ProRes 4:4:4 possible?

[Grug]

I vaguely seem to remember 4:4:4 being mentioned as one of the ProRes recording options, but can't actually find any reference to it anywhere. Does anyone know whether it's been announced as one of the available flavours? Or is it just 4:2:2?

[Barry Green]

ProRes does offer a 4:4:4 profile, but you can't get 4:4:4 off a single Bayer sensor, unless you downconverted to about 720p. You won't get anywhere near 4:4:4 at 2.5k. Count on a bit better than 4:2:0 at full res.

[adam777]

ProRes does offer a 4:4:4 profile, but you can't get 4:4:4 off a single Bayer sensor, unless you downconverted to about 720p. You won't get anywhere near 4:4:4 at 2.5k. Count on a bit better than 4:2:0 at full res.

Can CinemaDNG 2.5k from this camera be converted to 2.5k prores444 and 2.5k prores422HQ through DaVinci Resolve?

[Grug]

ProRes does offer a 4:4:4 profile, but you can't get 4:4:4 off a single Bayer sensor, unless you downconverted to about 720p. You won't get anywhere near 4:4:4 at 2.5k. Count on a bit better than 4:2:0 at full res.

Cheers Barry, I'll always forget about the damned Bayer pattern. 4:2:2 it is then.

[Andrew_HD]

Cheers Barry, I'll always forget about the damned Bayer pattern. 4:2:2 it is then.

Buy Sony F65 than- it uses different pattern (2x more data than Bayer)

[Barry Green]

Cheers Barry, I'll always forget about the damned Bayer pattern. 4:2:2 it is then.

It's not 4:2:2 either, it's more like 4:2:0.

Look at it like this -- the BMC has a sensor that's got an active area of 2432 x 1366. And it creates raw files that are 2432 x 1366. The question is what's the color resolution of those files, and is it capable of 4:4:4 or even 4:2:0?

Well, remember that the raw file is made up of green and red and blue pixels. No full-color RGB pixels, just individual red and green and blue. And if you wanted 4:4:4, you'd need 2432x1366 red, 2432x1366 green, and 2432x1366 blue. Is that what the BMC sensor has? Nope. It has 2432x1366 in a Bayer pattern, which means: 1216x1366 green, 1216x683 red, and 1216x683 blue. There's nowhere near enough color samples to make 4:4:4, or even 4:2:2. 4:2:2 would require 2432x1366 red and blue, and the BMC just doesn't have them. It has half that.

So, if we took the sensor at "face value" and used each sample as a sample, we could get a very close representation of 4:4:4 720p. It wouldn't be a true 1280x720, it'd be 1216x683, but come on, that's close enough, right?

Some sensors do work that way. For example, the Canon C300 sensor delivers 4:4:4 1080p because it has 1920x1080 red, 1920x1080 blue, and 1920x2160 green. It doesn't do a bayer de-mosaic, it just does a straight sample, one pixel to one pixel, and it doesn't try to make its image bigger than a 1:1 ratio. So that sensor delivers a true sampled 4:4:4 1080p (even though the recording format limits it to 4:2:2).

But other sensors don't try to do a direct 1:1 output. They instead combine the pixel samples with neighboring pixel samples in an attempt to synthesize a higher resolution image. This is called de-mosaic'ing, or debayering. During this process you can actually substantially increase the luma resolution (brightness) but you can't really increase the chroma resolution. So while on the BMC you'll have 1216x1366 green, the de-mosaic process will work some magic and use the sampled green and the sampled red and the sampled blue, working the detail from each to come up with an image that is probably going to be spot-on about 1920x1080 of true resolved luminance resolution. But it won't be 1920x1080 in color resolution. You still only have 1216x683 of color to go around, there's no way to get more. You can "guess" and "interpolate" and "synthesize" the missing color information, but actual sampled color resolution basically is what it is.

So -- in 1080p, 4:4:4 would require color resolution of 1920x1080. 4:2:2 would be color resolution of 960x1080. And 4:2:0 would be color resolution of 960x540. What will the BMC (likely) deliver? A 1920x1080 image with color resolution of 1216x683. Neither 4:2:2 nor 4:2:0. Chroma res that's higher horizontally than 4:2:2, but much less vertically than 4:2:2. Very much like 4:2:0 in color resolution vertically. So what is it, really? Hard to say. It's more than 4:2:0, less than 4:2:2 (based on the idea that 4:2:0 has 518,400 color samples, 4:2:2 has 1,036,800 color samples, and the BMC will have about 830,528 color samples. So about halfway inbetween 4:2:0 and 4:2:2. Better than 4:2:0, not as good as 4:2:2.

Keep in mind that de-mosaic'ing is a bit of a black art and a magical science, and different de-mosaic algorithms will deliver different results. We've typically always used about 70% as the target for how much real luminance resolution you can achieve after de-bayering (so, 2432 would yield about 1700 pixels of real detail) but Graeme at Red has worked some true wizardry that is resulting in legitimate results of about 80% of retained detail (hence why their 5K sensor delivers 4K of real resolution); using Graeme's demosaic technology the 2432-wide BMC sensor would deliver about 1945 pixels of real resolved detail -- in other words, a perfectly-resolved 1080p image. (in luminance detail, of course, but not in color).

