Michael Cioni prepping for a 4K world

[Jeff Whitehurst]

IS the video Back UP.

http://www.youtube.com/watch?v=Qztrb9arZl4&feature=player_embedded

Only in 2k, though ;-)

 

[Gary Paul Walker]

Ok, I just watched this - REDs next project should be to clone Mr Cioni.
Seriously. Make this happen.

 

[brett ogden]

Thay will do that after dragon upgrade :ihih:

 

[Marc Wielage]

Unfortunate that this is down ... his reasoning on not using LUT's and going with curves/etc was very interesting(I don't like using LUT's but have always felt like i'm a wimp for not using them, his argument kind supports how i like doing it).

There's many ways to skin a cat. There was a period when a certain LA facility was literally the "LUT of the Month Club," and colorists would routinely tweak stuff all over the map, between display LUTs, output LUTs, and tweaks inside the color corrector. What drove me crazy were people who would change LUTs inbetween reels. In one facility, a dozen different LUTs were being used, often with no record of who was using what, where, for which project.

Using a "lifted" (non crushed/clipped) balanced DPX log file will absolutely work, and we worked with film like that for a decade. The trick is finding a sweet spot for the color corrector to not have to grind the image down into a workable area. As long as you can believe the image you see on screen, and it can be reproduced in the appropriate color space (XYZ, P3, Rec709, whatever), it doesn't matter how you get there. RedLogFilm can work fine, as Michael says. I'm more concerned about making sure the client is happy, the session moves along at a reasonable speed, and everything gets delivered on time.

Michael's explanation on LUTs starts at about 28:40. Very interesting comments.

I gotta remember Michael's line: "A gig a second -- a gig, a gig, a gig." Yep, that's a lotta data.

 

[M Most]

Michael's explanation on LUTs starts at about 28:40. Very interesting comments.
.

Interesting, yes. Michael's a good speaker. I just have a couple of points to touch on..

First, P3 is not "linear." It is gamma encoded by definition, to a gamma of 2.6. I'm surprised Graeme, who usually corrects such oversights, hasn't chimed in on this one. In fact, to properly transform P3 to, say, Rec709, you need to "de-gamma" it - thus putting it in true linear space - and then re-transform to Rec709 gamma. I wish Michael didn't throw the term "linear" around so cavalierly as it leads to even more misuse of the term.

Second, 3D LUTs that are used for color space transforms are one thing, but simple 1D LUTs to implement, say, log to video gamma, or log to 2.6 gamma for P3, are very useful to many colorists becasue they are mathematically derived and represent the proper transform from the original log curve to the target gamma. If you're using a grading system that has a true exposure control (and hopefully a contrast control), this allows you to adjust exposure in a way that maintains the photographic contrast properly without having to guess as to what that was. While it's true that maximum and minimum values are predefined, in most cases this isn't really a "limitation" because every working space also has predefined max and min values and bit depth (well, all except for ACES, but that's a different topic and conversation). You can't go higher than 1023 in a 10 bit format no matter what you do. LUTs in that scenario are a very good way of getting a representative starting point for further grading, and when used in a properly designed pipeline can allow you to tailor the photography - which might be over or underexposed for various reasons, some intentional, some not - to the working space with little ambiguity. Curves and LUTs are essentially the same thing in a 1D scenario, it's just that one is predefined and the other can be altered by the user. However, using a curve requires a bit of knowledge and experience with the formats you're deaing with, and an improperly designed curve can cause a lot of unnatural effects in terms of the grey scale of the result. It's nice to say things like "no creative limitations," but in actual practice, a curve approach can lead to problems as well as solutions depending on the experience of the one using them. And for formats other than R3D, where color matrices are applied after the debayer has occurred, a 3D LUT becomes something of a necessity if one wants to use a properly designed color matrix. So while I don't disagree with some of Michael's sentiments, there's more to that story than the rather simple notion that "LUTs are bad because they limit creativity." They can be very useful in many, many situations.

 

[Nick Pasquariello]

Yeah Michael does know his S^#@%...

+1.


Also, personal favorite line: "It's a box that vomits data. It barfs out all sorts of video transcodes." I never would have thought that I would want something who's primary feature is "barfing".

 

[SeanBrown]

Great Job

 

[Michael Cioni]

Interesting, yes. Michael's a good speaker. I just have a couple of points to touch on..

First, P3 is not "linear." It is gamma encoded by definition, to a gamma of 2.6. I'm surprised Graeme, who usually corrects such oversights, hasn't chimed in on this one. In fact, to properly transform P3 to, say, Rec709, you need to "de-gamma" it - thus putting it in true linear space - and then re-transform to Rec709 gamma. I wish Michael didn't throw the term "linear" around so cavalierly as it leads to even more misuse of the term.

