Highlight Roll Off
- Thread starter sgttom
- Start date
Cary Knoop
Veteran
There are two kind of highlight roll off situations, one in-camera and one-in post.
Good in-camera highlight roll off is when there is a smooth transition from light to (almost) clipped light levels. Bad highlight roll off is like having sharp white holes in the video.
Highlight roll-off in post may be used to change good highlight roll off gradations but in addition highlight roll off is needed to map large color space volumes, for instance a log recording, onto small color space volumes, for instance Rec.709.
In order to maintain realistic contrast in the mid-tones one needs to roll off the black and white extremes.
I am really don't fully understand all this so bear with me.
I heard a couple people talk about using diffusion filters for better roll off, which is really just making the highlights glow and leak into non clipped areas. That isn't really the same as nice highlight roll off though is it?
Cary Knoop
Veteran
Exactly!
From what I understand the reason the Alexa seems to capture better highlight roll off is because it captures more information in highlights then in the shadows. So there are more stops until clipping then most cameras, which results in a more gradual roll off? or is there more to it then simply dynamic range in the highlights?
morgan_moore
Major Contributor
This Alexa claim has always mysified me. In recording any scene the operator can always iris/ND down until the highlights are protected.
The concept of a camera capturing '10 stops above mid grey and 6 stops below mid grey' is a flawed one to my mind. The operator chooses at what level to put mid grey and sorts out the visual brightness in post by using an off speed LUT or exposure sliders.
If you are an arrogant and ignorant DP coming from film who refuses to meter in any manner but by putting an incident meter by the actresses face then one is not shooting to protect ones higlights but nailing mid grey at a certain point.
The arri could be described as being well set up for arrogant and ignorant DPs.
That does not preclude it from being an excellent camera of course - logically it is an excellent camera - it has a small amount of large pixels and can record a bit depth of 12 (or more?) - I have issue with the description of a 'stops above mid grey' spec- not the camera.
----
As for highlights in general there seem to be many factors at play.
The scene DR, the operators exposure choices, the cameras core capability (DR of the chip) the cameras conversion ('scene file) of the core data to the codec it records (compression and bit depth) before hitting post where a poor LUT or other interpretation of the image data may clip data or embugger it in some other manner.
So highlight rendering is lead by..
Scene
Exposure choice
Camera inherent capability
'scene file'
Camera codec
Choices in post.
Last edited:
morgan_moore
Major Contributor
If one seeks a filmic image - this can be a legit approach. Film when over exposed 'bleeds' the white into surrounding grains - diffusion can kind of replicate this effect - Im not personally convinced that doing it in camera is the wisest approch though.
You are both correct in thinking that this does not inherently give the camera better highlight handling per se. Better? No. More visually appealing? Possibly.
morgan_moore
Major Contributor
Yes it can. It might be less critical than many imagine. Ultimately post can never get back what was not recorded in the first place though.
AndreeOnline
Veteran
Each pixel on the imaging sensor is made up of a small photo diode that captures light (incoming photons that create an electrical charge). The photo diode can be compared to a well that you can fill with water. Each photon would be a drop of water and as long as you're exposing the sensor, you're adding drops to the well. A full well represents the max charge for that diode and that, in sensor terms, determines the max dynamic range you can achieve. But as you know, the bottom of the well has sand, dirt and sediment, so the water at the very bottom can't really be used. Same with the sensor: the "photons at the very bottom" will get mixed in with electrical noise, often referred to as the noise floor. So, the 'usable dynamic range' is determined by the user who will "drink the water in the well". At some point close to the bottom we all stop drinking, as it's just to muddy.
The reason I frame it like this, is that it's important to understand that the entire top half of the well—all that clear water— is made up of the single last stop of dynamic range. One single stop. It's light's exponential nature vs how it's captured: linearly on the sensor level.
A RED frame as seen through IPP2 (medium contrast and soft roll off):
The same frame, as seen by the sensor in linear light:
And now to the actual point... sorry: the reason why great highlight roll off is so luxurious in nature, is that in order to have a nice and gradual roll off at the top you need to compress a few stops of light so that we get a gentle transition to 'clip'. But that means that not only to we take the whole top half of the well, but we keep taking more and more of that "clear water", only to reserve it for "roll off". Meanwhile, the actual meat of the image will come from water closer to the bottom of the well.