[Barry Green]

Buy Sony F65 than- it uses different pattern (2x more data than Bayer)

Yep, the F65 grossly oversamples its 8k sensor (like the C300 does its 4k sensor) to deliver truly-resolved images.

Interestingly about the C300 -- it can deliver a true 1920x1080x4:4:4 out of its 4K sensor. And a Red One delivers a resolved 3.2k at 4:2:0+ish out of its 4K sensor. Both systems use the exact same bayer pattern! It's just a different way of reading the data. If Red chose to, they could avoid de-mosaic'ing their sensor and instead process it as sampled points, and it would deliver a fully-resolved 1920x1080x4:4:4. And if Canon chose to enable it (which is probably what they're doing in the C500) they could get a 3.2k image (at, again, 4:2:0+) out of the same sensor that they've got.

That's one of the fun things about shooting raw -- when better technologies come along, you could reprocess your raw footage and get different results.

[Andrew_HD]

Wait here- C300 has some "problems" with resolving motion. I read comments that when it comes to static picture C300 is great, but when motion comes into play than it's not that great anymore. Is it to do with the way how it reds data form the sensor? This would make some sense

[laco]

some people complaining about the C300, because it looks "videoish" when held in hand, compared to the Alexa.
So I don't think anyone will notice anything, until you always shoot with alexa

[tfg]

It's not 4:2:2 either, it's more like 4:2:0.

So while on the BMC you'll have 1216x1366 green, the de-mosaic process will work some magic and use the sampled green and the sampled red and the sampled blue, working the detail from each to come up with an image that is probably going to be spot-on about 1920x1080 of true resolved luminance resolution. What will the BMC (likely) deliver? A 1920x1080 image with color resolution of 1216x683. Neither 4:2:2 nor 4:2:0. Chroma res that's higher horizontally than 4:2:2, but much less vertically than 4:2:2. Very much like 4:2:0 in color resolution vertically. So what is it, really? Hard to say. It's more than 4:2:0, less than 4:2:2 (based on the idea that 4:2:0 has 518,400 color samples, 4:2:2 has 1,036,800 color samples, and the BMC will have about 830,528 color samples. So about halfway inbetween 4:2:0 and 4:2:2. Better than 4:2:0, not as good as 4:2:2.

Still trying to visualize this, but in essence you are saying that a 1920x1080 4:2:2 (prores/DNxHD, etc) will look strikingly similar to the full RAW images, correct? That's because the RAW image, once debayered, isn't going to resolve anything better (perhaps a fraction more horizontal chroma resolution, but that's it) ? Very informative stuff.

[tfg]

It's not 4:2:2 either, it's more like 4:2:0.

So while on the BMC you'll have 1216x1366 green, the de-mosaic process will work some magic and use the sampled green and the sampled red and the sampled blue, working the detail from each to come up with an image that is probably going to be spot-on about 1920x1080 of true resolved luminance resolution. What will the BMC (likely) deliver? A 1920x1080 image with color resolution of 1216x683. Neither 4:2:2 nor 4:2:0. Chroma res that's higher horizontally than 4:2:2, but much less vertically than 4:2:2. Very much like 4:2:0 in color resolution vertically. So what is it, really? Hard to say. It's more than 4:2:0, less than 4:2:2 (based on the idea that 4:2:0 has 518,400 color samples, 4:2:2 has 1,036,800 color samples, and the BMC will have about 830,528 color samples. So about halfway inbetween 4:2:0 and 4:2:2. Better than 4:2:0, not as good as 4:2:2.

Still trying to visualize this, but in essence you are saying that a 1920x1080 4:2:2 (prores/DNxHD, etc) will look strikingly similar to the full RAW images, correct? That's because the RAW image, once debayered, isn't going to resolve anything better (perhaps a fraction more horizontal chroma resolution, but that's it) ? Very informative stuff.

[manakiin]

Strikingly similar if you don't plan on heavy post work. Onboard compressed may even look better out of the box (sharpening, processing, etc.).

Overall RAW will be noticeably better than onboard ProRes / DNxHD if properly processed.

[bumkicho]

Strikingly similar if you don't plan on heavy post work. Onboard compressed may even look better out of the box (sharpening, processing, etc.).

Overall RAW will be noticeably better than onboard ProRes / DNxHD if properly processed.

What would be involved to properly process Raw out of this camera? Would Davinci Resolve suffice or something else should be involved in this process?

[Andrew_HD]

Resolve will do everything what you need at high quality, but there are other ways also.

[Barry Green]

Still trying to visualize this, but in essence you are saying that a 1920x1080 4:2:2 (prores/DNxHD, etc) will look strikingly similar to the full RAW images, correct?

That is my guess, yes.

That's because the RAW image, once debayered, isn't going to resolve anything better (perhaps a fraction more horizontal chroma resolution, but that's it) ?