Second, 3D LUTs that are used for color space transforms are one thing, but simple 1D LUTs to implement, say, log to video gamma, or log to 2.6 gamma for P3, are very useful to many colorists becasue they are mathematically derived and represent the proper transform from the original log curve to the target gamma. If you're using a grading system that has a true exposure control (and hopefully a contrast control), this allows you to adjust exposure in a way that maintains the photographic contrast properly without having to guess as to what that was. While it's true that maximum and minimum values are predefined, in most cases this isn't really a "limitation" because every working space also has predefined max and min values and bit depth (well, all except for ACES, but that's a different topic and conversation). You can't go higher than 1023 in a 10 bit format no matter what you do. LUTs in that scenario are a very good way of getting a representative starting point for further grading, and when used in a properly designed pipeline can allow you to tailor the photography - which might be over or underexposed for various reasons, some intentional, some not - to the working space with little ambiguity. Curves and LUTs are essentially the same thing in a 1D scenario, it's just that one is predefined and the other can be altered by the user. However, using a curve requires a bit of knowledge and experience with the formats you're deaing with, and an improperly designed curve can cause a lot of unnatural effects in terms of the grey scale of the result. It's nice to say things like "no creative limitations," but in actual practice, a curve approach can lead to problems as well as solutions depending on the experience of the one using them. And for formats other than R3D, where color matrices are applied after the debayer has occurred, a 3D LUT becomes something of a necessity if one wants to use a properly designed color matrix. So while I don't disagree with some of Michael's sentiments, there's more to that story than the rather simple notion that "LUTs are bad because they limit creativity." They can be very useful in many, many situations.

Mike is right about the word "linear" being flippantly swapped for "gamma corrected" or "gamma encoded." -Trying to explain this concept to 200 people at the Final Cut Pro User Group meeting without losing the point: notta gonna happen.
LUTs are not evil. I probably shouldn't have said they limit creativity (on every level) because I don't necessarily believe that in all contexts. However, given the context of the discussion and the audience of that evening, the statement was intended to remove the fear and concern that "LUTs are how you get things to look like the pros." -That is a muth. Patrick said exactly how I used to feel when we started color correcting in 2005: like we were wimps! I hated that feeling and wasn't buying into the system that I was being sold. But the more we dug, the more we realized that the average LUT (based on the parameters of digital and often Quicktime-based projects) had destructive affects on the pictures.

I'm currently at the HPA Technology Retreat in Palm Springs and gave a few presentations here including the 2012 Event Keynote. My favorite moment, however, was when my good friend and mentor, Leon Silverman, (paraphrased) said this about LUTs to the entire group:

"Custom LUTs that are built to for in-house, proprietary, secret-tsauce systems that you "need" to make your pictures look good are about as sturdy a platform as trying to stand up in a hammock."

m

 

[Brad Allen]

Great to see you've found this thread Michael!

Had a quick question for you. You mentioned in your presentation that shooting 5k for 4k was mostly useful for shooting an "overscan" area that can be used to give more framing options in post + stabilzation etc. However, I always thought shootring 5k for 4k had to do with the fact that a 5k bayer image will only resolve to roughly 80% of it's native resolution - thus down sampling 5k to 4k gives you a legitimate 4k worth of resolution. Could you comment on that at all? Do we need to be shooting at 6k to accommodate down-sampling and overscan?

Thanks for all your good work and instruction! Would love to have a chat to you at NAB this year if we get the chance

 

[mikeburton]

It's always inspiring to hear you speak Michael.

 

[Roberto Lequeux]

With 4k delivery in mind, figuring out the Bayer resolution necessary for flawless 4k in all channels (even if this is major overkill for the human eye) would provide the 'right' figure to frame for, with any additional resolution RED gave us serving for the wonderful overscan benefits: stabilization, stitching performances, re-framing, etc.

I wonder what resolution Dragon will provide... With Epic's engine being capable of 120fps at 5k, the hard ceiling is pretty freakin' high...

 

[Michael Cioni]

Great to see you've found this thread Michael!

Had a quick question for you. You mentioned in your presentation that shooting 5k for 4k was mostly useful for shooting an "overscan" area that can be used to give more framing options in post + stabilzation etc. However, I always thought shootring 5k for 4k had to do with the fact that a 5k bayer image will only resolve to roughly 80% of it's native resolution - thus down sampling 5k to 4k gives you a legitimate 4k worth of resolution. Could you comment on that at all? Do we need to be shooting at 6k to accommodate down-sampling and overscan?

