Epic is the Ultimate "Film Slayer".... IF

[Daniel Browning]

That's very interesting. I've never heard that argument before.

Thanks, I enjoy sharing useful information.

So you are saying that given two sensors that are the same size and that process images the same way the one with MORE pixels will have LESS noise?

Yes. The greatest density of photosites are the most efficient and the lowest density are the least efficient. By that I mean the signal to noise ratio.

Larger sensors do not have a proportionately less amount of noise than smaller sensors, relative to area, which is light gathering ability. In other words, a sensor that is 23 times larger does not have that much less noise. As size drops, efficiency (per unit of area) goes *up*. More on this below.

Everything I've experienced myself and have encountered in the work of others seems to work counter to your argument. You state that the 1DsMarkIII has the lowest noise of any DSLR, but isn't the Nikon D3 the current low noise king?

The D3 is the per-pixel low-noise king, but the 1Dsm3 is the per-image low-noise king, by a small margin. Only the latter title has any meaning.

It has half the photosites of the 1DsMarkIII and achieves lower noise with less severe noise reduction. I'm just pointing to one comparison here but there are far more on the net with the same conlcusions: http://luminous-landscape.com/essays/comparison.shtml

Your edit is correct, this is the kind of flawed comparison that I was talking about when I said "People think smaller photosites are worse because they make flawed comparisons, such as viewing 100% pixel size instead of correctly resampling to the same resolution with a sinc algorithm".

To illustrate, imagine a 200 MP camera, let's call it ACS (I love Hubble!). If we compared 100% crops from ACS and D3, and the ACS had three times more noise, this flawed comparison would indicate that the D3 is better. Now imagine resizing those 200 megapixels down to just 12. The reduction in noise would be far greater than just a magnitude of 3, and the D3 would appear to be much noisier.

Now change the illustration from ACS to 1Ds3 and 200 megapixels to 21. See how it reduces apparent noise?

The comparison is further flawed by using inequal exposures (1/40 vs. 1/50), not accounting for Nikon's black point clipping (not apparent in their image anyway) or especially their chroma desaturation.

In it Michael Reichman reiterates the point that I thought was common knowledge: "The 1D MKIII's sensor at 10MP is closer to the D3's 12MP than the 1Ds MKIII's 21MP, and the pixels of the Nikon are bigger, so theoretically of lower noise. The D3 is also closer in price to the 1D MKIII than to the 1Ds MKIII."

The D3 is an excellent camera. I think it has as much as a full stop less image noise than the 10 MP 1Dm3, quite an achievement.

I thought that sensors with larger and fewer photosites were inherently less noisy than ones that are more denseley packed with smaller photosites.

Again, true at the pixel level but not when viewing the image as a whole. Another illustration of the 100% comparison flaw: take a beautiful 12 MP image: the 100% crops look great and the 8x10 prints look great. Now resize it to 21 MP. Add just a little noise in photoshop, enough to be visible at 100%, but not enough to be visible in the 8x10 print. Print both at 8x10 and compare: they are identical. But at 100% crops, one looks worse than the other.

In real life, are you going to crop a 4x6 out of your 21 MP image? Or are you going to resize it to 4x6 for printing? (People do all sorts of things online that they wouldn't in real life.)

Isn't this why 12mp digicams are far noisier than 6mp digicams at high isos?

A common misconception. Resize the 12 MP down to 6 MP and it will look better. Especially if the 6 MP is an older model.

Isn't this why Pentax had to reengineer the way their sensors worked in order to achieve the same level of noise in their 14mp K20d as they had in the 10mp K10d.

Right. Engineers are increasing density and efficiency at the same time. The big sensors are still nowhere near the digicam efficiency, but they've been following Moore's Law so far, which means we have a lot to look forward to.

Does the D300 have lower noise then the D3 then?

No.

You claim that people aren't resizing and resampling images properly when doing these tests. So why not try to compare images from cameras with the same number of photosites but in different sized sensors.

Let's compare the Canon G9 with the Canon 5D. Both 12 megapixels. The G9 packs 12 megapixels in 37 square mm. The 5D spreads itself out over a luxurious 864 square mm. That means the 5D has the opportunity to capture 23 times more light (a function of area). But does it? And does it have 23 times less noise? No, because it is far less efficient than the digicam.

