Knowing that the Arri Alexa has 15 or so stops of dynamic range is nice, but it’s hardly useful information unless you know how those stops are arranged at different ISO settings. Fortunately I’ve done the hard work for you.
Or rather I should say that Adam Wilt and I did the hard work, as we both traveled to Chater Camera in Berkeley recently with a DSC Labs DX-1 17-stop dynamic range chart intent on discovering Alexa’s hidden secrets. (The biggest secret was “Why is this Alexa just sitting here?” as we’d managed to luck upon the only day during a multiweek period when it was actually available for testing.)
Dynamic range charts are a lot of fun if you want to play games like “My camera’s dynamic range is bigger than yours!” but they don’t tell us how to exploit dynamic range in our work. My goal in these tests is to find how that dynamic range is distributed, which is much more useful to me as a cinematographer.
The only solid exposure reference that an HD camera offers is the point where the sensor clips, and the DSC 17-stop chart is designed with that in mind: by exposing the chart such that the first chip just hits white clip we can count how many chips are visible until they blend into the black background. As each chip is one stop, counting the visible chips to the right of white (don’t count the first one for this test) gives us a number for the camera’s full dynamic range in lens stops.
When using a camera “film style” we don’t set the exposure based on clipping, but with an eye to exposing the image properly as a whole. My preference is to use the Zone System and expose certain objects in the scene so that they reproduce with a specific tonality. That can only be done by knowing how a camera or film stock responds to different reflective values of light and using that knowledge in concert with a spot meter. The Zone System hinges around 18% gray, and as there isn’t a high dynamic range chart in existence that is set up to measure exposure in relationship to that middle tone, I set out to make one.
I set the Alexa to record ProRes 4444 to SxS cards in “legal” mode. I set the Alexa’s ISO at Arri’s recommended “sweet spot” of 800 and, while watching a waveform monitor, I set the exposure on the DSC Chart so that the first chip just flattened out into clipping. Then I looked at the waveform monitor and determined which step fell closest to 18% gray (generally 42% to 45% on a waveform), and marked that chip by putting a thin strip of black tape vertically through it. The tape creates a “notch” in the waveform trace and allows us to watch what happens to that middle gray reference as we change the camera’s ISO settings.
Adam and I only tested four ISO settings-1600, 800, 400 and 200-because the chart is calibrated in one stop increments. (We tried a Stouffer wedge chart with 1/3 stop increments on a prior occasion and beyond a certain point the steps became too fine to count in the shadows.) We shot the chart at each ISO setting in both LogC and Rec 709 mode, and then did an exposure test simply to measure overall dynamic range. This resulted in two sets of results: one that showed how many total stops were visible below white clip on the chart, and another that showed how those stops were distributed above and below 18% gray on the waveform monitor. The first tells us how much range the camera has overall at a specific ISO, and the other tells us how much exposure latitude we have before we hit white clip or the noise floor (black).
Turn the page to get started…
Yay! Don’t you love looking at charts? It’s a necessary evil, sadly, as there’s too much unknown detail cluttering the average shot to really understand what a camera is doing. A test chart is a known quantity: we know how it is supposed to work, we know how it should appear, and if something happens that we don’t expect then we can be fairly sure that there’s something going on with the camera because charts-over the short term-don’t change.
Here’s the DSC DX-1 17-stop chart photographed by an Arri Alexa using an 85mm Zeiss Ultra Prime. The stop is T2.3. The chips are placed at the top of the frame following Adam Wilt’s advice concerning flare: placing the bright chips in the center creates flare in the center of the frame that overlaps the chips and affects how they look, whereas framing the chips high, low, left or right throws most of the flare to the opposite side of the frame, away from the chips. In other words, placing the important part of the test chart to one side throws any flare to the OTHER side. (Very clever.)
In the top image you can see the piece of tape that I ran down the surface of chip #7. We can see the effect of that piece of tape on the waveform next to the label that says “Notched Reference.” That chip fell exactly on 45% in LogC mode, so I picked it as my middle gray reference. It floats up to 50% in Rec 709 mode, so in Rec 709 we’ll call 50% middle gray for ease of analysis.
Middle gray stays at nearly the same value when toggling between LogC and Rec 709, with the upshot that the LogC image–which is not designed to be viewable on any kind of HD monitor–still looks okay when viewed on a Rec 709 monitor. (Arri’s FAQ states that an 18% gray card should read at 39% in LogC mode and 38% in Rec 709, but none of my regular technical sources believes that 38% is a proper 18% gray value in Rec 709. Answers vary but range from 42% to 46%. The bottom line is that middle gray changes very little when toggling between LogC and Rec 709, and this seems to make LogC more “monitor friendly” than other log curves.)
