DR ..a good discussion!!

[Anton Esteban]

Luigi thanks again for doing these tests, extremely helpful. You are doing everyone on Reduser such an amazing service. Hats off.

I have a question about the Venice. How much heavier/bulkier is it then the Monstro? Do you think you could mount a Venice on a MoVi Pro if you powered it from another source? Or has anyone on Reduser figured this out? It says its 8lbs completely bare. Tried a few searches online and not much came up.

I've been playing with your clips and I'm finding the Monstro definitely has the most versatile Dynamic Range to play with in post but I'm really loving the color science of the Venice. It just feels so much more cinematic to me and somehow has more perceivable depth right off the bat. I can get the Monstro footage to get there but it seems to take a lot more work.

 

[luigivaltulini]

Luigi thanks again for doing these tests, extremely helpful. You are doing everyone on Reduser such an amazing service. Hats off.

I have a question about the Venice. How much heavier/bulkier is it then the Monstro? Do you think you could mount a Venice on a MoVi Pro if you powered it from another source? Or has anyone on Reduser figured this out? It says its 8lbs completely bare. Tried a few searches online and not much came up.

I've been playing with your clips and I'm finding the Monstro definitely has the most versatile Dynamic Range to play with in post but I'm really loving the color science of the Venice. It just feels so much more cinematic to me and somehow has more perceivable depth right off the bat. I can get the Monstro footage to get there but it seems to take a lot more work.

Ciao Anton,
it weighs a lot, the camera body is beautiful robust. I also had the RAW recorder. What I can tell you is that it weighs + - a few grams to an ALexa XT / Ev machine body.
Another important thing that I did not write the fans is noisier than the RED ones. But I think it's due to the Raw recorder, I did not try to take it off.
However there are several ways to set up the fan similar to RED.
For color science, of course that depends on each of us, who prefers ALexa who RED who Sony if you feel good it is right that you go in your direction

 

[luigivaltulini]

It is also curious that Panavision rates the DXL2 a 1600 native ISO. They probably feel confortable with the amount of noise at this ISO while also gaining 1stop in the highlights compared to Red's recommended ISO 800 (though they do not state anywhere I could find what they consider the native ISO).

Fabricio.
Panavison knows what he wants.
At 1600 ISO you have the same dynamic range and the same behavior in the highlights of an Alexa with a little less noise given the resolution.
Dlx2 is declared 16 stops, as we can see from the xyla charts made here on Monstro and everything coincides.
What interests me about Panavision is the science of color, I would be happy if I sold the M form to the public.

 

[Anton Esteban]

Luigi, thanks again for you quick reply. Wow, Venice seems like a heavy brick even though it actually doesn't look so huge. I'm sure its sturdy as hell but I think my dreams of fitting it on a MoVi pro are fantasy at best. In the end, nothing beats RED in terms of its small size and light weight. It's amazing how RED has fit in so much tech in such a tiny camera body.

 

[Alexander Nikishin]

Is anyone else noticing how the Gemini natively in STD or LL mode starts out with "milky" blacks similar to the low con look of an Alexa? Leads me to question whether or not there really is more range in the shadows.

 

[Michael Tiemann]

Is anyone else noticing how the Gemini natively in STD or LL mode starts out with "milky" blacks similar to the low con look of an Alexa? Leads me to question whether or not there really is more range in the shadows.

The chart gives more information about this. First off, the Gemini tests benefited from an exposure that was about half a stop greater than other contestants. You can see this because the distance between the clipped chip on the leftmost side and the highlight chip next to it is the smallest of any test. If Luigi (or anybody else) shot the Xyla charts again, I would recommend that one should shoot a range of exposures, not simply trusting the stoplights to tell you when the camera is clipping. The stoplights don't come on until a reasonable fraction of photosites are clipping. What is that reasonable fraction? Could well depend on the camera. Shooting when they first come on, and then shooting at +1/4 stop, +1/2 stop, +3/4 stop, and +1 stop means you will capture a chart that really, really has a decisive clip, and really, really has a maximum but not clipped second chip.

