I'm going to snip most of the text here, because this is getting long.
LetTheRightLensIn said:
But you say that suddenly for read noise we can only compare photosite to photosite and that you can't normalize as you did above to see if the higher MP might do as well compared at the same scale. Why? How is this noise magically behaving differently and scale invariant? It sounds like are mixing up the relative invariance between how each photosite does for read noise with the way the noise results scale which have nothing to do with one another.
I mean what if sensor A has 40MP of photosites that have the same read as sensor B that has 10MP of photosites that have the same read. Now you say to look at only the ScreenDR and compare the tiny photosite of the 40MP camera to the big one of the 10MP camera and you say they do about the same and get about the same score. But you are comparing photosites that are at only 1/2 the scale of the others as if they were the same scale so if you want to know what the relative comparison between the two was in terms of how you could do in this regard on an even basis, well you are not doing that. You are just seeing how the 40MP camera would do if you took full advantage of the extra res of the 40MP, but nobody says you have to take advantage of the extra res. If you compare the 40MP camera the 10MP scale it would probably give a slightly better overall result or just call it the same to be simple as the 10MP camera. The PrintDR numbers would tell you that. But just using your ScreenDR numbers you'd think that there is no way you could ever be able to use the new camera to match the old camera's result.
LetTheRightLensIn said:
And I still await to hear why it makes sense to only compare mid-tone noise at normalized scale and NOT at 100% view, but it makes sense to only compare read noise at 100% view. Why does it make sense to compare apples to apples for mid-tone noise but makes sense to compare apples to oranges for deep tone noise???
and for that matter, when you want to simply see how a full RAW does, taking advantage of the full res of each RAW, you want to still normalize mid-tone noise, when in that case you should not be normalizing either mid-tone noise or shadow noise.
I don't get the point in comparing one always normalized and one never normalized. I mean you can do that if you want, who knows why, but whatever, but that is one thing, to then say that DxO is raw BS and the PrintDR is fake, has no bearing on reality, and nobody should ever compare in any way using that, blah blah blah, is just flat out wrong.
There is noise, and there is editing latitude. In pretty much every thread on here that ends up with the DR debate, what do people talk about? The amount of noise they can see, or the amount they can lift the shadows? In all the threads I've been party to, it all ultimately comes down to how much you can lift shadows. In none of the DR debates I've ever been party to has anyone ever said "You don't see as much noise with an Exmor sensor." No, the thing everyone always says, and the thing everyone always tries to demonstrate, is "Look how much I lifted the shadows! Look, I have a fully detailed sun, and detailed shadows, in this landscape photo. Oh, and look over here, the Canon sensor has tons of nasty red banding noise when I lift the shadows."
As far as I can tell, as far as most consumers are concerned, DR all boils down to EDITING LATITUDE. It means more shadow lifting.
So, why do I treat them differently? First, you are not wrong about comparing cameras...you need to normalize. And normalization affects read noise as much as it affects photon shot noise. But that is comparing final IQ. That's fine and dandy...but I believe people are misusing final IQ (of UNEDITED images) to refer to editing latitude. It's THAT, the use of normalized images to refer to editing latitude, that I believe is wrong. I believe DXO's publishing of Print DR exclusively on their ratings, completely ignoring Screen DR entirely, has lead to the misinterpretation of REAL WORLD editing latitude in a RAW.
Since everyone always ultimately arrives at "pulling shadows" in DR debates at one point or another, I'm always harping on that point. More DR means less noise, and normalizing means even less noise for larger sensors, but we cannot edit normalized images. We only edit full size images. So, from an editing latitude standpoint...I believe Print DR is invalid. For one thing, the Print DR numbers at DXO are only valid if you downsample an image to exactly that size. I don't think many people actually downsample their images to exactly 8x12 @ 300PPI all the time...hell, I think it's actually probably quite rare. So always referring to 14.4 stop of DR when discussing shadow lifting ability is plain and simply wrong in all the infinite other possibilities for image size. And that's not even mentioning that when it comes to the types of photography that tons of DR are most useful for, say landscapes, your probably UPsampling, rather than downsampling, which makes it even more invalid.
