How do they get all this dynamic range?

[Lee Jay]

I feel my 5D has about 8-9 stops of dynamic range. The claim of RED is that their FF-sized sensor has 13 stops of dynamic range.

To me, getting 4 stops more dynamic range requires one or more of the following things:

- 16 times more well capacity. This means base ISO drops from 100 to 6.25. However, Jim stated it would be in the 320-500 range, not 6.
- 16 times more quantum efficiency. This would put it well over 100%, and so would violate the laws of physics.
- A different way of measuring dynamic range. If it's defined as the ratio of full-well to black-frame noise, it could be that high, but that is largely a useless method of measuring dynamic range for imaging in my opinion.
- Something else I don't understand at all, hence this thread.

 

[Carlos Echenique]

TBH, I was wondering that myself. The Canon 5D sports 8.2 micron photosites, larger than the Scarlet FF35. But the Scarlet FF35 has more bits per channel.

I would like to know how they estimated those values.

 

[J. Bernard Vallon]

- 16 times more quantum efficiency. This would put it well over 100%, and so would violate the laws of physics.

This is false. If the efficiency is 1/16 what it was error still exists, but it is much smaller. Information is still lost to heat, the laws of entropy still hold.

Just look at the DR of the human eye. Much more efficient on the low end, much less error, still good physics. Smaller 'pixel wells' too i believe.

 

[Lee Jay]

This is false. If the efficiency is 1/16 what it was error still exists, but it is much smaller. Information is still lost to heat, the laws of entropy still hold.

Huh? The laws we're looking at here are shot noise and dark-frame noise (a combination of noise sources). More light = more shot noise, but also more signal. The result is more signal-to-noise ratio. In order to have more signal-to-noise ratio in deep shadows (few photons impinging) you need to capture a larger fraction of them. Current silicon sensors are in the 20-40% range. Increasing that is difficult (especially in a Bayer sensor, which absorbs a lot of the light on purpose), and there's no room to get enough more to increase the SNR enough to match the claims anyway.

Just look at the DR of the human eye. Much more efficient on the low end, much less error, still good physics. Smaller 'pixel wells' too i believe.

The eye works so well because it can alter both sensitivity and integration time *locally* on the sensor (retina). Dark portions can have longer integration times and higher sensitivities, while brighter ones can have lower sensitivities and shorter integration times. This results in extreme dynamic range. As far as I know, the RED sensors can do neither one.

 

[Daniel Browning]

Lee,

It's great to see you participating here more and more. I have appreciated some of your contributions to DPR in the past.

To me, getting 4 stops more dynamic range requires...one or more of the following things:

[snip good commentary on well capacity and quantum efficiency]

- Something else I don't understand at all, hence this thread.

Very few commercial sensors have implemented any kind of in-sensor dynamic range hardware. Fuji is the only one I can think of, and it's not terribly impressive. Eric Fossum said that there are dozens of excellent ideas (probably all patented) for extending dynamic range, so it's possible that RED has done something like that.

My personal guess, however, is that RED has simply improved the read noise.

- A different way of measuring dynamic range. If it's defined as the ratio of full-well to black-frame noise, it could be that high, but that is largely a useless method of measuring dynamic range for imaging in my opinion.

You are exactly correct. RED follows the same convention that all sensor designers, manufacturers, engineers, and everyone else uses: FWC/read noise, where SNR is 1:1. Here is a post that describes RED's measurement:

Understanding Mysterium

We've had a variety of threads about dynamic range recently, and one of them seems to address exactly what you're saying here:

13+ stops....

I highly suggest that you read it. If it does not address your thoughts, then I would ask you to describe your objections with using SNR of 1:1 for measuring dynamic range.

I would describe the utility of that method this way. Everyone has a different level of personal taste, desired output, and circumstances. It sounds like you would never go below SNR of 8:1, but other people use 4:1 just fine. I even go *below* 1:1 sometimes, such as for youtube output (1:2, 1:4, or worse!).

So how is RED, or any manufacturer, supposed to give a number that will be useful to everyone? Should they poll the photographers for their favorite number and use the average (say, SNR 8:1)? Or follow my advice and quote the number after it's been downsampled to DVD (SNR 1:8)? Or just follow the convention everyone else uses: SNR 1:1?

The answer is that it doesn't matter. As long as RED states the nature of their dynamic range tests, and they do, then we can look at them and extrapolate for our own needs, or, even better, do our own tests.

In order to have more signal-to-noise ratio in deep shadows (few photons impinging) you need to capture a larger fraction of them.

Or, simply reduce noise. As you said, there is very little room for improvement in quantum efficiency, so DNR improvements from here on out will mostly be from read noise improvement, sensor size increase, or on-sensor DNR hardware.

Kind regards,

 

[Lee Jay]

Lee,
You are exactly correct. RED follows the same convention that all sensor designers, manufacturers, engineers, and everyone else uses: FWC/read noise, where SNR is 1:1. Here is a post that describes RED's measurement:

Understanding Mysterium

We've had a variety of threads about dynamic range recently, and one of them seems to address exactly what you're saying here:

13+ stops....

I highly suggest that you read it. If it does not address your thoughts, then I would ask you to describe your objections with using SNR of 1:1 for measuring dynamic range.

Okay. The reason I object is what Jim stated in the other thread:

"Measure DNR and actual shooting latitude are not necessarily the same thing."