Imagine someone took the Canon 5D sensor and wanted to make a 1/1.7" digicam out of it. So they got an exacto knife out and cut the center 37 square mm out of the 5D sensor and put it in a little camera. On that area, the 5D had 0.5 megapixels. But they're *good* pixels. If you look at the 0.5 MP image at 100%, it's clean and nice, whereas the 12 MP G9 image at 100% looks very noisey. Now resize the 12 MP G9 down to 0.5 MP... the 5D crop looks terrible in comparison. If you don't have an exacto knife, you can do the same experiment by cropping the center 0.5 MP in photoshop out of the Canon 5D. Thus proving that sensors can have *more* pixels in the same area to capture more light and less noise.

If they made the 5D with the same pixels as the digicam, it would have 280 megapixels. And by the time you resize that back down to 12 MP to compare with the real-life 5D, it would have a lot less noise. Obviously, they would if they could. But it dispels the fear that more pixels will mean more noise.

I know for a fact that the images I get off my 6mp DSLR are far cleaner than those that I get off of my 6mp Canon Powershot. Would you attribute this to the dslr having heavier noise reduction? I always thought it was because the photosites in the dslr were larger and less densley packed therefore generating less heat.

I have the same experience with my Canon DSLR and digicams. In addition to the comparison flaws I have explained, I think there is another big reason why so many have the misconception about dense pixels: digicams stink. They have small sensors and terrible processing (no raw, or if they have it it's dog slow). Thank goodness RED got it right.

 

[Steve White]

Daniel Browning - your posts in this thread have been nothing short of wonderful. Thank you for clearing up the misconceptions about noise.

I think the legitimate fear of the smaller photosites has to do with clipping and dynamic range. For a given photosite, you can only bin so many photons before you've filled it. The total number of photons you can represent in your image is your dynamic range for that photosite. This argument is flawed too however, since what really limits you is the combination of how many photons you collect, how many photosites you have, and how many photons each photosite can hold (plus what the signal to noise of each photosite is).

Suppose for example, that you bring in 100 detectable photons with your lens and focus them onto a single photosite, that can hold 50 photons. In this case, you end up with clipping.

On the other hand, suppose you had 100 photons evenly focused into 4 smaller photosites that could collect 25 photons each, occupying the same area as your 50 photosite detector. In this system, you actually collect all 100 photons and have fully capured the image. While the dynamic range of each photosite is lower, the dynamic range of the image is higher.

Now the problem is in how you represent the data. It's easy to make mistakes of course, but if you sampled RAW, you're at a good starting point so that someone with better algorithms can get better results.

I know for a fact that the images I get off my 6mp DSLR are far cleaner than those that I get off of my 6mp Canon Powershot. Would you attribute this to the dslr having heavier noise reduction? I always thought it was because the photosites in the dslr were larger and less densley packed therefore generating less heat.

Keep in mind that the DSLR probably has a significantly larger lens than your digicam. Suppose they had the exact same sensor, but the DSLR had a 72 mm outside diameter lens, and the digicam had a 36 mm outside diameter lens. In this case, the 72 mm lens collects 4x as much light as the 36 mm lens. Thus, the sensor has 4x as much information to produce the image. This is of course, why telescopes have such large lenses - they need to collect a lot of light.

 

[Bing Bailey]

wasn't this an EPIC thread ?

 

[Steven Caesare]

Daniel Browning - your posts in this thread have been nothing short of wonderful. Thank you for clearing up the misconceptions about noise.

I think the legitimate fear of the smaller photosites has to do with clipping and dynamic range. For a given photosite, you can only bin so many photons before you've filled it. The total number of photons you can represent in your image is your dynamic range for that photosite. This argument is flawed too however, since what really limits you is the combination of how many photons you collect, how many photosites you have, and how many photons each photosite can hold (plus what the signal to noise of each photosite is).

Suppose for example, that you bring in 100 detectable photons with your lens and focus them onto a single photosite, that can hold 50 photons. In this case, you end up with clipping.

On the other hand, suppose you had 100 photons evenly focused into 4 smaller photosites that could collect 25 photons each, occupying the same area as your 50 photosite detector. In this system, you actually collect all 100 photons and have fully capured the image. While the dynamic range of each photosite is lower, the dynamic range of the image is higher.

Now the problem is in how you represent the data. It's easy to make mistakes of course, but if you sampled RAW, you're at a good starting point so that someone with better algorithms can get better results.


