Aaron Green wrote:How do you guys know that the 2 methods don't produce similar results? I've tried both that are mentioned and 9 times out of 10 they're very similar. Only shots that may be an exception are very bright backgrounds.
Mihail Moskov wrote:Aharon Rothschild wrote:Let's look at overexposure on skin for a minute. Yes it's less of a problem than underexposure but the UM4.6 and every other manufacturer is playing highlights vs shadows vs midtones on the sensor, it's a balance between how much range on the curve we give each at base ISO. The toe and shoulder on that curve are designed to roll off highlights etc but the center of the curve is setup to maximise information in skin ie. Midtones.
You are mixing up film and tone mapped delivery curves (i.e. rec709 with a soft clip, or anything else meant for delivery) on one side, with linear and log signal (meant for post). There is no shoulder with any linear signal, or any real log curve (except Cineon and its derivatives, which are supposed to encode film scans and do replicate the shoulder of the negative): there is no shoulder over log-exposure on Arri's LogC, or Sony's S-log, or any other log curve with a true log upper part.
The consequence is that with most digital cameras shooting raw or log you will get the cleanest signal using ETTR (and yes, this includes the best and densest skin). Whether this is the best approach for your colorist, who will then need to match exposures shot-to-shot, is another story.
Thanks so much for bringing this up as it's a great way for us to get deeper on this! I'm advocating two things
That DP's use a meter or false color to hit target exposures they arrives at through lighting tests
They don't subscribe to the oversaturating the sensor and backing off of clipping in zebras philosophy as a means of exposure.
Lets do RAW first
The ETTR approach as you outline is to not use the last two stops of dynamic range and to place skin two stops higher on the sensor (linear, and I was oversimplifying earlier but I'll get into that in a sec) so as to avoid the noise floor. This is effectively just redistributing dynamic range, and to use Alexa as example moving the ISO down to 200 for +5 and -9 stops (roughly) and is a effective way to avoid noise. When we stop to think about how this will actually work in practice though things change. We shoot two stops over, avoid noise, bring it in and lower it two stops. Fine, great, clean image. But as we apply contrast and expand the data through the range we see that that redistribution is a problem in the brighter parts of skin. Yes shadows are clean but we really would have liked to have more range on the face because as the sunlight wraps around it looks a bit too pale and not dense enough and that's because we have two less stops of range to play with for contrast above midtones, while balancing out the sky we need to squeeze skin into a smaller range. We've only captured 5 stops above middle grey and that's the tradeoff.
Next we see sensor-linear, to Log-C ARRIRAW- LOG, to a REC-709 or camera or normalization LUT, we have a sensor conversion designed to maximise data in a RAW container with LOG and another conversion to look like film or the way a human eye sees the world with the inverse curve or the camera LUT. Here we see a problem because the whole thing is setup to accept a image where the middle of the curve matches skin. No one I know has ever watched a feature in LOG-C (although some directors might like to:)) so we need to talk about ETTR in this context of watching the thing in the real world on a display and when we come in to this very fine tuned workflow with a image that is two stops over we are landing skin in a place that is incorrect. The only way to get it back to where the workflow is expecting it at 18% grey is to window out skin in every shot or to shoot a grey card for every shot. Well.... Wouldn't it be great if camera manufacturers had thought this through and made it easier on everyone where they matched all these complex input and output LUTS with settings that allowed for less noise? Yup. Just shoot at 200 ISO. The entire reasoning behind ISO on every camera not the Varicam (which does dual ISO on the hardware level) is exactly this. The base ISO as we all know is what the camera is reading on the sensor level and the lower ISO's are just gaining up and down.
Case in point when dragon was released it was supposed to be a base ISO 2000 camera, after noise issues the conventional wisdom became "shoot at ISO 320 with STHL OLPF on Dragon" entirely as a way to avoid noise.
NON RAW FORMATS:
Anything uncompressed means stretching and changing values as we gain up and down and that's very bad for the image, banding color changes etc Also we see the same issues as before with landing middle grey as expected for the conversion luts etc, and more of a issue as we use log to compress highlights to fit smaller containers. This is really a non starter at all.
Again I'm not saying there is any issue with bumping up a stop to get a cleaner image, thats been happening forever, I'm arguing against using zebras as a exposure tool and backing off from clipping as I feel that landing midtones in a way where we don't know how to get them back to midtones leads to real issues in the color workflow and that we should use ISO or adding a stop on the meter for this.
Would love your feedback and we can definitely get into raw log and bit depths for the real fun;))))