The second thing the Gemini charts show is that Gemini has lots of noise in the blacks (especially blue noise). It is much more prominent than any of the other cameras. If you ramp up the Gain (blowing out highlights--who cares?!) you will see the hair on the blacks in Gemini much sooner than you do on the others. You can read that noise as if it were lifting the blacks, because ON AVERAGE that's what it's doing. But at the pixel level, there are plenty of deep blacks mixed in with many not-so-black pixels. So, technically, not lifting the blacks.

 

[rand thompson]

Regrade of Monstro

[/url]A001_C026_0115C6.0000043F by rand thompson, on Flickr[/IMG]

[/url]Studio 4.3 by rand thompson, on Flickr[/IMG]

 

[Brian Boyer]

I think Red chose the right approach, Focusing more on Lowlight and less so on Highlights. You can use a ND Filter or more diffused lighting to hold on to more highlight info., but being able to pull up less noisy Shadows that keep more info. is something more desirable IMHO.

Everyone has their opinion on this but I think focusing solely on low light performance is a mistake, of sorts (granted, they were commissioned to do so). So many people are enamored with the idea of lighting a scene with three fireflies and having it look like they pumped 3,000 watts in the room.

I'm not saying a sensor like Gemini isn't an impressive technical achievement or shouldn't exist in their lineup. I just think the history of motion picture imagery supports the case for highlights much more strongly than a case for shadows.

Motion picture film was not, and still isn't, a stellar low light performer and yet it became the "gold standard" for narrative storytelling over video (its contemporary) for decades. What it had going for it was highlight latitude, color and resolution.

Also, 24fps motion, but that's not relevant to this conversation since digital video also uses that frame rate.

We've been psychologically conditioned to subjectively and (mostly) unconsciously associate the superior highlight performance of film, which comes closer to how our eyes see bright things, with a better quality image.

Blown out skies, windows, etc. are a relatively recent development in human visual experience, at least going back to the emergence of homo sapiens, I'm surmising. Unless we stare directly at the sun, normal vision allows us to look at any sky without losing detail in the highlights.

How many naturalistic paintings do you know of with blown out skies? I'm not talking overcast skies but vast white areas of canvas with no gradations. People painted them the way they saw them. It wasn't until we developed mediums that couldn't quite capture the range that we started to see those artifacts.

More highlight detail = better image to our brains.

In contrast, there's no equivalent association with shadow detail, given that shadow detail was usually missing unless you expressly lit for it because of film's relatively poor light sensitivity.

From an audience's perspective, other than the appeal of a well composed and artfully lit low light shot, there's nothing that psychologically signals quality in the shadows, which are rarely the focus.

Humans are very tolerant of noise and are many times drawn to its seemingly chaotic and organic nature.

From a technical standpoint, we know a certain type of picture quality exists with high sensitivity. However, with a few notable exceptions, the viewer can't tell if a scene was shot in substantial lighting with Gemini in STD mode or in very low light in LL mode.

Both scenarios would simply produce what appears to be an adequately lit shot.

Unless you told the viewers or they were on set during recording, they wouldn't know how dark it was and wouldn't be able to tell the difference between shooting situations in most cases. It benefits the production much more so than the viewer.

In addition, if low light performance is what the film and tv industries most desired, Alexa would've been supplanted years ago, as there are better low light performers. So what accounts for Alexa's and film's continued appeal?

Several things can be argued but I think the most overlooked is the beauty of and psychological attachment we have to being able to see into the brightness the way we do. The sensors that come closer to that in the way of greater saturation capacity will be on to something.

 

[David J. Buchanan]

It's not always about highlight retention as much as it's about the gradient between dark and light.

In painting, chiaroscuro. Most video cameras just don't have the latitude or accuracy to replicate it.

 

[Michael Tiemann]

Let's again look at what this really means (which, again, means referring to what Ansel Adams teaches in his series, The Camera, The Negative, and The Print).

[...]

So the first thing to keep in mind is that whatever the sensor's DR, the top stop and the bottom stop have to be tied to the display's top stop and bottom stop.