There are a couple of things we can discern by looking at the relationships of the different chips to middle gray on the LogC waveform:
Looking at highlight latitude, I see seven stops of exposure latitude before clipping occurs, which is unheard of in an electronic camera. Four or fives stops is typical. This is clearly why Alexa is so good at handling highlights, which have long been HD’s weak spot.
Looking at shadow latitude, I see eight stops of exposure latitude before hitting solid black. This can be seen in the chart itself, where the 15th chip is just barely discernible from black, but whether you’d ever want to expose something important at that value is questionable. My guess is that I’d plan on -5 being the lowest really usable stop, and know that in order to achieve real black I should probably push the exposure more than eight stops below middle gray. Odds are that everything below stop -4 or -5 will be crushed in the grading process, but if you need to hold detail in those tones it’ll probably be there. Just don’t try to pull them up much or you may find more noise in the image than you like.
Log curves are designed for efficient storage, and nothing more. They are not raw, and you are not excused from white balancing when using them. What they do is preserve the maximum amount of usable grading information by storing tonal values according to perceptual steps. If you want to learn more about how Log curves work, read this article. (And stay tuned, as I’ll be talking to Arri shortly about how their LogC technology works.) LogC is a great way to observe what kind of information is really available to the camera, and if you’re headed towards color correction for a broadcast master this is definitely the gamma you’ll want to record in. And do note that LogC is a gamma curve only: the color in the image is unaffected.
The most interesting thing about LogC is that the only time the values compress is when they descend into noise at the right side of the waveform. It is unknown whether this is due to the slope of that part of the log curve or due to the “dog leg” that appears as sensor response drops near the noise floor. Otherwise every stop in the camera’s dynamic range stores the same number of bits as any other, which is seen in the equal steps between most of the chips. This helps to prevent banding during the color grade.
Let’s look at the Rec 709 curve:
This is the same chart, exposed exactly the same way, but with Rec 709 gamma applied. Wow, what a difference. The good news is that this is the Alexa’s what-you-see-is-what-you-get mode, where you can trust a properly calibrated broadcast monitor and vectorscope to show you how the final image will look without grading. It’s shocking, though, to see what you give up at the extreme ends of exposure.
The important thing to remember about Rec 709 is that it was created at a time when cameras could deliver only about five stops of dynamic range–so that’s all Rec 709 is designed to hold. And it’s rare to find a modern camera with a true Rec 709 curve in it, because there’s not a single camera out there that only delivers five stops of dynamic range. Instead, manufacturers create curves that look good in a Rec 709 environment, in spite of the fact that the cameras capture more dynamic range than Rec 709 is supposed to hold.
The trick has become cramming 10+ stops of dynamic range into a small five stop “bit bucket” in such a way that it still looks good. As you’ve probably noticed from the image above, Alexa does this by applying quite a strong S-curve to the gamma, stretching out the mid-tones so they are nice and contrasty while compressing the highlights and shadows much the same way film does.
If you’re curious as to what happens when the mid-tones aren’t emphasized, look at the log curve above. Cramming 15 stops into a five stop “bit bucket” results in a very flat-looking image: whereas five stops have plenty of room to stretch out, fifteen stops end up crammed closely together and the contrast between each step is reduced. Stretching out the mid-tones, which is where most of the important visual information is, and compressing the highlights and shadows, where our eyes naturally discard detail, is often the best way to shoehorn all that latitude into a small “bit bucket.”
By analyzing the Rec 709 curve and comparing it to LogC, my guess is that stop -5 is about as low as you’d want to expose something that should have some presence in the frame. Looking at the chart itself, chip #12 is the last one I can clearly see, and that corresponds to -5 on the waveform. Stops -6, -7 and -8 have some presence but they’ve already been compressed to the point where pulling any detail out of them in post is probably impossible. They are so close together that there won’t be much contrast between those values and, while they’ll be visible, they may blend together to the point where image detail in those shadows becomes indistinct due to lack of contrast.
What’s interesting is that the LogC and Rec 709 curves look similar between step -5 and black. The black levels are different between the two curves but the lowest stops show about the same amount of distance above black.
The real difference between the curves is in the ability to access highlight detail in post: Rec 709 rolls off and compresses highlights the way film would, but LogC better preserves the differences between the tones and allows a colorist to massage them more easily later. We can see these differences in the charts: LogC shows the brightness steps at the top of the scale very clearly, but Rec 709 shows the difference between steps +5 and +6 (chart steps 1 and 2) as the last discernible step, with the difference between steps +6 and +7 (chart steps 0 and 1) being too small to see.
