Wally Pfister talks about RED

[Kemalettin Sert]

http://www.youtube.com/watch?v=NRvnj0tCmWk

Resolution Fetishists

 

[Brian Timmons]

I was bracing for a rather surly comment but that was a relatively down to earth response.

Brian Timmons
britim-media

 

[Shawn Nelson]

I really wonder about when people compare film to resolution.

Jim says 35mm film at best is 3.2k (if I recall correctly, it's below 4k)

IMAX is just over 3x as wide as 35mm film http://en.wikipedia.org/wiki/File:Imaxcomparison.png

So IMAX should be about 10k, while Wally says "16 or 18k". For that to be true, 35mm film would have to be around 6k actual resolution, which I've heard from others. Is there any published whitepaper or independently verified test which actually measures the resolution recorded in modern film stock?

 

[David Mullen ASC]

I think there is some confusion over optimal scanning resolution for film versus measurable resolution. Film should be scanned at a higher resolution than it records or else you get into aliasing and other issues, hence why 4-6K is optimal for 35mm film that resolves more like 3-4K, 18K would be optimal for something like IMAX that resolves more like 9-12K.

 

[Sanjin Jukic]

Actually he said that RED1 resolution of 4K is a joke to compare with 16K (!!!) of IMAX.

Also I can't get where he found those numbers just because there is no any 16K film scanner available :laugh: .

He is really very good cinematographer but also wants to be like a 16K big guy in terms of resolution :laugh: .

Btw, I did't like Inception because it has sort of "abuse" of dreams in filmmaking and can't compare with

best European tradition (Bunuel, Fellini, Tarkovsky...) or even the latest Asian use of dreams in movies

by Apichatpong Weerasethakul in "Uncle Boonmee Who Can Recall His Past Lives" (Cannes Palme d'Or 2010).

It looks pretty cheap and enforced in overload of dreams just to get a sort of action drama but...

 

[Graeme Nattress]

David, that sounds about right. Film responds better to being scanned at a higher resolution than the level of detail the film itself contains. Somewhere over 3k is about the level of detail in 35mm - similar to that of a RED One, so around 9k for Imax (scanning higher to retrieve that detail) again would be reasonable. Although when we talk about 3.2k measured RED One resolution, that's system resolution including lens and R3D recording.

Always remember though that resolution is necessary, but not sufficient. Other image qualities are important too, but that doesn't mean resolution isn't necessary.

Graeme

 

[Leo Sardello]

He doesn’t really “talk” about RED, he just tosses a bit out about IMAX film having higher resolution than a RED one. The other parts of the interview are somewhat interesting though. I do think it’s a telling blurb that he say’s “…at a time when everyone is shooting on RED Cameras…”, at least he admits it. I mean geez, if I had the kind of backing these guys have I can’t say I wouldn’t try to shoot some IMAX footage too.

 

[Graeme Nattress]

I bet it's a hell of a lot of fun to shoot on Imax, and the good Imax movies I've seen in real Imax cinemas have been most impressive and immersive indeed.

I'm happy that some people don't feel that they want to be limited by "good enough" and embrace their own ideals and make movies to their specifications.

Graeme

 

[M Most]

I'm happy that some people don't feel that they want to be limited by "good enough" and embrace their own ideals and make movies to their specifications.

It's a lot easier to "embrace your own ideals and make movies to your specifications" when you have an amazing track record and a blank check.

The desire to do something and the practical and financial ability to do it are usually two completely separate things.

 

[Graeme Nattress]

Indeed

But limited practical and financial considerations can cause some highly creative solutions to problems to come about, and that's a positive outcome too.

Graeme

 

[Gunleik Groven]

I know it is silly to use bandwith on, but I really did LOL when I read the two last comments on this thread...

 

[Leo Sardello]

I bet it's a hell of a lot of fun to shoot on Imax, and the good Imax movies I've seen in real Imax cinemas have been most impressive and immersive indeed.

Graeme

Yeah, it’s amazing to me that a format that blew me away when I was a kid can still hold that effect so many years later.


It’s also amazing to me that IMAX was willing to tarnish their name by prompting so many LIEMAX theaters. First time I was in one of those I had to leave and ask for my money back out of principle.

