RAW workflow - a discussion

[Blair S. Paulsen]

To me the core issue of this thread is whether staying RAW as long as possible is really a significant benefit over a proper transcode to a robust RGB format.

To determine that there are several issues that need to be examined in a comparative manner to find any pitfalls or benefits in terms of image quality. The debate over the practicality of staying RAW longer is typically more about the particular production environment one encounters on a given project. It would be great if we could cite universally beneficial practices that would apply to any project but as the spectrum of comments on this forum proves, that's not reality.

The reality I see is that budget, turnaround time, sophistication of personnel, post tools, etc are all factors that need to be considered all the way to how a scene is lit and exposed. I also believe that native scene DR and whether elements in the frame with lower exposure levels need to "play" in the finish have to be considered in choosing a strategy. For example, for a dimly lit scene I might add more light to get a cleaner "RAW negative" and then restore the shadows in the grade to keep the noise from intruding. On a sunny exterior I might underexpose a little to hold more highlight information knowing that with all that blue light I should be able to dig whatever I need out of the shadows with minimal noise.

ISO on a digital camera can be counterintuitive, so I suggest we think in terms of signal to noise ratio. One of the things I noticed immediately with MX sensor performance in the Epic vs the R1 was the increased precision in the shadows (which I attribute to the added bit depth) changed the noise "signature", I would wager that's one of the things that Gunleik notices when comparing them. The lighter compression options made possible by the higher data rates also seems to give more texture to areas of relatively consistent density/chroma which I assume is due to less interpolating in the wavelet encode - but that's a Graeme issue that he may be constrained from discussing in detail publicly. In any case, to add a more seat of the pants perspective to all this tech talk, I have a story:

My friend, and legendary RedUser, Evin Grant, embraced the noise right from the earliest Mysterium/R1 days. He liked the texture and thought of it as digital grain. He loved the look of film and I think he found the "clean" digital negative too sterile, so he sought out some noise to add some guts and grit. He made some very evocative imagery with that approach, though his skill as a DP was probably more to "blame" (hehe) than his ardor for noise. I bring this up, because at the time I was firmly in the other camp trying to kill noise because I disliked the way the type of noise we had in 2008 seemed to look roughly the same no matter what the other image metrics might be - it was easier to see and harder to kill under 3000K, but its character was roughly the same as in dark shadows under daylight.
With the introduction of the MX sensor and the concurrent release of new color science (RC2/RG2 IIRC) the noise "character" changed rather dramatically. Instead of chunky it was now a smooth gray bottom end and, even better, we had another 2 stops of genuine DR before it even went into the noise floor. For grain aficionados like Evin, there were now several post plugs that could dial in the amount and type of grain desired - I believe there was even one company that used scans of popular film stocks to create a random noise pattern that could be comped in during DI/finishing.
With the Epic/MX package I am now seeing noise "signatures" that track more accurately to the specific scene and subtle differences in the deep shadows. This makes me more comfortable with letting the noise "play" in certain material for some of the same goals Evin had. I still think adding random grain in post is a better way to get that grittier look, but that takes some serious computational horsepower in post AFAIK. In any case, its worth noting that all noise is not created equal and so determining a specific numerical value of S/N in db is not as purely empirical as we camera nerds would like.

FWIW, even with the newest color science I still pop on a 1/4 CTB anytime the light sources are under 3000K. I know that Graeme has gone to great lengths to optimize the reconstruction of balanced images even when the color temp is near the extremes, but it seems to help (perhaps its just my bias clouding my judgement). I bring it up in this thread because I believe it makes the choice to transcode to an RGB format at the one light stage easier to do.

Cheers - #19

[Martin Stevens]

Graeme ,

Is it your opinion that you or someone can take an R3D file and convert it to an RGB file within RedCine-X Pro
and preserve ALL of the image data?

Also, I know there is REDlogFilm, but why not a REDlinFilm?

[Graeme Nattress]

Marting, isn't that linear light EXR?

Graeme

[Martin Stevens]

Marting, isn't that linear light EXR?

Graeme

Thanks Graeme.

Also...

I mentioned to you here on RedUser, perhaps more than a year ago, that the 4K in your signature should be 5K.

Well, with the Dragon sensor coming out, perhaps we will soon see 6K in your signature.