Thanks for all your good work and instruction! Would love to have a chat to you at NAB this year if we get the chance

Good question. In this case, 5K for 4K was used for the reasons I mentioned. I also mentioned it because I believe it's a fantastic way everyone should consider shooting because it is littered with benefits from all directions. I speak more about it in this December blog entry (about halfway through).

http://michaelcioni.tumblr.com/post/14725750331/4k-digital-intermediate

With regard to resolution; how much is enough is not an answer anyone can qualify. If you subscribe to the Nyquest approach, you need to double the resolution at the source in order to end up with the desired resolution at the display. In other words, Nyquest theorizes that in order to have real 4K pictures, you need to have an 8K source. I agree with this theory in terms of mathematics. So per your question, based on the Nyquest approach (which many subscribe to) 6K isn't technically enough to deliver true 4K. But it helps. 5K is a major improvement over 4K which is why Epic looks noticeably better than RED in 4K. But when it comes to desire, consider this;
Even when RED releases a 9K sourced sensor, you may not want it to look like 4K (for creative reasons). I believe taking it in small strides is having a more positive affect on the market (professional and consumer). In other words, 4K center extraction is really a 3K picture at best. In the case of GDT, it is not using every possible pixel to make up the image. But that was this filmmaker's choice and will be the choices of many filmmakers as we creatively navigate this. So that's why I mentioned in my blog for people to experiment and evaluate. I didn't invent this idea by any means, but I totally love what it does to the pictures.
Either way, by doing it this way, even if the images from GDT are not full 5K, the are undeniably better than 2K!
m

 

[Brad Allen]

Thanks for that Michael

Imagine that - 9k for 4k! It will happen one day I suppose - maybe not as far off as we think

Totally understand that it was the filmmakers choice to essentially only end up with a 3k worth of resolution in a 4k image. Just wanted to check that their wasn't something I was missing. One other choice I was surprised about was to see that the broadcast release would have more source image used than the cinema release. I would have expected that a director would aim to have the cinema release featuring the most image with the broadcast version then being cropped if necessary. Interesting to see that that isn't the case here. Always great to get a behind the scenes scoop of what went on

 

[AJ Aguirre]

Awesome, really really informative and I feel this hits the nail on the head for so many people wondering what the heck is going on. I have a presentation for class I am going to base on this. Great presentation Michael!

 

[Marc Wielage]

One other choice I was surprised about was to see that the broadcast release would have more source image used than the cinema release. I would have expected that a director would aim to have the cinema release featuring the most image with the broadcast version then being cropped if necessary.

That has often been the case with all the Super 35mm films shot over the past 15 years. Generally, the DP makes the decision to zoom out slightly for the home video transfer, showing more foot room. Headroom stays the same, but you see maybe 10-15% more image on left, right, and the bottom of the frame. This is a pretty effective compromise for 16x9 compared to the 2.40 theatrical version, and it also makes the inevitable 4x3 pan/scan (which still exists for most delivery contracts) faster and easier to accomplish.

As long as Michael is looking in on this discussion: I'd like to know how both Pablo rooms were used simultaneously on the project. I assume each room mainly did every other reel, which makes sense on a lengthy project that had to get done in a set amount of time. I'm also surprised nobody has made much of a fuss of the fact that they opted not to use R3D files and just debayer everything out to uncompressed 4K/5K DPX log files and go from there.

Did the client ask you to pull along a second simultaneous timeline of the original dailies color decisions, just as a reference? I'm curious how and when that was used in the workflow, in terms of the final color correction. I'm also curious when you chose to color-correct in 2K vs. 4K on the available projectors, and how much of an effect that had on the decision-making process.

 

[Graeme Nattress]

If you subscribe to the Nyquest approach, you need to double the resolution at the source in order to end up with the desired resolution at the display.

That's a mis-characterization of sampling theory (Nyquist-Shannon sampling theory: http://en.wikipedia.org/wiki/Nyquist–Shannon_sampling_theorem). Sampling theory is often stated in terms of frequencies: "If a function x(t) contains no frequencies higher than B hertz, it is completely determined by giving its ordinates at a series of points spaced 1/(2B) seconds apart."

But with images we deal with pixels (which are the samples) rather than frequencies. As seen by above you need at least twice the number of samples as the frequency to properly reconstruct the signal. In images, frequency is the number of line-pairs in the image, so it's saying you need at least twice the number of pixels as the line-pairs. Each line-pair is made up of two pixels, so now it's saying you need at least twice the number of pixels as two pixels, or dividing throughout by two, you get at least the number of pixels as pixels. So for an image container to hold 4k resolution the theory states you need at least 4k samples.