Having read noise of zero, as in some astro sensors, doesn't equate to a DNR of infinity on the images, unless you call 1 LSB changes in the image "shadow detail". Shot noise matters or those doing that astro imaging wouldn't be using long integration times and stacking to get huge dynamic range (sometimes up to 30 stops) in their final images. Also, this method doesn't reward large wells the same as it does low read noise, whereas in actual images, large wells are more important for perceived image latitude.

As you said, there is very little room for improvement in quantum efficiency, so DNR improvements from here on out will mostly be from read noise improvement, sensor size increase, or on-sensor DNR hardware.

Kind regards,

I don't agree. The concept of continuous cell reset allows infinite dynamic range through effectively infinite well capacity, and therefore arbitrarily low ISOs and long integration times without blowing highlights. Unfortunately, REDs sensors seem to go the opposite direction from this ideal by having high base ISOs, and therefore small well capacities. I would *guess* that a standard Canon or Sony sensor at ISO 100 would have more real dynamic range than any RED sensor at ISO 500 even if REDs sensors have slightly lower read noise (and it has to be slightly since we're already talking about 1-3 counts in those other sensors). Perhaps that doesn't matter because REDs are always shot in poor light and integration times are more-or-less fixed anyway. To me, capturing 4 photons with 2 counts of shot noise does not constitute a usable portion of the image, even if the read noise is zero. Like I said, shot noise matters in any real image.

 

[Jannard]

Lee Jay... how do you explain that Mysterium-X (12-bit, same ISO and same pixel size) has 1 full stop more DNR and less noise... and is twice as fast as Mysterium?

Jim

 

[Daniel Browning]

Having read noise of zero, as in some astro sensors, doesn't equate to a DNR of infinity on the images, unless you call 1 LSB changes in the image "shadow detail".

I've never really thought about that because it's so far away from happening in the real world; I'll have to mull it over.

Anyway, I thought sub-electron read noise was only achieved in laboratory conditions. The SBIG astro sensors still have high read noise. What astro sensor are you referring to?

Shot noise matters

I agree completely, but QE+FWC tells you exactly how much shot noise you'll get, which is why it's often mentioned in the same breath as DNR. I don't see the need to work that metric into DNR, but how would you do it?

I don't agree. The concept of continuous cell reset...

That is what I was referring to by "on-sensor DNR hardware". Continuous cell reset is just one of dozens of ideas for improving dynamic range in the sensor.

Unfortunately, REDs sensors seem to go the opposite direction from this ideal by having high base ISOs, and therefore small well capacities.

What do you think RED's base ISO is, and what are you basing that on? Are you aware that RED ONE does not have a variable gain amplifier? And that ISO is metadata? Have you considered that RED rated for ISO 320 may not be 3.5 stops below FWC? Have you looked at any R3D files? There is a lot of footage at RedRelay.net and the RED raw conversion software is a free download.

I would *guess* that a standard Canon or Sony sensor at ISO 100 would have more real dynamic range than any RED sensor at ISO 500

I would guess the opposite.

Like I said, shot noise matters in any real image.

Agreed.

 

[Evin Grant]

Me thinks it's the drinks at Red...

 

[Lee Jay]

Lee Jay... how do you explain that Mysterium-X (12-bit, same ISO and same pixel size) has 1 full stop more DNR and less noise... and is twice as fast as Mysterium?

Jim

Well, more DNR by the definition you're using would come from a tiny reduction in dark-frame noise (say, from 3 counts to 1.5 counts). "Less noise" is a bit of a nebulous term, but generally less overall image noise comes from either larger microlens total aperture or better quantum efficiency. Speed is another issue, and one I haven't addressed.

 

[Lee Jay]

I've never really thought about that because it's so far away from happening in the real world; I'll have to mull it over.

Anyway, I thought sub-electron read noise was only achieved in laboratory conditions. The SBIG astro sensors still have high read noise. What astro sensor are you referring to?

There is a technique available now, for which I can't presently find the info, for reducing read noise by an arbitrary fraction (often several thousand) thereby effectively reducing it to zero. I've read that such sensors are available now and are in use in professional astro applications. I'll keep digging for the info, but it's hard doing a search when you can't remember the name!

That is what I was referring to by "on-sensor DNR hardware". Continuous cell reset is just one of dozens of ideas for improving dynamic range in the sensor.

Oh, okay, and I agree that there are other options.

What do you think RED's base ISO is, and what are you basing that on?

A post by Jim saying it might be 500 instead of 320 (can't find it right now).

 

[Lee Jay]

By the way, I now have an idea as to how RED could use their unique technology to produce enormous *real* dynamic range in still images. Interested Jim?

 

[Daniel Browning]

A post by Jim saying it might be 500 instead of 320 (can't find it right now).

You must have thought Jim was talking about mid-gray being is 3.5 stops below FWC; he wasn't.

 

[tj williams]

why not just take one frame @ 10 ASA and one frame at2000 ASA and shoot 48fps then combine them in post and have twice the dynamic range?

 

[fde101]

A post by Jim saying it might be 500 instead of 320 (can't find it right now).

R1 is natively 320, that's when it's at its best. RED has been saying that it still looks good at 500 -- meaning that it is professionally usable and you need not be afraid to use 500 if called for -- but the native rating is 320.