New idea (?) for using CDLs with LUTs in IPP2

[Michael Tiemann]

After spending some time looking carefully at a wide range of inputs to Phil Hollands philmColor LUTs, and also taking a much closer look at the functionality of 3D LUT creator, an idea occurred to me. But first things first...

philmColor LUTs are designed to fit into the Creative LUT slot of the IPP2 image processing pipeline, which means after initial RAW development (Kelvin, Tint, ISO, Exposure Adjust) and after CDL processing, but before Contrast and Curves. An important point that Phil makes is that most of the LUTs are middle-tone gray neutral , meaning that a Log3G10 value of 1/3 will remain 1/3, whereas darker and lighter tones will be pushed or pulled according to the LUT. In case it's not obvious, that means it's pretty damn important to know where middle-tone gray is in your input image, and also pretty damn important to use appropriate Kelvin, Tint, ISO, and Exposure Adjust parameters to put that middle tone gray value where it belongs. WHICH IS NOT TO SAY YOU CANNOT GET COOL RESULTS BY BREAKING THIS RULE! But, as with many things, it's best to know the rules before you go around breaking them. And knowing how the LUT and the image agree or disagree on the central fulcrum is important, I think!

But there's more we can do than just lining up the middles and having the LUT express itself with the balance that Phil intended. A second thing we can do is to squash or stretch the contrast range so that the LUT has a greater or lesser effect on the shadows and highlights of the image. If we know that the lift or the crush or the cooling or the warming really starts to take effect 3-4 stops away from middle-tone gray, we can use a CDL to send more or less of the image through the strongest part of the LUT. We can even do this asymmetrically, creating a larger effect on the dark side of middle-tone gray while having less effect on the bright side, or vice versa. We can then use the contrast and curves to effectively invert the CDL, keeping the effects of the properly stimulated LUT while returning our Log3G10 image data back to "normal" exposure.

It very much seems that this is precisely the idea of Stage 2 of the IPP2 pipeline. What is missing is some guidelines on the basic best way to use the parameters of the CDLs to most effectively aim at the most active regions of the LUTs in question. Yes, of course one could just push parameters around until things look right. But rather than taking a trial-and-error approach, I think it would be most interesting to know that if the LUT's greatest 2nd derivatives are at 22 and 77 IRE, and if most of my image ranges from 15 to 66 IRE, with a large bulge toward the left, not the right, what's a sensible starting place to normalize my image for my LUT, and the best way to then renormalize the image back to the tonal range I shot? Again, I'm not expecting math to solve this problem 100%, but rather knowing enough basic math that I can get to a good place where parameter changes I make from there have sensible, rather than surprising consequences.

 

[Michael Tiemann]

Since nobody took the bait (and gave me the simple answers), I had to go and learn a few things for myself. Here's what I've learned:

1. The most important controls for making RED images work with philmColor LUTs (or other creative LUTs) are Kelvin, Tint, ISO, and Exposure Adjust. If the input image can be brought into "normal" range using primary RAW development, then problem solved. In general, whether one is exposing to the right (DRAGON), exposing to the left (Helium), shooting low-key, or shooting high-key, if one aims middle tone gray at a Log3G10 value of 1/3, then the image is hitting the LUT right in the sweet spot. There are plenty of options post-LUT to make things extra-dark, extra-bright.

2. Failing that, if the tonal range of the image is somehow compressed (that never happens on REDUSER.net!), then increasing the SLOPE of the CDL above 1.0 will both raise the value of middle-gray and extend the dynamic range further into the highlights. However much SLOPE is applied to reach the tonal range goals, middle-gray must then be adjusted. A SLOPE greater than 1.0 could bring an under-exposed middle-gray up almost enough, too much, or just the right amount. Such a SLOPE will bring an over-exposed middle-gray well past the 1/3 value. Regardless, the best way to re-align middle-gray is with ISO/Exposure Adjust, not OFFSET, since negative offsets will do strange things to the left side of the histogram, especially if the subtraction of OFFSET results in zero or negative values. And because positive offsets might cause highlights to clip or high-intensity colors to go way outside the gamut.