The other thing is that we don't edit RAW images after downsampling them. They would no longer be RAW. We edit RAW images at native size. So we deal with the native noise levels and noise frequencies
when we are pulling shadows. So yes, normalization will reduce all noise, including read noise, we can't downsample our RAWs...we must edit them at native size. If we take DXO's Print DR numbers as a reference for editing latitude, they would have you believe that you have more than an additional stop of editing latitude with a D800, and nearly two additional stops of editing latitude with a D810. That is plain and simply false. Hence the reason I treat read noise uniquely in the case of editing latitude.
Finally, as far as
visible noise goes from a human perception standpoint...read noise only exists in the deep shadows. It doesn't really matter if you have 38e- worth of read noise (as in the 1D X), or ~3e- worth (D800) or ~6e- worth (D810). You can't see it at native size. You still can't see it after normalization. Perceptually, read noise is inconsequential from a visual standpoint. Photon shot noise, on the other hand, or what you have called midtone noise (I think that's a bit of a misnomer...photon shot noise exists at every level of the signal, from the highlights down to the utter depths of the shadows, well below the read noise floor), affects the entire signal, and is the prime source of noise that we actually perceive in our photographs.
So, I don't believe that read noise frequencies are consequential to normalization...you can't see them anyway. That leaves photon shot noise as the primary noise culprit we are dealing with when normalizing images for comparison. Sure, read noise frequencies get normalized as well, but only a computer algorithm can tell the difference, so as far as I'm concerned, normalization of read noise frequencies is immaterial. When it comes to photon shot noise, well that is a part of quantization of the incoming photonic wavefront itself. Photon shot noise is ultimately determined by frame size, as your gathering the same amount of photons regardless of what your pixel size is. If your using a 1D X, each pixel is gathering more photons than the pixels of a D800. At native size, the D800 will appear noisier on a per-pixel basis, however after normalization there won't be any significant difference (in noise...the D800 still certainly maintains the advantage in overall detail, no question). There won't be any difference, because the amount of visible noise in those two photographs really has nothing to do with read noise...it has to do with the total quantity of light gathered. Now, the D800 has higher Q.E., so it should have an advantage...but at the same time, it also has a lower fill factor (more pixels, still an FSI design, so more die space has to be reserved for wiring and readout transistors.) I think in a normalized context, where a D800 image was downsampled to the same size as a 1D X image at ISO 100, the D800 would probably have a slight edge, as I don't think it's fill factor is going to entirely negate it's increased Q.E. However, fundamentally, the overall amount of perceptual noise (photon shot noise) in the images has to do with total sensor area. Read noise is there, and it is less in the D800...but we can't see the difference with our eyes. The shadows are shadows, both cameras have more dynamic range than any computer screen can handle anyway (~11stops for Canon, ~13 some stops for Nikon)...so the stuff in the shadows is buried several stops below the limit of a computer screen regardless.
The only time read noise becomes a meaningful factor is when your pulling shadows. THEN, and only then, does the advantage of having LESS read noise really become a meaningful issue. In that case, the D800, and any other Sony Exmor based camera, wins, hands down, no contest. However, and here is where DXO comes in again...A D800, D810, A7r, A7s, etc. don't have an 8x light gathering advantage over a Canon camera (as DXO's PrintDR numbers would have you believe). In fact, its about HALF that, one stop less, or a 4x advantage. Personally, I think being off by 100% is a meaningful thing. If DXO was saying the D800 had a 4.1x advantage over Canon cameras, I'd have never said a peep. But saying the D800 has nearly an 8x advantage over Canon cameras...yeah, I have a problem with that.
So...I keep read noise levels in the context of discussions on DR (which pretty much ALWAYS end up referring to shadow lifting ability at some point), distinct from the whole concept of normalization. Because were talking about editing latitude, something that cannot be compared in a normalized context (at least, as far as I see it.)
Well, I don't think I can explain my stance any better than that. I'm guessing you still disagree, but that's ok. Nothing either of us can do about that at this point.