 

[Matt W.]

Is there any published whitepaper or independently verified test which actually measures the resolution recorded in modern film stock?

Kind of! And the numbers are readily available on the spec sheets of any film (see the modulation transfer function chart on the right of page three):

http://motion.kodak.com/motion/uploadedFiles/TI2647.pdf

The mtf chart indicates that 5219 approaches 100% mtf (full-on indisputable detail) to about 20 cycles/mm. Each cycle is two pixels and so we get 40 pixels a millimeter. So for super 35, 24mm * 13mm * 40 pixels horizontally * 40 pixels vertically = about 500,000 pixels (about 650,000 for anamorphic). Not so good. But these pixels are sharper than sharp.

The mtf chart indicates that 5219 hits 30% mtf (where fuzzy detail becomes pure fuzz) at 80 cycles/mm. Same math again gives 8,000,000 pixels (about 10,500,000 for anamorphic). Grain is way smaller than this but doesn't contribute much to the image beyond this point. At this point it's really, really mushy with almost no image at all.

So the answer is that 5219 has "somewhere between 0.5 megapixels and 8 megapixels," assuming a great lens. IMAX is then somewhere between 5 and 85 megapixels.

But then the issue is: what's the mtf curve of digital? Is 4k digital 11 good megapixels (it's not) or 11 fuzzy ones? If fuzzy then how fuzzy? Hard to say, few manufacturers release this information. What's known (or at least oft-repeated), though, is that digital mtf travels along at about 100% to nearly extinction (at about 60-70% of the stated linear resolution in the case of a Bayer sensor, slightly higher in a 3 ccd or striped design), rather than fading out slowly, like film. The 5D and 7D reach well over 100% and then drop to zero due to their weird design and sampling, hence mega-aliasing. What's this mean? It means with digital you can get a lot of "fake" resolution (i.e. aliasing) but also a seemingly much sharper image. Red uses pretty aggressive filtering to avoid aliasing, commendably, but it also results in a 4k that sometimes looks softer than 1080p. Unless you apply sharpening, re-introducing faking resolution.

So the real answer is: Ks, ps, megapixels, etc. are all poor metrics of resolution. The integral of the mtf chart times the surface area of the sensor seems like a better metric, but even then perceptual sharpness is mostly related to medium-frequency detail, meaning digital in theory should have the perceptual edge, but then film grain contributes to perceived sharpness, too, so....

?

And that's without introducing lenses to the equation. Or the mtf of the scanner.

 

[David Mullen ASC]

The mtf chart indicates that 5219 approaches 100% mtf (full-on indisputable detail) to about 20 cycles/mm. Each cycle is two pixels and so we get 40 pixels a millimeter. So for super 35, 24mm * 13mm * 40 pixels horizontally * 40 pixels vertically = about 500,000 pixels (about 650,000 for anamorphic). Not so good. But these pixels are sharper than sharp.

The mtf chart indicates that 5219 hits 30% mtf (where fuzzy detail becomes pure fuzz) at 80 cycles/mm. Same math again gives 8,000,000 pixels (about 10,500,000 for anamorphic). Grain is way smaller than this but doesn't contribute much to the image beyond this point. At this point it's really, really mushy with almost no image at all.

So, trying to understand this, just below 3.8K for 35mm if you make this 80 cycles/mm as the cutoff point? Doesn't that seem to jibe with what many people's tests show shooting line resolution charts?

 

[Matt W.]

That's what I get, too, although I'll admit 30% is a fairly generous and somewhat arbitrary number (I've read anywhere between 20% and 50% offered as the "cut of point"). And that's assuming you're using a really, really good lens with perfect focus.

On the other hand, I've read that film has more resolution at higher contrast ratios, and I saw a test in which scans of Ektar and Tmax looked equally as sharp despite Tmax having like twice the linear resolution according to its mtf curves, so I don't really know if there is a concrete cut off point. Plus digital should hold onto a higher mtf for longer and then more suddenly slope downward, which means more area under the mtf curve and more coarse detail, thus more "sharpness" assuming the same extinction in resolution. But again I don't really know, it depends too much on in-camera sharpening, etc.--I've just had too much free time lately to research this so I figured I'd chime in.