Now, in between those two we can look at how well the sensor's DR maps to the display's DR, and why not all stops are created equal. When looking way to the right of the XYLA chart luma values (preferably in a waveform monitor), we see that there's very little (or no room) for meaningful intermediate values to exist. I previously talked about digital "eye" charts, which show (or don't show) the decisive space between reading something as a 1 or reading something as a zero. If the noise of the 1 signal overlaps the noise of the 0 signal, there is no "eye", which means that there's no real discrimination possible between signal and noise. Of course one can draw a digital line in the sand and say "everthing above this line is a 1 and everything below this line is a zero", but that means that up to half the time we'll be rendering a 1 as a 0, and vice-versa. And all those guesses will lead to a dithering of 1s and 0s, creating a kind of digital gray. That gray will be 50% gray if the 1s and 0s are full scale, and it will be a much darker gray if the 1s and 0s map to the bottom limit of the DR of the display device.

If we pick a chip where the eye is 6x the height of the 1 noise and the 0 noise, then there are 6 distinct gradations of values that can be teased from the sensor between the 1 value and the 0 value, 8 levels if we include the values of the 1 and the 0. That's 3-bit color! Now, not all part of the image need to be rendered in fully expressed 10-bit color gradients to look good. But if the blockiness of 3-bit color is going to be a problem for a given part of the tonal range, then even if we have the DR to distinguish one stop from another, the image rendered with that stop can fail subjective tests (blockiness in the shadows). So we must ask: how much do we have to crush the shadows before the limitations of gradations disappear? Alternatively, what chip of the XYLA chart gives us an eye that's large enough to hold enough gradations to render acceptably given the noise measured at the top and bottom of those scales?

The same calculations can be done in the highlights: when does blockiness in the highlights become a problem, and how large is the eye (vs the noise) large enough at the top level to deal with that?

[...]

TL;DR:

It's not always about highlight retention as much as it's about the gradient between dark and light.

In painting, chiaroscuro. Most video cameras just don't have the latitude or accuracy to replicate it.

Beautiful word and an important concept!

 

[Brian Boyer]

It's not always about highlight retention as much as it's about the gradient between dark and light.

In painting, chiaroscuro.

Well, it's never always about any one thing. I wasn't suggesting it's always about the highlights. Dark and light are relative terms. You can maintain the gradient between the two while shifting luminance 10 stops brighter.

The light levels under the shade of a tree that constitute the darkest parts of an image on a bright, sunny day can be the same levels that seem like really bright areas during unshaded evening hours. Our sun looks pretty dim in comparison to stars that are much brighter.

Of course, with the exception of the sun, such extreme light intensities aren't relevant to our practical, day-to-day shooting situations. All I'm asking for is the gradient range from light to dark we've had with film for so long. Let's not throw the baby out with the bathwater. We should strive to achieve that performance with digital as well.

Most video cameras just don't have the latitude or accuracy to replicate it.

Yet. But they will. And when they do, everyone will sing their praises. And it doesn't have to be the entire range from where film clips on the high end to where Gemini sees into the shadows.

For me, it would be cool if RED and everyone else eased up on the "I shot this scene at 11pm with a single LED street light from half a block away and it looks like midday" low light chasing and put more effort into sensor saturation capacity.

Folks have Gemini and A7S and whatever else is out there with fantastic low light specs from which to currently choose.

 

[Brian Boyer]

Beautiful word and an important concept!

Michael, being a knowledgeable person who's comfortable with math, please have a look at a post of mine where I pose some questions about the nature of dynamic range at the top end of light intensity and whether more could be obtained within the current DR limits of around 15 or 16 total stops, if not even less.

http://www.reduser.net/forum/showth...onger-on-Red&p=1811385&viewfull=1#post1811385

 

[Alexander Nikishin]

The chart gives more information about this. First off, the Gemini tests benefited from an exposure that was about half a stop greater than other contestants. You can see this because the distance between the clipped chip on the leftmost side and the highlight chip next to it is the smallest of any test. If Luigi (or anybody else) shot the Xyla charts again, I would recommend that one should shoot a range of exposures, not simply trusting the stoplights to tell you when the camera is clipping. The stoplights don't come on until a reasonable fraction of photosites are clipping. What is that reasonable fraction? Could well depend on the camera. Shooting when they first come on, and then shooting at +1/4 stop, +1/2 stop, +3/4 stop, and +1 stop means you will capture a chart that really, really has a decisive clip, and really, really has a maximum but not clipped second chip.