:)

[Gunleik Groven]

Gunleik, you have said in the past (please correct me if I misremember) that you find you retain more highlight detail, even in REDlogFilm, when you drop the ISO from 800 to 320. I have to say I have not found this to be the case, and as Graeme says it should not be the case. Can you elaborate on your findings on this?

I think maybe the big difference occured when we went from a 10bit log to a 16 bit log pipeline.

Haven't had the same need to look that cloesly at the highlights after that, though I still try to keep them not too compressed.

technically - of course - Graeme is 100% right. We just had some more luck and quicker results by doing it this way.
But don't write a book on that.

Test, check what works.
If you have issues with your highlights, this may well be a way to approach it...

Kinda

G

[Gunleik Groven]

To me the core issue of this thread is whether staying RAW as long as possible is really a significant benefit over a proper transcode to a robust RGB format.

To determine that there are several issues that need to be examined in a comparative manner to find any pitfalls or benefits in terms of image quality. The debate over the practicality of staying RAW longer is typically more about the particular production environment one encounters on a given project. It would be great if we could cite universally beneficial practices that would apply to any project but as the spectrum of comments on this forum proves, that's not reality.

The reality I see is that budget, turnaround time, sophistication of personnel, post tools, etc are all factors that need to be considered all the way to how a scene is lit and exposed. I also believe that native scene DR and whether elements in the frame with lower exposure levels need to "play" in the finish have to be considered in choosing a strategy. For example, for a dimly lit scene I might add more light to get a cleaner "RAW negative" and then restore the shadows in the grade to keep the noise from intruding. On a sunny exterior I might underexpose a little to hold more highlight information knowing that with all that blue light I should be able to dig whatever I need out of the shadows with minimal noise.

ISO on a digital camera can be counterintuitive, so I suggest we think in terms of signal to noise ratio. One of the things I noticed immediately with MX sensor performance in the Epic vs the R1 was the increased precision in the shadows (which I attribute to the added bit depth) changed the noise "signature", I would wager that's one of the things that Gunleik notices when comparing them. The lighter compression options made possible by the higher data rates also seems to give more texture to areas of relatively consistent density/chroma which I assume is due to less interpolating in the wavelet encode - but that's a Graeme issue that he may be constrained from discussing in detail publicly. In any case, to add a more seat of the pants perspective to all this tech talk, I have a story:

My friend, and legendary RedUser, Evin Grant, embraced the noise right from the earliest Mysterium/R1 days. He liked the texture and thought of it as digital grain. He loved the look of film and I think he found the "clean" digital negative too sterile, so he sought out some noise to add some guts and grit. He made some very evocative imagery with that approach, though his skill as a DP was probably more to "blame" (hehe) than his ardor for noise. I bring this up, because at the time I was firmly in the other camp trying to kill noise because I disliked the way the type of noise we had in 2008 seemed to look roughly the same no matter what the other image metrics might be - it was easier to see and harder to kill under 3000K, but its character was roughly the same as in dark shadows under daylight.
With the introduction of the MX sensor and the concurrent release of new color science (RC2/RG2 IIRC) the noise "character" changed rather dramatically. Instead of chunky it was now a smooth gray bottom end and, even better, we had another 2 stops of genuine DR before it even went into the noise floor. For grain aficionados like Evin, there were now several post plugs that could dial in the amount and type of grain desired - I believe there was even one company that used scans of popular film stocks to create a random noise pattern that could be comped in during DI/finishing.
With the Epic/MX package I am now seeing noise "signatures" that track more accurately to the specific scene and subtle differences in the deep shadows. This makes me more comfortable with letting the noise "play" in certain material for some of the same goals Evin had. I still think adding random grain in post is a better way to get that grittier look, but that takes some serious computational horsepower in post AFAIK. In any case, its worth noting that all noise is not created equal and so determining a specific numerical value of S/N in db is not as purely empirical as we camera nerds would like.

FWIW, even with the newest color science I still pop on a 1/4 CTB anytime the light sources are under 3000K. I know that Graeme has gone to great lengths to optimize the reconstruction of balanced images even when the color temp is near the extremes, but it seems to help (perhaps its just my bias clouding my judgement). I bring it up in this thread because I believe it makes the choice to transcode to an RGB format at the one light stage easier to do.