Since pixels and resolution don't really seem to have a one-to-one relationship, I think it still boils down to "what looks good" being most important, as it should be. A 2k scan should get all the most important detail from film, but a 4k scan should get all the detail there is. To some extent that supports arguments forwarded by both Red and companies pushing 1080p. It also explains why a blu-ray from IMAX looks better than a blu-ray from super35 and yet a 4k film scan from super35 looks better than a 2k scan. But it's comforting (at least for a nerd like me) to see that mtf data correlates somewhat with test charts.

 

[Gavin Greenwalt]

I would be interested if anyone has done temporal testing. There's been displays in the past which 'interlace' a higher resolution display by vibrating the display slightly to offset the pixels creating the illusion of sub-pixels.

If you had for instance a static shot with a RED the pixel sites would be temporally static. Grain would mean you're sampling from different sub-pixel locations. Could give the illusion of increased resolution as well.

/* wild speculation

 

[Chuck Z]

I would be interested if anyone has done temporal testing. There's been displays in the past which 'interlace' a higher resolution display by vibrating the display slightly to offset the pixels creating the illusion of sub-pixels.

If you had for instance a static shot with a RED the pixel sites would be temporally static. Grain would mean you're sampling from different sub-pixel locations. Could give the illusion of increased resolution as well.

/* wild speculation

Never heard of that before.

Some older DLP mirrors have half horizontal resolution 960x1080 and use wobulation for full 1080p.

 

[Mark Collins]

It's a lot easier to "embrace your own ideals and make movies to your specifications" when you have an amazing track record and a blank check.

The desire to do something and the practical and financial ability to do it are usually two completely separate things.

Not necessarily true. I just wrapped my second short film this past weekend, and though it was a student short and we had almost no budget, it all came down to my production designer being amazing and my DP being even more amazing.

When you have people that know what they are doing, and are very good at it, money is no issue. Granted we had SWAT teams in the film, so we did have to budget accordingly for that, but our producer and I had no problem doing so because it sold the film.

Wally is truly an amazing cinematographer though, I would say he's the next Roger Deakins for sure. Working with IMAX sounds like a lot of fun, and very intense work. I'm glad that he chooses to shoot film though. Film truly is amazing and just has a look to it that no digital camera can touch.

 

[Daniel Browning]

Kind of! And the numbers are readily available on the spec sheets of any film (see the modulation transfer function chart on the right of page three):

We should keep in mind that these MTF charts include the benefits of high accutance developers, which is similar to sharpening, but with chemicals.

But then the issue is: what's the mtf curve of digital?

It's the sensor MTF combined with the OLPF MTF.

Is 4k digital 11 good megapixels (it's not) or 11 fuzzy ones?

It's not "good" vs "fuzzy". It's "aliased" vs "anti-aliased". You may consider aliased as "good", but I do not. You cannot have anti-aliasing without what you call "fuzzy".

If fuzzy then how fuzzy? Hard to say, few manufacturers release this information.

The sensor MTF has two primary components: OLPF, and the bare sensor.

The MTF of the bare sensor is well known and it's the same for all manufacturers that use square pixels. The only variable that changes this is the effective optical fill factor, but with microlenses, most manufacturers are in the same ballpark.

The MTF of the OLPF does vary, so that one will have more aliasing and more contrast; while another will choose an MTF that has less aliasing and less contrast. If they were all following good practice, they would all be the same in this. But they aren't, because some manufacturers think that their customers have such poor taste that they would actually prefer a sharp image with nasty aliasing artifacts over one that was slightly lower contrast but natural and free of the artifacts.

For example, MTF at Nyquist for 100% fill factor is 67% for all square-pixel image sensors: (MTF=(sin(pi*a*f))/(pi*a*f), where a is pixel width in microns and f is the spatial frequency in cycles/mm. Use f=1/(2*p) to add in fill factor (a=p for 100% fill factor).

Raw conversion and post processing have a tremendous effect on MTF, of course, and with sharpening it's often possible to completely reverse the loss of contrast from the MTF without introducing aliasing artifacts.