One interesting thing is that while the Rec 709 curve fills the entire “bit bucket” from 0% to 100%, LogC only extends from 0% to 95%. In extended mode, where the bits from 100-109% are made available, LogC still only rises to a maximum value of 104%. During our tests Adam and I discovered that there’s only one ISO setting that uses the full range of the “bit bucket,” but you’ll have to wait a bit before we reveal which one.
Here’s a better look at how many stops are available overall in ISO 800:
The vertical charts are recorded in LogC. Chip #15 is just barely visible as brighter than the black dividing line between sections.
Arri has asked me to clearly state that their factory considers the camera to have a maximum dynamic range of 14 stops. Based on what I see in this chart that measurement is slightly conservative as the 15th stop is clearly visible on the chart, but it’s so faint it’s likely not usable. This is a case where practicality wins over marketing, as marketing dictates shouting “15-stop dynamic range!” from the rooftops while practicality recognizes this 15th stop isn’t something that should be relied on.
In a previous article I noted the difference between what I call “paycheck” stops and “gravy” stops:
Paycheck stops are values that you can bet your paycheck on. They’ll appear in the image with enough contrast and detail that objects exposed at that value will be easily discernible.
Gravy stops are the values that lie beyond paycheck stops: they are the tonalities at either end of the dynamic range scale where some presence will be seen, along with some discernible detail, but they shouldn’t be trusted. If there’s some information held in those tones then that’s great, but don’t count on it.
That 15th stop is definitely a gravy stop. Don’t bet your paycheck on it.
NOTE: In some of these chart images you might see some magenta or green fringing around highlights. That appears to be longitudinal chromatic aberration, as changing lens focus caused the colors to shift from green to magenta. This means the color shifts are lens related, not camera related. (The vertical charts were shot on a 24mm Ultra Prime, and I’ve zoomed in to the image to make it visible. Why a 24mm Ultra Prime? Because the 85mm was too long and all the other Ultra Primes were out on rental that day.)
Incidentally, there’s no spectrum of light that’s magenta-so what we’re really seeing is that a portion of the image has either too much green (which appears green) or too little green (which appears magenta).
Turn the page to see what happens when we deviate from Arri’s recommended ISO “sweet spot” of 800…
Behold! I bring you ISO 400:
This is where things start to get interesting.
I haven’t changed the T-stop on the lens or the lighting on the chart. Notice on the waveform that the notched chip, which used to be middle gray, is now one stop darker than middle gray (at -1). Chip #6, in the image of the chart above, is now middle gray, a distinction originally held by chip #7. Changing the ISO from 800 to 400 costs one stop of overexposure latitude and shifts all the steps down one toward black.
Making all the stops darker means there should be one more stop in the shadows, but it’s way down there just past -8 and it’s barely visible. What I do see is that noise decreases noticeably:
Noise shows up in a waveform trace as a fuzzy line, and by comparing these two images of the darkest part of the frame we can see that the toe of the ISO 800 trace is fuzzier than the toe of the ISO 400 trace. Not only that, but the black level itself is a couple of units lower. Switching to ISO 400 cost us a stop of overexposure latitude but gave us cleaner and crisper blacks.
Let’s look at the Rec 709 results:
This is a bit more interesting. What I’ve noticed is that step +5 appears in roughly the same position in relation to white clip at both ISO 800 and ISO 400:
Although there’s another stop of overexposure latitude under ISO/EI 800, the distance of stop +5 from clipping is about the same. In fact, when I go back and look at the ISO 800 Rec 709 chart, I can’t tell the difference between stops +6 and +7-and I can tell the difference between +5 and +6 at both ISO 800 and ISO 400. What this seems to imply is that, at ISO 800 or 400 in Rec 709 mode, stop +5 is the brightest stop that will hold significant detail. At ISO 800, stops +6 and +7 are so close together that they will most likely appear the same, and at ISO 400 stop +6 is the clipping point.
It appears that ISO 400 might offer cleaner blacks and less noise, with hardly any compromise in the highlights, when shooting Rec 709. That’s very interesting to keep in mind. But, at the same time, it’s worth noting that maximum white in LogC is down to about 90% (or, in extended mode, just under 100%). As the ISO decreases our “bit bucket” is getting smaller. Will that make a difference in the grade? Most likely not… but it’s interesting to note.
Let’s proceed down the ISO scale to 200:
We’ve now lost another stop of overexposure latitude, and our notched reference is down two stops from where it started at ISO 800. More on that in a moment. First, let’s look at noise levels:
Once again, lowering the ISO gives us crisper and cleaner blacks, as indicated by the ISO 200 trace looking less fuzzy than the ISO 400 trace and the black level sinking a bit lower.