Keep in mind that the DSLR probably has a significantly larger lens than your digicam. Suppose they had the exact same sensor, but the DSLR had a 72 mm outside diameter lens, and the digicam had a 36 mm outside diameter lens. In this case, the 72 mm lens collects 4x as much light as the 36 mm lens. Thus, the sensor has 4x as much information to produce the image. This is of course, why telescopes have such large lenses - they need to collect a lot of light.

But your photon count per area ratio doesn't change. If the smaller photosite can only handle 50 photon, but you can fit twice as many sites in the same physical area, you still can handle detecting all ths photons.

Two five gallon buckets fit in my bath tub. I can pour 10 gallons of water in and hold it all. replace the buckets with ones that can only hold 2.5 gallons each. However, four of these smaller buckets fit in my bathtub. I can still pur 10 gallons of water...

 

[Zack Birlew]

Epic is going to be awesome, that much is clear. The more interesting question is how is the RED One going to upgrade its sensor and how much will that cost?

Once that gets into the picture, then things will really get interesting. Also, the accessories are obviously going to be different, otherwise, why would you buy the Epic if you can save *possibly* a few thousand by buying a RED One with an upgraded sensor? Perhaps there's something about Epic that hasn't been revealed yet?

Well, enough speculation. I agree, from what I've been reading about the RED One so far, the dynamic range has got to get better. Having 11 stops or whatever at a high ISO isn't very helpful if you can't even see the images through the garbled video gain. Another thing that could change the game is if RED actually does implement or allows for the IMS interchangeable lens system by default. Then the camera would truly be open to independent and pro filmmaker alike, because then you can use any lense with the RED cameras.

Workflow is another big issue. This year's NAB and REDUser party were very eye opening as far as the RED workflow goes. I thought my Mac Pro would have handled everything just fine, well, now I know that I would need a Kona card or equivalent, a production LCD, and a big RAID array in order to do the kinds of things I planned on doing with my RED. Even then, I would need at least a Macbook Pro in the field with me just so I could offload footage, not a big deal but it didn't hit me until then. Also, Scratch would almost be a neccessity. But, granted, you can get around a lot of this by working in 2K or lower resolution. However, I think the biggest problem right now is that there still isn't support for 4K 16:9. Obviously most of this can be solved with software support and bigger storage devices (ie. RED Flash, fast 32gb CF cards, etc.) but that's just the way it is at the moment and a big part of what needs to be overcome to become a true replacement to film.

 

[Zakaree Sandberg]

honestly.. I think Red One, is the "film killer" that we were all looking for (not literally, but maybe..) I would hate to see people off the bat shy away from using RED ONE because Epic is a bigger badder version. The New sensor will be a must have, wether you stick with red one and opt for the upgrade or go ahead and grab the epic. I believe Epic is designed for HIGH HIGH end work and will be amazing on the effects shots, but RED ONE will still be a work horse for most features, commercials, and tv shows (hopefully)
How much better can things get? i mean I had the HD laser eye surgery and i see 20/13 and yet i thought superbad (shot on genesis) looked good. I havent seen RED stuff Printed to film and projected yet, But i cant wait.

 

[Steve White]

But your photon count per area ratio doesn't change.

That is not a fact, that is an assumption. If you were designing photosites, you would want to increase this ratio, lower the area, and decrease the noise floor.

 

[Tom Lowe]

Zakaree, I never thought Red One would be the "nail the in coffin" for film. I have always speculated (openly here and at cinematography.com, etc) that Red's second offering would be the beast to slay chemical film for motion pictures. Epic seems tailored made for this role, if only the DR can be brought closer to film. But as some have pointed out, even high-end DSLRs have not completely matched film's DR, but that did not stop them from putting film still cameras out to pasture. (I realize that some photogs do continue to shoot chemical film stills, but they are few and far between these days.)

Daniel Browning, thanks for your posts. I think I actually learned something here today! What you seem to be saying is that a Canon 350D (which I shoot on) will have less visible noise than a 450D, when viewed pixel for pixel, but once you downsample the 450D's image to 8MP, it will appear cleaner than the 350D. So, for example, if your end goal was a 1920x1080 pixel frame, the 450D downsampled to 1920x1080 would show less noise than the 350D downsampled to 1920x1080.

 

[Daniel Browning]

Daniel Browning, thanks for your posts. I think I actually learned something here today! What you seem to be saying is that a Canon 350D (which I shoot on) will have less visible noise than a 450D, when viewed pixel for pixel, but once you downsample the 450D's image to 8MP, it will appear cleaner than the 350D. So, for example, if your end goal was a 1920x1080 pixel frame, the 450D downsampled to 1920x1080 would show less noise than the 350D downsampled to 1920x1080.