(at about 60-70% of the stated linear resolution in the case of a Bayer sensor, slightly higher in a 3 ccd or striped design)

No. Bayer, 3-chip, and striped are all the same in this regard. The only difference is that they *tend* to be different because of the different kind of artifacts you get if you don't filter them properly. On Bayer, you'll get chroma aliasing, which is so bad that no one would let a camera out the door without proper filtering (e.g. Canon 5D2). With 3-chip and striped, you might only get luma aliasing. To me, this is still terrible and should be avoided at all costs, but there are a lot of people who actually like those artifacts (widespread mental impairment is my best guess), so some cameras ship this way -- especially if they have a low pixel count (e.g. 1080p).

Red uses pretty aggressive filtering to avoid aliasing, commendably

Red's filtering is not "aggressive" at all; it's a well-reasoned balance between contrast and aliasing, and it's about the same as millions of other 9 MP Bayer image sensors.

, but it also results in a 4k that sometimes looks softer than 1080p.

No, that is incorrect. The 4K image downsampled to 1080p has an MTF that is as close to ideal as you can get. It's not even theoretically possible for a 1080p-native sensor to be that good. If you think it's "soft" because it doesn't have aliasing, then you can add them in post with a braindamaged downsampling filter. At least RED gives you the choice -- the aliased 1080p image has no option to remove the aliasing.

That's why 1080p sensors can never be ideal for 1080p delivery format.

Unless you apply sharpening, re-introducing faking resolution.

No, good sharpening just restores contrast at lower spatial frequencies that were affected by the OLPF. You're confusuing resolution with contrast. It's like high accuatance chemical developers in film -- it doesn't add any resolution, it just increases contrast. If you don't overdo it, it will look great.

 

[Matt W.]

Thanks for the corrections. I don't know much about sensor design, as I'm sure you've guessed. I'm not 100% clear on a few things, though:

For example, MTF at Nyquist for 100% fill factor is 67% for all square-pixel image sensors: (MTF=(sin(pi*a*f))/(pi*a*f), where a is pixel width in microns and f is the spatial frequency in cycles/mm. Use f=1/(2*p) to add in fill factor (a=p for 100% fill factor).

If I'm not mistaken, Nyquist is half the sampling frequency, so this would be about 46 cycles/mm for Red's sensor? So in theory would the red camera be resolving 67% mtf at 46 cycles/mm? Is that with or without taking the low pass filter into account? How then does bayer interpolation factor into this? Wouldn't the Nyquist frequency for red or blue, in absence of green data from which to interpolate, then fall as low as 13 cycles/mm (and 46 cycles/mm really be closer to 70% of that, or around 30 cycles/mm)? It would be so much simpler if digital camera manufacturers published mtf charts of their systems at default settings.

No, good sharpening just restores contrast at lower spatial frequencies that were affected by the OLPF.

That's interesting regarding sharpening; I always assumed sharpening increased aliasing, but what you wrote makes much more sense. It can't increase aliasing if there's no aliasing in the first place. I suppose, in theory, one could really boost the mtf of any system simply by sharpening the hell out of it. What's the mtf of hdr imagery? A lot of that stuff has rather insane amounts of wide-radius sharpening applied to it. As much as I'm not a fan of most hdr, I can see why it's so popular: tone mapping provides ridiculous dynamic range and apparent sharpness.

Red's filtering is not "aggressive" at all; it's a well-reasoned balance between contrast and aliasing, and it's about the same as millions of other 9 MP Bayer image sensors.

It may not be aggressive to your taste, but you must admit red's low pass filtering is pretty aggressive by industry standards. Look at the 7d, for instance, which looks way sharper than it has any right to until artifacts quickly appear. And most medium format digital backs for still photography (and the leica m9) have no low pass filter at all. Even my experience with old betacams is that many of them alias really quickly and really seriously. Whether or not it's smart to do as much, it seems camera companies have put a premium on apparent sharpness over a lack of aliasing for quite some time now. And none of the 1080p cameras out there have 1080p sensors, either; they're merely designed for 1080p post rather than 4k post, another reason why many of them look much sharper per-pixel than Red's 4k output (which, if properly sharpened and then scaled to 1080p, would no doubt be as good or much better).