3. Alternatively, if the tonal range of the image is so wide that the LUT is doing radical things that are neither technically shadows nor highlights, then decreasing the SLOPE of the CDL below 1.0 will lower the value of middle gray and compress the dynamic range. However much SLOPE is applied to reach the tonal range goals, middle-gray must then be adjusted. A SLOPE less than 1.0 could bring an over-exposed middle-gray down almost enough, too much, or just the right amount. Such a SLOPE will bring an under-exposed middle-gray well below the 1/3 value. Regardless, the best way to re-align middle-gray is with ISO/Exposure Adjust, not OFFSET, since negative offsets will do strange things to the left side of the histogram, especially if the subtraction of OFFSET results in zero or negative values. And because positive offsets might cause highlights to clip or high-intensity colors to go way outside the gamut.

Not that any change made using SLOPE to adapt the image for the LUT can be trivially reversed by using a reciprocal CONTRAST curve that is applied post-LUT in stage 2. Thus you can get the full coloring effect of the LUT even on very compressed tonal ranges, if that's your thing.

I leave the use of the POWER parameter as an exercise for the reader. My thinking is that there's not a trivial (i.e., mathematical) way to under the effects of a POWER function. Yes, an S-curve can undo the application of POWER, but you are on your own to draw the S-curve curve that undoes e.g. POWER=1.2 (or whatever). If you set a SLOPE of 1.25, you can undo that mathematically with a CONTRAST of 0.8.

So, the answer is: for most well-behaved images, Kelvin, Tint, ISO, and Exposure Adjust puts your image where the LUT expects it, and you can grade from there. In what I think is the next most common case, SLOPE (plus Kelvin, Tint, ISO, and Exposure Adjust) before the LUT and CONTRAST afterward will optimize the image response of the LUT while giving you the dynamic range you always intended.

If you want the LUT to apply much more strongly to one side of the dynamic range or the other, you can use POWER > 1 to cause more of the shadow side of the LUT to affect the image (while the highlight side has a reduced effect), or you can user POWER < 1 to cause more of the highlight side of the LUT to affect the image (while the shadow side has a reduced effect). In either case, you have to invert the POWER function using an S-curve that you draw yourself, which might not be the perfect mathematical inverse of the POWER parameter you set. Also in either case you can use SLOPE and/or Kelvin/Tint/ISO/Exposure Adjust to complete the desired impulse response.

I hope that helps somebody else pondering these questions. It helped me to write it up!

 

[Phil Holland]

Tapping in as this possibly involves me

So when I said philmColor is mostly focused on maintaining a consistent 18% gray (less so with some of the bolder looks found in R2 btw) what I should have also expanded on is generally they are focused around keeping the same general exposure values.

baseStocks will break that rule occasionally, but rarely.
ghostsStocks will always adhere to that rule.
printModels will break it depending on the "strength", though 18% is generally there.
toneAdjusts will push and pull as in their general descriptions.
achromicStocks are sort of their own thing for black and white.

The reason I have such a massive collection of test images and sequences from I don't even know how many people is there's a good deal of people out there who expose fairly differently than I do. One of my favorites is a guy who's usually 2-3 stops under key for a fairly moody style, but I have far crazier R3Ds for sure both on the top end (exposed heavily to the right) or the other way around.

So with all that in mind, I'll tackle the questions.

The most important 'control' for making RED images work with philmColor is A. To shoot images. B. Expose however you see fit for a given scene. That said from there usually ISO and Exposure Adjust work great if you need to generally adjust the exposure after the fact. I'll say this with the caveat that my general POV is to exposure more accurately than not. That's more of advice for people shooting in general. As you described there's things that happen before the Creative Cube to help you dial in your Exposure and Color Temp, all good. After the fact the Curves and CDL as well as the Output Tone Maps can be used to dial in the look further. This is all part of the Grading Pipeline in IPP2. Ideally you can use it "however you want". With CDLs in particular, outside of the common controls, it's a good place to dial in your saturation.