In a nutshell, yes. It sounds like raw should be "higher res" because it's 2432 pixels wide instead of 1920, but in reality that additional resolution isn't "real" until it's been debayered. And you lose anywhere from 20 to 30% of "nominal" resolution during the debayering process. Which is why a Red One, with the genius of Graeme Nattress behind it, can deliver at best about a 3.2k image out of its 4K sensor -- a 20% loss of the "nominal" resolution. And Graeme's debayer is extremely intelligent and extraordinarily efficient. It's probably more realistic to count on about a 25% loss from nominal to actual resolved detail. Using those numbers, the 2432 pixel wide sensor on the BMC should result in somewhere between 1945 and 1824 pixels of real, resolved detail. And 1920 is between them, so ... that's why I'm saying that in terms of actual resolved picture information, the BMC is pretty much ideally suited to make a great 1080p.

As for chroma resolution, yes it does look like you'll have more in the raw file. There's 1216 chroma samples on each line in the raw, whereas in a 4:2:2 codec that will be scaled to 960. In the vertical direction it doesn't matter, there's only 683 samples in the raw, and so the 4:2:2 codec will hold all that the sensor can give it. Technically a 4:2:2 codec can store 1080 lines worth of chroma resolution, but if the sensor can't deliver that much, then you can't store more than the sensor can give you, so 4:2:2 has enough space to hold all the vertical chroma res the sensor can feed it. Not as much room to store all the horizontal chroma res, but more than enough room to store all the vertical.

So the advantages to raw would be a) 12-bit instead of 10-bit, b) uncompressed instead of compressed, c) higher color resolution, and d) future flexibility.

The benefits of a) remain to be seen, as to how much actual detail and true color levels the sensor fits into the raw. I don't have any reason to suspect that we won't get full 12-bit 4,096 shades, but that is the absolute maximum. Noise levels or processing issues may reduce that total number in terms of what we can actually discern and what is actually "workable", but no matter what, we'll have 10 bits in the ProRes and DNxHD, which is better than any large-sensor camera under the F3 can give you, so it's a choice of "way better" or "way way way better".

As for b), hey, uncompressed is always better image-quality wise. However, to be fair, high-profile DNxHD and ProRes are really good codecs, and it's going to take a really trained eye to distinguish between those and uncompressed, especially when the frames are flying past at the rate of 24 per second. I'm not discounting the beauty of uncompressed, I'm just pointing out that for many practical shooting scenarios I think the vast majority of people are going to be plenty satisfied with the (excellent) compressed options as well.

As for c) hey, better is better. Now, the counterpoint is -- visually we don't see color resolution nearly as clearly as we see brightness resolution, which is why cameras with 4:2:0 codecs can still look strikingly good to us. On the other hand, the computer sees everything, and raw 1216x683 color res is going to be better for keying/compositing and heavy effects work than 960x540 4:2:0 would. And it will be better than the 960x683 you'd get from the 4:2:2 codec too. Better is always better, and more is better, so if you want to extract the absolute maximum performance from your BMC, raw will let you get at everything the sensor can deliver.

Which leads us to d) flexibility. Raw gives you more to work with, but note the key word there: work. You have to work raw, you can't just view it. Raw files look horrible. They're flat and mushy and soft and ugly and noisy. When working with raw it is incumbent on YOU to make them look good. You have to de-noise the footage, you have to color the footage, you have to de-bayer the footage, you have to apply curves and grading to the footage... the camera, in raw, will not deliver any "magic", whether "black" or otherwise. It's up to you and your usage of Resolve (or Magic Bullet Looks or whatever) to turn that pasty raw flatness into stunning gorgeous colorful bright beauty. And that will take time and storage space and effort. (Of course, the effort might be a one-click "look" like Magic Bullet, followed by the appropriate rendering, so it might be something you set up to run overnight, but the point is, the work still needs to be done).

Is all that work worth a couple of bits of color depth and a couple hundred chroma samples? That's the fun part of the BMC -- you get to decide. You can shoot raw and get it all, or you can use ProRes and/or DNxHD and get probably 90% of the goodness with very little effort, just drag and drop files to your timeline and start editing. It's up to you and what your preferences are. I personally feel that there are "get-it-right-in-camera" shooters, and there are "shoot-it-flat-and-grade-it-in-post" guys, and this BMC should appeal to both crowds (although I suspect that the in-camera options are going to be quite limited, it's not going to have menu structures like an EX3 or HPX3700 after all!)

But -- if you do shoot raw, and you ever want to revisit your footage (say you're James Cameron and twenty years later you want to re-release Titanic, for example) raw gives you the option to go back and re-process the footage. What if debayering techniques have advanced to the point where they're 90% efficient five years from now? You could go re-process your footage and -- voila, it'll be sharper than it was! That's one of the neat things about raw, is that you always have the raw data, not some locked-in look.

If you can afford the storage space and the processing time, always shoot raw. If your priority is a quick turnaround, you'll probably find ProRes (on a Mac) or DNxHD (on Windows) to give you massive space and time savings, while retaining practically all the quality.

Of course, once again, this is all somewhat theory until we get to put it to the test on real images.