The second thing the Gemini charts show is that Gemini has lots of noise in the blacks (especially blue noise). It is much more prominent than any of the other cameras. If you ramp up the Gain (blowing out highlights--who cares?!) you will see the hair on the blacks in Gemini much sooner than you do on the others. You can read that noise as if it were lifting the blacks, because ON AVERAGE that's what it's doing. But at the pixel level, there are plenty of deep blacks mixed in with many not-so-black pixels. So, technically, not lifting the blacks.

But not having deep blacks is the same as...well...not having blacks at all imo. I just see milky/grey shadows and no micro contrast in daylight scenes so far. Not all too impressed as everyone else is with Gemini just yet. Would love to do some real testing.

 

[Michael Tiemann]

Michael, being a knowledgeable person who's comfortable with math, please have a look at a post of mine where I pose some questions about the nature of dynamic range at the top end of light intensity and whether more could be obtained within the current DR limits of around 15 or 16 total stops, if not even less.

http://www.reduser.net/forum/showth...onger-on-Red&p=1811385&viewfull=1#post1811385

My reading of this post is that you both are saying the same thing textually, but that you seem to think that saturation capacity is some kind of magic bullet, which it is not. To pick this apart in greater detail:

I was asking is there mathematically a minimum total DR a sensor has to have in order to accommodate a light intensity of, say, 9 stops of exposure above middle gray before clipping? Or...

I cannot make any meaning out of this because it's not clear how many stops are supposed to be below middle gray, with what kind of noise criteria. Once you define that half of the equation, call it X, then the answer to your question is 9+X.

Can a sensor accommodate that same light intensity before clipping in as few as 13 stops, suggesting it has a high saturation point but also a relatively high noise floor as well?

If 9+X = 13, then X = 4, which means you have 4 stops of dynamic range to represent the chips below middle gray.

Let's say we have a sensor that has 9 stops of DR total, with enough saturation capacity to allow, at best, 4 stops of overexposure and a subsequent noise floor that allows 5 stops of underexposure, in keeping with the known sensitivities of the various ASA/ISO ratings.

Saturation capacity is not a magic bullet. It simply states where the sensor clips highlights. Definitionally, if you have 9 stops of DR (counting from zero), there exists an exposure level that's 4 stops below clipping and 5 stops above the noise floor. You can call that middle gray. When that middle gray is over-exposed by 4 stops, it clips to white, and when it's underexposed by 5 stops, it fades into the noise that's black. And what none of this tells you (yet) is the extent to which quantization gives you (or does not give you) latitude for getting nice gradations within any given range of stops.

Increasing the saturation capacity of that sensor's next gen replacement by the light intensity equivalent of 4 stops while not improving its noise specs yields a sensor with more DR on top and more total DR, but still only 13 stops total.

That's correct. Mathematically, if we can add more saturation capacity without affecting noise at the bottom (a big IF!), then mathematically, the higher stops do give us greater gradations in the highlights because those larger numbers of electrons can be quantized much more smoothly. If our previous top was 100,000 electrons and we increase the bucket size four stops to 1,600,000, then we have lots of gradation between 100,000 and 101,000 and 102,000 and 103,000 and ... and 1,598,000 and 1,599,000. Compare that to the bottom where our noise floor is at 195 (100,000 / 512), so say black = 200, our first stop above that is 400, and our second stop above that is 800. That's three stops of DR with a difference of fewer than 1,000 electrons. There are 300 1,000-electron gradations between 100,000 and 400,000, which can lead to much smoother gradations of color there.

As you know, manufacturers are increasing overall DR but they're doing it mainly by lowering the noise floor, not significantly increasing saturation capacity.

Saturation capacity is not a magic bullet. It's true that opening up more stops at the top gives you math that makes it look like you have much smoother color gradations at the top, but that's not necessarily where they are needed (with everything fading to white). Lowering the noise floor brings lots more color gradations into the middle tones, which is where the real money is.

The "DR we currently have" falls between 14-16 stops for the top rated cinema cameras. Do we need more than 15 total stops to get substantially higher saturation capacity (highlight latitude) or is that number sufficient?

Latitude comes out of your DR budget. Saturation capacity tells you where you clip whites. Once you nail down where your whites clip, DR tells you where your blacks become noise. There is no magic.

That's what I was asking. I hope it's clearer now. From your answers, I gather higher saturation capable sensors can conceivably exist within the 15 or 16 stops of DR we currently have available. What say you?