Once again, the “bit bucket” is getting smaller, as maximum white only reaches about 82% on the waveform (85% or so in extended mode). Does it matter that LogC doesn’t allow the brightest tones to reach white in this mode? Not really. The white level will be set in post, requiring that the entire range be stretched out a bit, but 10-bit 4:2:2 or 4:4:4 material should be able to handle that.
Let’s look at Rec 709:
Whereas stop +5 had been somewhat consistent in distance from white clip and separation from stop +4 between ISO 800 and ISO 400, stop +5 is now just plain white, with no detail beyond it. Let’s compare highlight steps between ISO 400 and ISO 200:
In ISO 200, step +4 is about where step +5 was before. It’s interesting to see that the chart shows very clear separation between stops +4 and +5, whereas other ISO’s under Rec 709 crushed the last couple of highlight steps together to the point where stops beyond +5 were almost meaningless. The good news is that highlights and brighter tones will have more contrast and appear snappier as the steps above middle gray are pulled farther apart, resulting in more contrast.
ISO 200 seems like a perfectly viable option in situations where low noise and less than five stops of overexposure latitude are acceptable. One situation that screams out for this configuration is green screen.
Let’s take a look at two vertical charts before we move on to our last ISO test. Here’s ISO 400:
And here’s ISO 200:
We can see by eye how the steps disappear off the dark part of the chart as the ISO decreases.
Let’s move on to ISO 1600, on the last page, for a final surprise…
Feast your eyes upon ISO 1600:
This is the only LogC ISO that takes full use of the “bit bucket.” White is at 100% (or 109% in extended mode). This is also the best ISO for protecting highlights, with an astounding eight stops of overexposure latitude. The price, however, is more noise:
The increased fuzziness of these lower steps shows that there’s a bit more noise in the shadows than at ISO 800, which is to be expected. It’s worth testing to see whether your own taste precludes or embraces this much “texture.”
Once again, Arri’s Rec 709 curve puts stop +5 in a familiar place:
While stop +6 is a bit more differentiated, stop +5 seems consistent from ISO 1600 to ISO 400 in distance from the white clip. This is a good thing to know, and could very much simplify working with Alexa. For example, for ISO’s 1600, 800 and 400 in Rec 709 mode, it’s safe to assume that we’re holding detail (or as Ansel Adams might have said, “preserving textured white”) at +5 stops over middle gray, and anything that’s brighter than that and still contains detail is a nice bonus–but don’t count on it.
Here’s the ISO 1600 vertical chart in LogC:
Based on all the information I collected from this test, I’ve created some tables that may help the first time you use the Alexa. The numbers on the charts are stops above and below 18% (middle) gray and are intended to be judged with a spot meter. USE AT YOUR OWN RISK. I take no responsibility for your results unless they are excellent.
I should note some definitions:
“White Clip” means the sensor is saturated, or clipped. There is no detail retained at or above this point. Every tone at or above this point will be a featureless white.
“Textured White” is the brightest white you can expose and still retain enough contrast to see visible texture. I’ve determined this value by looking at compression in the brightest chips in Rec 709 mode and finding the one that has a step that is really distinct, by a few waveform units, from the next one above it. This really only counts in Rec 709 mode where you’re baking in a look. You’ll still have additional stops of exposure (subtract the “textured white” number from the “white clip” number to find out how many) but these tones will be very close together in value and may not resolve detail well due to the low contrast between them.
In LogC you’ll retain texture right up to the clipping point while shooting, although it’ll be up to your colorist to retain and preserve highlight detail during the grade.
“Textured Black” is the darkest tone in which you should retain good tonal separation and detail. At ISO’s 800 and 1600 I’ve been a bit conservative as to which tone I define as textured black: I’m looking not just at the distance of the waveform step above black but also at how fuzzy/noisy it is. If a step’s fuzziness causes its waveform trace to merge with the noise floor (black) then I don’t consider it to be a solid bet for textured black and suggest pushing exposure one stop up the scale. There may still be visible detail below this point, but consider it a bonus.
“Noise Floor (Black)” means that tones at or below this level will be lost in the noise floor, or black. Typically LogC tones in this range get crushed during the grade to eliminate noise.
Please note that there are two tables on this chart: one for Rec 709 and one for LogC. Pick the right one for the mode you’re working in. Click here to download a PDF of the chart.
I hope you’ve enjoyed this technical look into the Arri Alexa. It’s a wonderful camera to work with, and I do worry that it’ll be the final nail in film’s coffin. I know people have been saying things like that for years, but… finally an HD camera is that good.
Disclosure: ARRI gave me a 320gb hard drive, for free, after my Alexa prototype shoot, because the Firewire 400 drive I brought for file storage was ridiculously slow.
Art Adams is a DP who stops at nothing, either above or below 18% gray. His website is at www.artadamsdp.com.