You're welcome. Yes, that's what I'm saying. The 450D is no slouch when it comes to per-pixel noise, either, but I think the 350D still beats it by a small margin. (I haven't got a 450D yet, I'm waiting for the 1000D ;-).)

Your remark about downsampling to 1920x1080 highlights an important difference between stills and digital cinema. With still photography, you can choose between getting less noise by printing smaller (8x10), or getting more resolution by printing larger (30x20).

That choice doesn't exist in cinema. When I first saw a top-notch 2K digital projection, the difference from film projection was so striking that I vowed to never go back to a regular 35mm theater again. Unfortunately, 2K digital projection is rare and 4K is rarer still. Sure, filmouts are ubiquitous, and the whitepapers cite nearly 1.4K, but my eyes tell me the comparison is even more disparate. Anyway, 1080p, 720, DVD9, and web distribution will all benefit from the resize. Remember the ISO 12,000 stills that Jim posted?

By the way, Tom, I really enjoyed your Joshua Tree astrophoto timelapse. Last time I was there I had telescopes but no cameras. Next weekend I'm shooting some star-and-landscape timelapse in the Oregon desert with a 5D, 20D, 350D, and the 24mm f/1.4.

 

[Dj Joofa]

Daniel, in general there is a trade off between resolution and dynamic range. However, it is not always clear how to achieve the most optimal trade off.

Again, true at the pixel level but not when viewing the image as a whole. Another illustration of the 100% comparison flaw: take a beautiful 12 MP image: the 100% crops look great and the 8x10 prints look great. Now resize it to 21 MP. Add just a little noise in photoshop, enough to be visible at 100%, but not enough to be visible in the 8x10 print. Print both at 8x10 and compare: they are identical. But at 100% crops, one looks worse than the other.

I don't think the above example is a good reflection of what you want to say. Consider this. You started with a 12 MP image, which had some spectrum, and you upsampled it to 21 MP. Now the upsampled image has a compressed spectrum in relation to the original 12 MP image, in addition to the replicas of the compressed spectrum. (The replicas don't concern us too much as you would going to get rid of them when you later downsize.)

Now you added noise to the 21 MP image. Typically noise will manifest itself as high frequencies and probably will do damage in the areas after (or starting nearby) the first compressed spectrum finishes. So at 1-1 at this stage you of course see that noise. Now you downsized. The downsizing will remove the noise you added, and stretch the compressed spectrum to more or less the same size back as you started.

You see the problem, you got back to where you started from. I don't think this is a good example. A real example would be when you start from a high-resolution image with noise, as in this case the noise spectrum is overlapping the signal spectrum, because the signal spectrum now is unmodified and not compressed in relation to some other image. Now you downsized. The downsizing will remove high frequencies and stretch the spectrum.

However, the difference here is that unlike the first case where the removed high-frequencies were mostly noise, in this case it is noise + some actual signal spectrum. Therefore, the image becomes a little blurred / smoothed, and that is the price you pay for downsizing.

Now that is a separate question that is the blurred downsized image still better in visual quality than a (smaller) sized image of the same size as the downsized image, i.e., acquired with a sensor with fewer number of photosites?

 

[Phil Bates]

If there was no Phantom, film's last hold out for me would be frame rate. When I shoot on film, I rent a 435 with the capability to shoot 160fps. I am hoping Epic can do better than that at 2K.

Phil
www.artbeats.com

 

[Tom Lowe]

By the way, Tom, I really enjoyed your Joshua Tree astrophoto timelapse. Last time I was there I had telescopes but no cameras. Next weekend I'm shooting some star-and-landscape timelapse in the Oregon desert with a 5D, 20D, 350D, and the 24mm f/1.4.

very cool. my mom lives up in the Salem area, so I was going to check out the deserts there late this summer. do me a favor and post the results from that shoot here or over at the timescapes forum (see my signature) if you get a chance. i'll be curious to see how your shots turn out. good luck!

 

[Dan Hudgins]

22 stop range

22 stop range

It seems to me that Epic is the perfect storm to finally put chemical film out to pasture, in much the same way that high-end DSLRs did over the last few years for stills. It will shoot a "real" 4K, RAW, s35mm, cine glass, overcranking onboard, solid-state or harddrive recording, etc.

