1 Like
The difference is how far to the left is the toe. Base curves I’ve seen finish the dip almost a quarter into the histogram. The original Duiker filmic curve when plotted in full in normal display resolutions doesn’t show the toe, you have to zoom in on the lower tenth or so to start to see it.
Yes that’s no surprise that they have different shape. Interestingly my criticisms of dt filmic is that the toe does what you claim the Duiker doesn’t. I find it difficult to make the toe small enough while the falloff large enough using filmic. You have to use other tools to get there which again is sort of strange considering the tonemapping goal of filmic.
Where you prefer your crushing to be has nothing to do with math, math comes into play deciding how to crush your shadows or burn your whites.
As far as this image is concerned, first thing I did was add 2EV exposure (so 2.7EV total) to “fill the histogram”. So this image was certainly not ETTR looking at the raw data (causing part of the noise problems in the shadows).
However, when applying the Nikon basecurve, just correcting the in-camera exposure compensation is already too much… And let’s not talk about what happens to the shadows…
Afaik, that basecurve mimics the result of the default in-camera processing (as represented by the embedded jpg).
1 Like
Sorry I wasn’t communicating clearly I was referring to my photos not this example.
No, the sooc curve toe goes further in. Its a custom curve with extra small toe. The file has also been tonenapped using a log module. Thats why I thought it was a good comparison to filmic. Similar, but not identical, modules. With the normal more flexible curve you can do without local editing.
I think there’s a different between bit-depth and dynamic range. 8-bit sRGB, because of non-linear encoding, can encode values down to and slightly under -12EV, not -8 (lower threshold under darktable 4.0 user manual - clipping warning and also the reason for filmic’s display target black being not 0, but also not 1/256 - target black luminance under darktable 4.0 user manual - filmic rgb).
1 Like
Are the axes scaled the same way in those comparisons? The same curve with different visualisations:
1 Like
Are you asking if dt filmic and RT tone curve can be visually compared? I don’t know but I assume, perhaps wrongly, you can’t. My intention was to phrase it in relative terms.
When I talk about the size of the toe I mean how blacks are compressed not literally comparing curves. But perhaps the filmic curve is just a tone curve. If thats the case I’m even more confused about the choice to lock it down.
Yes, it is - centred around the mid-grey, and you define what range you want to cover on the x axis (black and white relative exposure), how steep the mid-section should be (contrast), how wide the constant-steepness section should be (latitude), and where it should be located (shadows/highlights balance).
Much of the discussion under New Sigmoid Scene to Display mapping is how such curves should/could be implemented. Aurélien gives his reasons in that topic, others suggest different approaches.
2 Likes
Yes, but in darktable it also functions as the gateway between scene-referred algorithms (all processing before and up to filmic) and display-referred algorithms (all processing after filmic).
The first non-linear tone curve of any type in the processing pipeline is that gateway.
Recently, I’ve been playing with other tone curve strategies, particularly a log curve followed by a contorted control-point curve. Shooting into the sun for instance produces an extremely large dynamic range, one that can only be tamed with decidedly non-continuous transfer functions (curves are a transfer function). So, what I’ve tried in these cases is to first apply a log curve that lifts the shadows out of the bottom of the histogram and spreads them out a bit, followed by a tortuous control-point curve to lift where lifting is needed, depress where depressing is needed. What I’ve found is that if I pay attention to where I put the steep slopes, I can almost tone-map HDR-wise. The strictures of the spline curve still constrain me, so the success varies image-to-image.
All of it is just messing with the linear recording of the camera where it doesn’t satisfy our whims…
By abusing the black relative exposure and setting it to -14EV you can lift the shadows to decent levels.
This is maxing out though and a bit more would have been useful. It also seems to me a bit odd for this d7100 to have 19EV dynamic range 
I guess one should just push the sliders until the results appear but the scene tab and those exposure values feel sort of real.
DSC_2481.NEF.xmp (38.1 KB)
edit: exposure is also maxed out at +3EV
A question, as a beginner: couldn’t you have a curve (transfer function) that does not max out at 100%? One could (I think) apply such a curve (even if non-linear) and still remain in scene-referred space.
You can right click the slider and type in a value.
That gives us more options! Those relative exposure can look a bit more sane as well.
So it’s kind of the same old filmic lesson. Just mash that exposure beyond what you would expect.
The Duiker filmic equation doesn’t top off at 100%; I had to add a ‘normalize’ option to get it to do so. However, it’s not that characteristic that affects linearity; it’s the departure of the function from a straight line anywhere in the tone range that does that.
‘linear’ refers to the energy relationship of the light values in the scene. Any curve that isn’t a straight line (gee, what goofy semantics there… ) causes values to lose that original energy relationship.
Right. But my understanding is that linearity and scale are orthogonal concepts – in other words, one can have linear (or non-linear) data in each space. Or does display-referred imply non-linearity?
I’m probably going to biff this, math people are welcome to correct or clarify…
The recording of light by the camera is done with respect to the light’s energy relationship, that is, producing numbers whose magnitude corresponds to the energy magnitudes, e.g. 2x a measurement yields a number that represents twice the corresponding energy. Scaling doesn’t disturb that; an EV adjustment is an example of an operation that changes the numbers but preserves that relationship.
When you say, “have linear (or non-linear) data in each space”, I can posit a transfer function that leaves part of the tone range alone but curves another - is this what you’re talking about?
“Display-referred” is definitely about non-linear, either our perception, or display characteristics, which both depart from the energy of light in a scene. Well, a LCD doesn’t have to, I’m told, but I can’t make mine behave… 
Your edit is the type of edit I attempted in the start of this thread. I think that this example shows that it’s quite expressive what you can achieve with filmic by pushing it to the limit(s)…
. Of course, you have do some further editing tuning the walls and the roof.
I forgot to mention that you can key in a black relative exposure value of max -16…
1 Like