The caveat with sampling theory is that we're dealing with bandwidth limited signals. In images and cameras this is rarely the case, but we do try. The optical low pass filter (OLPF) tries to limit the incoming signal so that aliases never occur. But OLPFs are quite "soft" in that as well as reducing the frequencies that cause aliasing, they soften the lower frequencies we wish to keep as well. Getting a balance between maintaining over-all MTF and reduction of aliasing to a point where it's rarely, if-ever provoked in the real world is a tricky balancing act.

So when you talk about starting with more resolution than you need, you're not talking so much from a Nyquist-Shannon sampling theory approach, but from "giving the OLPF room to work in" approach so that you have very strong MTF at the desired resolution along with minimal aliasing. That is why the real way to generate a superb 1080p image is to start from a higher resolution source rather than trying to achieve a full measured 1080p from a 1080p sensor. Similar issues occur in audio where sampling is performed at a much higher frequency than the target 44.1kh to allow room for gentler low pass filters which exhibit less ringing than the traditional "brick wall" filter that was used in the early years of CD.

Graeme

 

[Brad Allen]

That has often been the case with all the Super 35mm films shot over the past 15 years. Generally, the DP makes the decision to zoom out slightly for the home video transfer, showing more foot room. Headroom stays the same, but you see maybe 10-15% more image on left, right, and the bottom of the frame. This is a pretty effective compromise for 16x9 compared to the 2.40 theatrical version, and it also makes the inevitable 4x3 pan/scan (which still exists for most delivery contracts) faster and easier to accomplish.

Now that I think about it - I think I knew that this happened. It just seems bizarre in my way of thinking. If it was my choice I think I would be aiming to have my cinema release as the standard against which all others are measured. Max resolution, best colours, least compression, best sound and optimum framing

Not a criticsm for doing it another way of course.

 

[Brad Allen]

The caveat with sampling theory is that we're dealing with bandwidth limited signals. In images and cameras this is rarely the case, but we do try. The optical low pass filter (OLPF) tries to limit the incoming signal so that aliases never occur. But OLPFs are quite "soft" in that as well as reducing the frequencies that cause aliasing, they soften the lower frequencies we wish to keep as well. Getting a balance between maintaining over-all MTF and reduction of aliasing to a point where it's rarely, if-ever provoked in the real world is a tricky balancing act.

So when you talk about starting with more resolution than you need, you're not talking so much from a Nyquist-Shannon sampling theory approach, but from "giving the OLPF room to work in" approach so that you have very strong MTF at the desired resolution along with minimal aliasing. That is why the real way to generate a superb 1080p image is to start from a higher resolution source rather than trying to achieve a full measured 1080p from a 1080p sensor. Similar issues occur in audio where sampling is performed at a much higher frequency than the target 44.1kh to allow room for gentler low pass filters which exhibit less ringing than the traditional "brick wall" filter that was used in the early years of CD.

Graeme

Hi Graeme - thanks once again for your incredibly informative posts.

Just a quick question. Are you saying that it is the OLPF that is primarily responsible for measured resolution loss on the camera? Does the Bayer process also affect measured resolution? I could have almost sworn I'd read this in other posts - but would always prefer to be set straight if I've got this wrong

 

[Graeme Nattress]

Brad, OLPF is the primary factor - the Bayer CFA is secondary. Most cameras that don't use a Bayer array let through more aliasing than I'd consider acceptable, but they generally have a low "native" resolution, so they try to get as much out of it as possible through also allowing through significant aliasing.

There's a trend in still photography to allow much more aliasing than we'd accept in the motion side of the industry. This is not least because if there's an issue, it's much easier to fix seamlessly on a single frame than on thousands. Although with Bayer there's often a concern about chroma moire, that's actually the most easily fixable aspect both in motion and stills, so that leaves luma aliasing (which effects all types of sensors) as the one to worry about!

Basically, if you want higher un-corrupted measured resolution, got for higher-still sensor resolution. The close you want the measured resolution to be to the sensor resolution, the more likely you're going to get the corrupting influence of aliasing.

Graeme

 

[Brad Allen]

Got it . That makes perfect sense.

It's also probably one reason I'm preferring the stills coming out of my scarlet than DSLR's I've used in the past.

Thanks once again for sharing - would love to shout you a drink/meal during NAB to thankyou for all the help you have been over the years