The ONE THING Epic needs to do in order to fulfill it's destiny, though, is match (or come close to) chemical film's dynamic range. Aside from nostalgia, DR is really the last leg film is standing on right now, IMO.

Similar to Clinton '92's precept "It's the economy, stupid," the Epic team should have a poster on the wall reading, "It's the dynamic range, stupid." :weight_lift:

With a sensor upgrade next year, I think Red One will do the same.

If you put two sensors in and a 45 degree Pellicle mirror you can overlap the curves and get more than 20 stop range to the tones, in other words more than the lens can take even with good coating. You could make a cemented lens that might make 18 stops for the camera. The image would look flat though on the screen since that dynamic range would need to be put in what can be projected.

I went through this with my brother using special films to get detail in the shadow under a car and in the reflections of the sun on its chrome at the same time, after all the work to get that he said it looked too flat and he wanted more contrast! Be careful what you ask for!

You can see what this would look like, make two shots with your RED ONE (tm) one +8 stops and one -8 stops, then use my fuse command to mix the two images together. To get better results the top end should be rolled off before you mix the images so there is not a hard clip, and the shadow should be rolled off so the blocking does not get mixed in, i.e. two S curves before mixing the frames.

There is no reason that they cannot make the Epic have a 18 stop or better range with two sensors since the streaking issue for the hot low end sensor seems under control and they know how to make a IR free ND filter for the high end sensor.

If they use 6 mono sensors they can make a camera with better color and more range than film, that WOULD end the issue once and for all time(?)

 

[Daniel Browning]

Daniel, in general there is a trade off between resolution and dynamic range. However, it is not always clear how to achieve the most optimal trade off.

I think the trade off is pretty clear. Image noise, for a constant size image, has decreased far more in recent years due to higher numbers of efficient pixels, than it has for reducing noise at the pixel level. You can see it by looking at any given sensor size and comparing performance at the pixel and image level. Both have improved, but the latter has improved faster and has more proven future potential (on 1/1.7" now, and hopefully in larger sensors in the future).

I don't think the above example is a good reflection of what you want to say. [...] You see the problem, you got back to where you started from. [...] However, the difference here is that unlike the first case where the removed high-frequencies were mostly noise, in this case it is noise + some actual signal spectrum. Therefore, the image becomes a little blurred / smoothed, and that is the price you pay for downsizing.

You're quite right. Thank you for pointing out my mistake.

Now that is a separate question that is the blurred downsized image still better in visual quality than a (smaller) sized image of the same size as the downsized image, i.e., acquired with a sensor with fewer number of photosites?

The downsize also has the benefit of removing the softening effect of the OLPF (to the degree that it is resized by the amount of the optical filtration), whereas the native image does not.

There are, of course, a great variety of downsizing and sharpening algorithms and their varied effect on subject matter, depending on the sizes and ratios involved and individual taste. And, as always, YMMV; the FZ50, for example, bungled the analog ISO so that 200 pushed three stops has far less noise than the analog 1600.

There are a lot of same-sized sensors out there, so it's easy to make your own comparison. To me, the downsized photos have clearly better detail and noise characteristics.

 

[Dj Joofa]

I think the trade off is pretty clear. Image noise, for a constant size image, has decreased far more in recent years due to higher numbers of efficient pixels, than it has for reducing noise at the pixel level.

Daniel, it is a simple matter if you are doing things manually as many software provide fixed length filters with fixed co-efficients with a few selectable options, and even then they are not optimal.

An automated software that is always "optimal" without manual supervision is a whole different beast. Before we go forward we need an understanding of what is the definition of being "optimal?" And, for that we are going to dig on the surface of some estimation theory here.

Whenever there is a mentioning of the word "optimal", it should all always come with the question but in what sense? And, typically answer boils down to a certain metric (objective function).

Suppose you are downsizing. So you are given a filter and at each location you convolve it with certain neighboring pixel values. Convolution basically means that you multiply filter coefficients with pixel values and sum them up. Now this step can be seen as linear prediction in the classical estimation theory consideration:

Given n samples (pixels), to go with a filter with n number of co-efficients, with what is the best estimate of a quantity (your current pixel value) at hand. Least squares is a widely used metric in signal processing. Now under certain assumptions (which are very realistic) the arithmetic mean is the least square estimator for a number of samples. Hence, an optimal estimator in a widely used sense.