MIT is developing a sensor with much larger dynamic range by essentially counting how many times saturation capacity is reached during an exposure. PDF paper here.

BTW, I'm not advocating for a noisier sensor, I'm just willing to accept a little more noise if it were the trade-off for gaining substantially better highlight latitude in the same number of total stops that are achievable today.

To me, that ask is meaningless. Total DR = top highlight - noise stop. If you want better highlight latitude for a given DR, you have to accept more noise, in a 1:1 equation. 4 more stops of highlight latitude will cost you four stops of noise.

 

[Mike P.]

Haven't read this thread so forgive me if I'm beating a dead horse, but I found Cioni's/Panavision's perspective on the DR of Montro to be reasonable (and something I think most of us have expected for quite some time).

Basically says, 'the sensor may see ~20stops, but we have found "Sweet 16" to be the most useable; any more than that and you're trying to pull stops out of the weeds [noise]...'

Maybe with Gemini, because it's so clean, the DR seems more apparent/easy to play with, but I'm pretty sure from Dragon onwards we've been dealing with ~15-16 useable stops (with the VV sensors having an edge because the downscale and/or pixel pitch makes noise less noticeable/destructive).

 

[Jon Dishler]

One thing I have noticed with the Helium sensor (and I think others) is that it clips green first, so even if there are brightness "steps" beyond the green clip point in total luminosity, they are really not usable stops since all 3 colors are not correctly represented. So at the high end one has to be careful where you start counting. In the shadows, conversely, I have noticed that in really dark shadows there is very little noise, black is black, but it is in the darker shadows below middle grey where at higher ISO one sees color noise, most apparent with 100% zoom. When the R3D is reduced to ProRes in camera (or post), especially from 8k to 2k, the shadow noise is much diminished with the color noise component virtually gone. This is to say that there are not only these noisy shadow stops, but also partially blown out highlight stops that should not be counted as well.

 

[Mike P.]

One thing I have noticed with the Helium sensor (and I think others) is that it clips green first, so even if there are brightness "steps" beyond the green clip point in total luminosity, they are really not usable stops since all 3 colors are not correctly represented. So at the high end one has to be careful where you start counting. In the shadows, conversely, I have noticed that in really dark shadows there is very little noise, black is black, but it is in the darker shadows below middle grey where at higher ISO one sees color noise, most apparent with 100% zoom. When the R3D is reduced to ProRes in camera (or post), especially from 8k to 2k, the shadow noise is much diminished with the color noise component virtually gone. This is to say that there are not only these noisy shadow stops, but also partially blown out highlight stops that should not be counted as well.

I think IPP2 handles that highlight behaviour really well actually. I believe it desaturates the one or two non-clipped channels to a more neutral tone, and then uses that information as very faint detail/texture in what would otherwise be completely blown out highlight.

And yes, that is exactly what's happening at the low-end (re: downscale, how comfortable you are with grain, noise reduction all play factors in deciding what one considers usable.)

 

[Anton Shavlik]

A good magenta filter, like CC40M, gets you a little more out of the sensor by overcoming the fact that green clips first. That's probably true on most bayer sensors, and I'm sure to balance the CFA you'd have to lose a stop overall and that's no good.

It's also easy to see in the Xyla R3Ds posted that if you set gemini to 3200k +0 tint you get blue noise in your blacks. But set the white balance to 5600k and about +20 green and you get a much nicer looking noise and noise floor, even if the chips are now colored.

In my testing on helium the STD OLPF was 5600k while the STH OLPF was more like 4200k, but both need about a stop reduction in the green channel, which seems to be compensated by about +20green.

 

[David J. Buchanan]

A good magenta filter, like CC40M, gets you a little more out of the sensor by overcoming the fact that green clips first. That's probably true on most bayer sensors, and I'm sure to balance the CFA you'd have to lose a stop overall and that's no good.

It's also easy to see in the Xyla R3Ds posted that if you set gemini to 3200k +0 tint you get blue noise in your blacks. But set the white balance to 5600k and about +20 green and you get a much nicer looking noise and noise floor, even if the chips are now colored.

In my testing on helium the STD OLPF was 5600k while the STH OLPF was more like 4200k, but both need about a stop reduction in the green channel, which seems to be compensated by about +20green.

That's doing too much.

 

[rand thompson]