Cheers - #19

Yup, that's where I started off.
I am happy to see this thread developing into something valuable and non-dogmatic.
There are a lot of RAW issues treated here in very informative ways. And it is great to hear different peoples approaches!

Thanks Blair
(And Graeme and everyone else contributing)

BUT I'll go back to the original subject a bit later on, as there seems to be some confusion as to the signalchain/pipeline. WHat better then than to try to document it.

My main point is that transcodes/non-transcodes perform identical when done right, (or wrong...) but that there are good reasons - depending on project and environment - to choose one over the other, fully or partly.

Cheers!
G

[Marko Hocevar]

BUT I'll go back to the original subject a bit later on, as there seems to be some confusion as to the signalchain/pipeline. WHat better then than to try to document it.

Great thread. Thanks!
I hesitated to write anything because it wouldn't follow the main theme of the thread, but this "invitation" has pushed me over the edge.
I often found the discussions on RED workflow to ambiguous for my taste - and as it seems that creates confusion. I wish we could build some axioms and bring more formalism into discussions, but that is a rant that deserves it's own thread.
Anyway, I'm still trying to come to peace with the following group of statements (which describe exposure & raw development basics):
1. ISO is just metadata and it doesn't effect raw image data
2. Higher ISO gives you more highlight protection and moves middle gray point to lower setting
(article on red.com even states higher ISO gives you more clipping protection - I consider this not to be correct if the first statement is valid - I base this on the fact that clip white point is fixed and kept the same regardless of the ISO setting)

First I would like to describe my understanding in case of two options:
a) I used high ISO setting and a light meter or screen&histogram to expose - in this case high ISO would give me additional highlight and clipping protection because I was using lower exposure. Stated differently ISO setting has indirectly influenced raw image data.
b) I used ETTR technique and RAW view to expose - in this case higher ISO setting does NOT give me additional clipping protection - it just gives additional highlight protection. In this case it is correct to say that ISO doesn't influence raw image data.

Of course after exposure is recorded changing the ISO will not have any effect/"influence" on raw image data in both a) and b).
Now it is also a good time to define what is meant by "highlight protection" after the exposure has been recorded ignoring the ISO setting. (obviosly option a) gives both highlight and clipping protection so here we talk only about option b)
For the purpose of discussion let's assume 13 stops of scene dynamic range is captured by EPIC/Scarlet.
We develop this linear scene capture with redgamma3 - meaning preparing it for display on a REC 709 device - meaning 13 stops of dynamic range needs to be compressed to i.e. 6 stops.
When we use a higher ISO setting we are slidding the possition of middle gray onto the lower stops of 13 stop scene DR. Then all the stops from the middle gray postion up to the clippping stop from 13 stop scene DR will be tone mapped to the 3 stops of higlights in the Rec709 6 stops DR. And the rest of the 13 stop scene DR will be mapped to the lower 3 stops of REC709 6 stops.
Now the more highlight stops we use in this tone mapping the more our highlights will be protected - color precision and gradations will be better/nicer.
Vice-versa the less highlight stops we use in this tone mapping the worse our highlight will look on our REC 709 displays.
High/Low ISO - more/less highlight stops go into tone mapping and better/worse the highlights look like.

Is the above what is meant with highlight protection?

3. Can there be any "perceptual" highlight clipping hapenning when ISO is raised for development?
According to Graeme, highlight information is kept even under "extreme" high ISO settings.

When we map high dynamic range data from the R3d down to the low DR of the display we use a tone curve. That tone curve "pegs" shadow and highlight points and remaps in-between. I can test this on fake sensor data with known detail in the highlights and see that such detail which may appear lost at first glance in a ISO6400 export is actually there still, although at a reduced precision than it is at ISO800 and lower ISOs offer no real increase in precision.

I just re-wrote to put a small pattern in the brightest area of my test data at 0.2 stops down from the main pattern and it's still utterly visible at ISO6400 and hasn't been clipped away. Yes, it's now represented at less precision, but it's still there and still visible (even to the naked eye without adjustment).