However, we know that a filter that downsamples by averaging has poor visual response. Why did that discrepancy happen? Because, this particular estimator did not employ the photometric visual response.

If you are going to estimate a quantity from n samples with a linear combination such as c_i * x_i, where, c_i are co-efficients, and x_i are samples, and i are index, then based upon the values of c_i you can realize many different types of filters.

All c_i = 1/n gives a mean filter, just seen above. You can even have a (truncated) sinc filter, lanczos, etc. However, we just saw that in the least squares sense mean is the best one. Even better than lanczos?

Interestingly, though least square sense gave you a sense of the nature of the estimator, it did not say that th co-efficients c_i are the "best" ones. It just minimized the objective function in terms of "error". If you have some a priori notion about the probability distribution of c_i's then you can even use a Bayesian estimator, for a quadratic loss function for finding a set of c_i's that are closer to their "actual unknown values."

Now to make matters more complicated, if we assume stochastic behavior, then the beast starts getting unmanageable, as a whole new set of filters pop up, Wiener filter, Kalman filter, etc.

None of them is particularly optimized for visual response. On the other hand those filters that work visually may not be the most optimal under certain sense.

The problem then boils down to incorporating photometric visual weights into objective functions. Now a whole, separate area has sprung up, as we are not always clear that what is the best way that such visual weights can be incorporated into a purely number crunching statistical metric.

I shall leave it at there as it is already giving me nasty aftertastes.

 

[Daniel Browning]

Now to make matters more complicated...

That whooshing sound was your post going over my head.

 

[Radoslav Karapetkov]

What about randomized pixels?

That's another characteristic of film that I don't see in digital.

Is there some good software solution to achieve the random-pixels effect of film?

 

[Steven Caesare]

That is not a fact, that is an assumption. If you were designing photosites, you would want to increase this ratio, lower the area, and decrease the noise floor.

Isn't the light intensity (i.e. photon count per unit of surface area) determined by how much light your lens gathers and focuses on a given area?

So if you maintain the same sensor size (i.e. surface area), and use the same lenses, then how can the sensor's photosite size/count change that intensity?

 

[Steve White]

So if you maintain the same sensor size (i.e. surface area), and use the same lenses, then how can the sensor's photosite size/count change that intensity?

The photosites can't change the intensity, they measure it. Every photosite you have can measure the intensity over some absolute range, with some uncertainty. For example, you could have a photosite that can generate a signal from 0 to 1 Volt. This is your dynamic range. This signal is sampled by a Analog to Digital converter with some bit-depth - say 4096 values (12-bit). These 4096 divisions are your precision.

Even if no desired photons struck the photo-site, you would still get a signal, say 100 +/- 50 mV. In this case, the 100 mV is called the noise floor, a signal value below this is not terribly useful, since you have difficulty telling if it's signal or not... so you would probably try and crush the noise to black. To do this, you would probably need to sluff off 150 mV or so, leaving you with 850 mV of useful information. You still have the +/- 50 mV noise in the rest of the image though.

In order to improve the dynamic range of a photosite, you want to do several things.
First off, 0 to 1 Volts is your total dynamic range. If a single photon generated 1 mV, then you could collect 1,000 photons. If it generated 1 μV, you could collect 1,000,000 photons. Alternatively, you could invert the math, and sample 1000 V. You can guess which approach uses less power. Therefore, if possible, you want to decrease the sensitivity/gain. Low sensitivity sounds bad, but it isn't if you also have low noise and high precision.

Next up, you want to decrease the +/- 50 mV uncertainty. By reducing your noise variance, you effectively lower the noise floor. If you brought the uncertainty down to +/- 10 mV, you would how have 890 mV of useful information. This has the added benefit of reducing your image noise too.

Finally you want to decrease the noise floor. If you got it down to +/- 20 mV, you would now have about 970 mV of useful information.

Once you optimize your sensor to a noise floor and uncertainty on the order of 1/4096 of your total dynamic range, you increase the bit-depth of your D/A and start over.

This analysis assumes that the signal you get off your photosite is linearly proportional to number of "photons". I'm pretty sure CMOS and CCDs are linearly proportional to some measure of intensity per unit area. These assumptions don't really work for film due to the way grains collect light.

 

[Steven Caesare]

Right. They can only measure.

Hence the example of concern of a smaller site being "overloaded" by 100 photons is offset by the fact that the site is smaller, and fewer photons fall on it.