The way image data gets translated from the dark image above (linear) to the brighter image, is via a transfer function. Normally we just call them gamma curves, like LogC, Canon Log, S-Log and so on. In the above example it's RED's Log3G10 plus a contrast LUT. Here's the straight Log3G10:
Not all camera systems capture data that is robust enough to have "data to spare" for lots of roll off. More basic camera systems don't have logarithmic curves but use more contrasty picture styles.
If you have a clippy, contrasty image, I think adding some glow or blooming in post can help with the top transition, or at least sell the illusion. At the time of capture you can either shoot RAW, or if you bake your image, at least use a Log profile (if available) to help redistribute the captured light intensity more appropriately.
Note: a log profile doesn't necessarily increase the dynamic range of an image, but it does control how light gets distributed. A standard picture style will have severely compressed shadows and highlights, and those are often captured in 8 bit. It's simply not possible to expand the shadows or pull down the highlights in post here, because the lack of data would cause excessive posterisation/banding.
The reason I frame it like this, is that it's important to understand that the entire top half of the well—all that clear water— is made up of the single last stop of dynamic range. One single stop. It's light's exponential nature vs how it's captured: linearly on the sensor level.
A RED frame as seen through IPP2 (medium contrast and soft roll off):
The same frame, as seen by the sensor in linear light:
And now to the actual point... sorry: the reason why great highlight roll off is so luxurious in nature, is that in order to have a nice and gradual roll off at the top you need to compress a few stops of light so that we get a gentle transition to 'clip'. But that means that not only to we take the whole top half of the well, but we keep taking more and more of that "clear water", only to reserve it for "roll off". Meanwhile, the actual meat of the image will come from water closer to the bottom of the well.
The way image data gets translated from the dark image above (linear) to the brighter image, is via a transfer function. Normally we just call them gamma curves, like LogC, Canon Log, S-Log and so on. In the above example it's RED's Log3G10 plus a contrast LUT. Here's the straight Log3G10:
Not all camera systems capture data that is robust enough to have "data to spare" for lots of roll off. More basic camera systems don't have logarithmic curves but use more contrasty picture styles.
If you have a clippy, contrasty image, I think adding some glow or blooming in post can help with the top transition, or at least sell the illusion. At the time of capture you can either shoot RAW, or if you bake your image, at least use a Log profile (if available) to help redistribute the captured light intensity more appropriately.
Note: a log profile doesn't necessarily increase the dynamic range of an image, but it does control how light gets distributed. A standard picture style will have severely compressed shadows and highlights, and those are often captured in 8 bit. It's simply not possible to expand the shadows or pull down the highlights in post here, because the lack of data would cause excessive posterisation/banding.
morgan_moore
Major Contributor
Yep - that last stop!
I didnt know a good way of illustrating that.. but I did allude to that with the 12bits or more comment.
When you are trying to distribute stops over 256 levels (!) 10 bits or even 1024 levels you have limited options for allocating a load of values to the last (brightest) two stops, - with 12bit 4096 levels things start to open up and with 16 bits (65,000 levels) you have proper options to create a camera with a fine highlight roll.
Oddly in 2004/5 I bought a haselblad digiback. It was the first digital device I saw that had decent highlights.. its a 16bit camera. Affordable motion cameras still lag a mile behind that - so sketchy highlight roll should only be expected at the highest level of motion cameras.
I didnt know a good way of illustrating that.. but I did allude to that with the 12bits or more comment.
When you are trying to distribute stops over 256 levels (!) 10 bits or even 1024 levels you have limited options for allocating a load of values to the last (brightest) two stops, - with 12bit 4096 levels things start to open up and with 16 bits (65,000 levels) you have proper options to create a camera with a fine highlight roll.
Oddly in 2004/5 I bought a haselblad digiback. It was the first digital device I saw that had decent highlights.. its a 16bit camera. Affordable motion cameras still lag a mile behind that - so sketchy highlight roll should only be expected at the highest level of motion cameras.