Graeme

Where would we plug into the story this statement? - After exposure is recorded with high ISO you can lower the ISO for development and reclaim the additional highlight stops.
I can only justify it if it means that we get nicer highlight gradation and more color precision in the highlights, but not if it means we are recovering clipped data. Well, I could even justify the un-clipping part if 3. is possible - but it contradicts Grame's response to 3.
All this was an attempt at unwinding of the ambiguous statements 1. and 2. (maybe I'll list some examples of different conclusions possible from those statement, but there are not that important)
I purposly ignored the discussion of shadows as that could become overwhelming to put into one post.

Do you see any flaws in my logic or some facts I got wrong?

[Graeme Nattress]

Marko, here goes.....

ISO doesn't effect the raw, it's "just metadata". But say you have a shot properly exposed for ISO800 and now set the camera to ISO1600. There is still no change in the RAW data, but the image looks over-exposed. Now you shut down the iris by a stop, the exposure looks right, but you now also have an extra stop of highlight protection, or, in other words, you have an extra stop of clipping protection in the highlights.

ETTR is a very different method of exposure setting. With that you monitor the raw data and set exposure to be just before the edge of clipping. Then the image will appear too bright or too dark, so you set the ISO accordingly.

With ISO, always try to relate your exposure to mid grey and it's recorded code value. When you increase ISO by a stop you lower the effective mid grey code value to a half of whatever it was previously. Now you have more code values above mid grey than you had before, and less below.

When we're talking about RAW development, we're mapping a large dynamic range linear light data set to a display monitor with limited dynamic range. To do this we use curves, and there's basically two options - a log curve (think REDLogFilm) that preserves all the DR, but the end result is very low contrast, or we increase the contrast to make the image look "right" (think REDGamma3) but either clip (as in loose) code values in the highlights or shadows, or compress the highlights and shadows, not throwing away information but reducing it's precision.

When you develop at a lower ISO you're not avoiding clipping as our curves generally don't do that, but you are re-distributing precision. That means what you may think it hard clipping in a REDGamma3 export is not actually clipped, just reduced in precision making it hard to see (as we need contrast to see) the tonal values you have stored there.

Graeme

[Blair S. Paulsen]

Perhaps Graeme or others will kill this theory but I think its worth considering:

(Note: this approach is based on the desirability of a "fat negative", which is perhaps more accurately a bigger data set.)

Put the scene values of greatest interest in a part of the exposure that has the most precision.

If I understand the linear assignment of code values correctly, the farther up the scale you go the more mathematical precision available. This would support the ETTR with no RAW clip approach to setting exposure. Based on personal experience I would have to say that some of the most robust (able to handle aggressive grading) images I have shot were at least 2 stops lower than a pure ETTR exposure would have been. Is that simply proof that if there is enough latitude so that nothing clips, or dies in the noise floor, that its really about all the other factors (lighting, etc)?

Perhaps only a scene with 12 stops of potentially interesting scene values should be considered in this discussion. If you know the audience is highly unlikely to see more than 8 stops in 99% of presentation scenarios do you bias your whole pipeline to service that?

To get back on topic, one might argue that if the exposure setting during acquisition is aimed at maximizing the discrete code values in the data set (fat neg) then doing a one light adjustment and creating high quality RGB media where middle grey, d-min and d-max is where the DP intended might be the holy grail.

Bottom line - IF, big IF for some, you have a DP and "first light" colorist (a savvy DIT is fine) working together to craft a robust RGB file that will travel through post; IMHO you will get the most consistent results with very little of the "RAW goodness" wasted. While paying for a good "first light" pass may seem like adding more cost to already stretched budgets, for shows with other image elements like VFX/crash cam footage/etc it should end up being less total expense.

Cheers - #19

[Graeme Nattress]

ETTR is not about precision, it's about noise. Darker code values are noisier because in a sensor, read noise is a predominant source of noise. I know some people, particularly my friend Michael Reichmann support the "precision theory" of ETTR (see http://www.luminous-landscape.com/tu...exposure.shtml) but really, it's incredibly rare (not sure off the top of my head of a camera that represents this) that the AtoD precision or the linear distribution of code values is the issue. We ETTR so that more light from the shadows hits the sensor and gets recorded at a code value that represents electrons in the sensor well that is less influenced by read noise when read out and sent to the AtoD.

Graeme