to help the potential light loss due to increased inter-pixel spacing.

As for the A7S - my preliminary results (so please don't quote me) indicate an approx. 8% higher signal in the Raw file over the A7R (same lens, aperture, shutter speed, ISO, lab lighting). That's about a 1/9 EV advantage just in terms of light gathering ability.

It's nice, but certainly not anything groundbreaking... my bet would be the noise optimizations are what really help.

HowaboutRAW: It's not a myth; one needs to go no further than our own studio scene to see this.

Have a look at this particular comparison of low resolution cameras on the left vs. high resolution cameras on the right at ISO 6400.

Although the low resolution cameras have lower noise at native resolutions, the leveling field evens out as soon as you click the 'Print' to normalize the comparisons by downsampling all cameras to the same (8 MP) resolution.

It can't get any clearer than that. Continuing to argue against what the visual results clearly show & the math dictates is just misleading our readers.

Are there limits to this? Yes. Is it a myth? No.

Rishi Sanyal:

And I have results that prove the fail, more than once.

For example the ISO 256000 raws from the A7r and A7s with the former downsampled. And that A7r is doing an okay job, but the A7r result is noiser--just as one would expect from the full sized image.

What makes you think I've never tried this?

I don't click "print" to downsample, I downsample in Photoshop. You can't make the A7(D610) shoot higher ISOs like the D4 by downsampling.

Now there is some possibility that "print" is introducing some other factor, and that, as a solution to the claims of NR, I'm open to. For example printing to paper, beyond 300ppi, does help with noise reduction.

Rishi Sanyal:

I checked again. Using the A7r and A7s ISO 25600 raws from the Seattle sky at night from the DPR test.

With the raws down-loaded, I extracted them to 8 bit tiffs with the same NR settings in ACR 8.5, within PhotoShop CS6.

Then I downsampled the A7r tiff so its pixel dimensions were the same and, I believe 4240 across. Result: The A7r tiff had a bit more grain, and much much more cyan and magenta blotching than the A7s tiff. (Let PhotoShop use the first resampling algorithm. I believe there are 6 to pick from.)

How many times do I have to repeat the test? Downsampling does not have some magic way of sorting signal from noise and holding on to signal. Now for that ISO, the A7r did surprisingly well for grain, not for blotching though.

We even talked about reasons for why the normalization might not be enough. E.g. the A7S might have the following going for it: lower aggregate read noise of 3x fewer pixels, potential QE increases, possibly decreased quantization error due to higher conversion gains, etc.

But as for reasons why the A7R can't keep with the A7S at higher ISOs, it's definitely not that 'downsampling does NOT reduce noise'. That's just incorrect, and a misleading statement. It's that downsamping/normalizing does not reduce noise enough to provide the same level of performance as the A7S for very, very low signals. Likely because of one or more of the reasons in the previous paragraph.

Rishi Sanyal:

Yes of course downsampling does reduce noise, it reduces the rest of the picture too. The cyan and magenta blotches get smaller, by pixel count.

What downsampling doesn't do is reduce noise as an overall fraction of the total pixels. Or: Downsampling can't select signal over noise and toss out the noise. Otherwise NR would be really easy.

That some of the downsampling options in Photoshop smear out noise differently, yes. And sometimes that can be helpful.

5D III doesn't shoot 4K video. 5D III can shoot raw video. You can frame grab from 4K video, equals quiet shutter anyway.

Jun2:

Raw to an external recorder, right?

No.... That's recording to a 1050x or faster CF card internally.