It's as simple as that overall. As always pay attention to your scopes if you're going deeper into a grade. For instance if you've landed on a flat or lifted black philmColor preset you can dial "where that lands" pretty easily with a quick curve, often 1 or two point adjustment.

 

[Hrvoje Simic]

Not that any change made using SLOPE to adapt the image for the LUT can be trivially reversed by using a reciprocal CONTRAST curve that is applied post-LUT in stage 2.

It cannot.

Because there is LUT between those two transformations additionally modifying the image and later in the signal chain you affect the joint outcome.

Thus you can get the full coloring effect of the LUT even on very compressed tonal ranges, if that's your thing.

Doing that can easily bring oodles of noise, artifacts and undesired signal deviations.

I hope that helps somebody else pondering these questions. It helped me to write it up!

There is creative range both before and after but mostly after so take caution.
Especially since it is a particulars-oriented, input/output-defined image transformation. The more complex the image transformation by the LUT the less room for manipulation before it.

Make sure you don't drift too far before, you may not be able to compensate that afterwards adequately.

 

[Marc Wielage]

The most important controls for making RED images work with philmColor LUTs (or other creative LUTs) are Kelvin, Tint, ISO, and Exposure Adjust. If the input image can be brought into "normal" range using primary RAW development, then problem solved.

This is also the idea most colorists I know generally do with Raw material: we "normalize" the Camera Raw settings in the color-correction program, making sure the exposure is correct and the color temperature is correct, so we then have the range to take the image the rest of the way necessary. Whether you start correcting from RedLogFilm or Log3G10 almost doesn't matter, as long as this starting point is reasonably flat and doesn't sacrifice any image detail. You can use a LUT or color-correction controls to get it to Rec709 (or whatever the delivery color/gamma space is), but there is no real magic there.

Better people than I have pointed out: the real creative part of color correction is the colorist, not the LUT. Don't think that a colorist can't duplicate that look, because they can. Noted Australian colorist Juan Melara on his website has some interesting thoughts about LUTs and when they can be precisely matched in the color-correction process:

http://juanmelara.com.au/blog/rebuilding-the-600-dollar-linny-lut-for-the-alexa-plus-free-download

 

[Phil Holland]

Better people than I have pointed out: the real creative part of color correction is the colorist, not the LUT.

I do wholeheartedly agree with that.

If you're curious my "metal stuff" in the toneAdjusts are variations of that particular style of grade without breaking the footage from say MX, which the Linny LUT will do a good chunk of the time.

In the case of these particular LUTs however a great deal of work has gone into scanning, studying, deconstructing, optimizing these looks for most material shot with all RED cameras at nearly every exposure, and producing them for RWG/Log3G10. That has been much of the work here.

At it's core philmColor is designed to be a "digital stock" concept that you can grade underneath and even above to a degree to refine your image. Quick to deploy, quickly transforming the footage into a consistent look that hopefully "just works" for most.

But you can't replace a good colorist. Especially in most typical production scenarios.

 

[Hrvoje Simic]

Whether you start correcting from RedLogFilm or Log3G10 almost doesn't matter, as long as this starting point is reasonably flat and doesn't sacrifice any image detail.

Generally speaking with similar gammas I'd agree but in the case of RLF and Log3G10, it matters more.

You can use a LUT or color-correction controls to get it to Rec709 (or whatever the delivery color/gamma space is), but there is no real magic there.

Oh yes there is.

: )

Better people than I have pointed out: the real creative part of color correction is the colorist, not the LUT.

Of course, but:

Don't think that a colorist can't duplicate that look, because they can.

That would pretty much depend on what that LUT actually does. Post production image manipulation depends on principles of image shaping tools and environment. They don't allow achieving everything which can LUT carry and vice versa.

Also the questions arising from that are: a) why would one want to waste time on that process if there is a more practical way to get to a specific goal and b) why would one want to copy someone else's look in the first place.


Especially because of things like these:

In the case of these particular LUTs however a great deal of work has gone into scanning, studying, deconstructing, optimizing these looks for most material shot with all RED cameras at nearly every exposure, and producing them for RWG/Log3G10.