Thanks for posting. As the module is still in progress, I do not provide migration of params, so everybody, please don’t be surprised if edits break between versions.
3 Likes
Don’t worry, I understand that you’re in the R&D phase, but the results are really promising and they’re arousing my curiosity.
I can see one or two sliders such as ‘clarity, texture or detail’.
1 Like
Hi, at this point no such sliders are planned. AgX is just a pixel-wise curve (with some pre-baked rotations), no sharpening or anything requiring context (e.g. the surrounding pixels or the whole image, etc.) is done.
Not just for Christian:
I have been thinking about splitting it, actually:
- separate out the log mapping: provide the relative black and white exposures, produce a 0 → 1 log-encoded response. On top of that, one could simply add any curve (e.g. a standard ‘base curve’ / ‘rgb curve’ / ‘Lab curve’), or a parametric one, such as a sigmoid
- provide the tunable sigmoid as a separate tool.
I may do that later. Right now, I’d like to get the maths to solve that reversal issue. I have a workaround, but that changes the contrast; what I’ll need to do is to curve the toe concave instead of convex (or the shoulder convex instead of concave). Sorry about the horrible sketches:
Probably trivial maths for others - if I can’t do it, I’ll ask for your help; for now, please let me struggle with this puzzle a bit more. 
I’ll let you know if I find the solution (smooth fitting to the linear portion, so fixed derivative at the join point; being able to take it from arbitrary (x0, y0) to (x1, y1) to connect the (0,0) or (1,1) to the linear portion) or if (when) I give up and need you to solve it for me. 
1 Like
Sure. I don’t want to burden the forum with large test files, so you can find them next to the executables:
https://tech.kovacs-telekes.org/dt-agx/2025-03-22-13-27-34-P1060905.RW2
https://tech.kovacs-telekes.org/dt-agx/2025-03-22-13-37-39-P1060919.RW2
I’ve added the default CC licence from PlayRAW, so have fun!
1 Like
Thank you!
Another question. Do we still need the Offset slider?
On my last test I merely used it but concentrated on black- and whitepoint setting.
What do the others think?
I’m a little busy this morning, but I’ll try to build a Windows version
2 Likes
Using Filmic Tonemapping with Piecewise Power Curves – Filmic Worlds
@Dave22152
You may want to wait a bit
@kofa - too late! 
Link to Windows executable:
1 Like
‘No problem’ (= thanks!), I need some time to code this up.
if you don’t want to make a version with sliders for each channel, perhaps an option to operate on all channels or just one, then people could use 3 instances if they want per channel control, this looks really interesting so far
Here’s another build. The only change is that it can now (hopefully) handle relative black/white exposure without weird side effects.
With the sigmoid curve that is in the original AgX, the toe and the shoulder are always concave / convex, respectively. That means they always have a slope that’s lower than that of the linear portion (the toe increases from a lower value and reaches the slope of the linear portion when it hits the toe transition point; the shoulder starts with the slope of the linear portion and then reduces it as it approaches the white limit). If the slope of the linear section is too low, we may not hit black/white; moving the white/black relative exposure closer to mid grey clips and darkens highlights / crushes and brightens blacks. This is illustrated above.
The solution is to detect if the slope is too low to reach black and/or white, and switch to a convex toe / concave shoulder. Technically, they can hardly be called toe and shoulder any more, and the curve is not an S shape, but it avoids the ugly artefacts.
So, instead of something like this:
You get this:
This fallback is only applied if needed, and only to the end of the curve that needs it (independently from the other end). Otherwise, the toe/shoulder curve of the original sigmoid curve used in AgX is applied.
Note: this new part of the curve does not take target black/white into account, and uses 0 and 1, respectively.
Do does anyone really need those settings? If not, I’d like to nuke them. Fewer sliders, less complex code.
I’m still unable to build for Windows (I’ve applied the recent change from master, but it did not help).
The Linux AppImage is here: https://tech.kovacs-telekes.org/dt-agx/Darktable-5.1.0%2B549~g84f72fcab8-x86_64.AppImage.
1 Like
If you mean target black and target white, yes, they are very important for certain analog styles.
1 Like
@kofa I haven’t played with this yet, but based upon the feedback here I hope to see it one day in DT. Keep up the good work.
@kofa, the amount of work you’ve put into this proof of concept is impressive. I wanted to ask what your objectives were for this exploratory tonemapper compared with what we have in Filmic and Sigmoid. There’s some potential to streamline tone mapping and bring multiple controls “under one house”, but I’ve also read that AgX applications have some unique color management properties that we might not have right now.
I don’t know if that’s true or if that’s the goal here, but some thoughts on where this is heading might help us focus on what important for this POC.
Many thanks
1 Like
@kofa…sorry if this is an ignorant question or if I missed something…but comparing the controls of the current DT module with the curve mock up that you shared…to me the shoulder power is working the opposite way I would expect…usually increasing power drops the curve and it works that way for the global power and the toe but not the shoulders?? in the curve mock-up it also drops the shoulder down but in my testing in DT things go the other way?? I said dumb and likely I am…
I see here that you explained the behaviour…
It just seems different here…
but I might be interpreting this wrong…
The tranistion behaviour matches your explanation ie that the transition is sharper but the values seem to drop while in DT things seem to go up…
But looking at this now it maybe this is because the curve mockup is above 1 and in DT I am going from below 1 towards 1??
Lower power means gentler transition (the contrast starts dropping immediately past the end of the linear region / pivot, but it does not drop so much, some contrast is retained also for the topmost region):
Higher power, more abrupt transition (high contrast for a longer section above the pivot, close to that of the linear section; then quite quickly, the gradient drops, and contrast is lost for the topmost region):
Here, the power is 1:
The clouds are not that bright, and as contrast (due to the low power) starts dropping off immediately above middle grey, those ‘mid-bright’ regions are not pushed too much.
Here, the power is much higher. The high contrast assigned to the midtones is kept for longer, and the ‘mid-bright’ regions are shifted up more:
I hope I got your question right and was able to provide an explanation. If not, please let me know!
2 Likes
Here are some results from the Study in Blue play raw, with base adjustments to exposure and tone mapper settings in AgX, Filmic, Sigmoid and Filmic:
AgX
Sigmoid
Filmic
The roll-off of the lights and details in the shadows in AgX was very impressive. I struggled with Filmic and probably could have done better with more time, but I was surprised that I could get good results so quickly with AgX.
2 Likes
The start was just an experiment:
- how does a darktable module work?
- can I use darktable to play with colours, tone mapping, gamut mapping?
Then, there was this overwhelming response, people seem to like it. So now I’d like to see if, in some shape or form, it’d actually be useful to get into darktable.
The ‘colour management properties’, I guess, are the rotations and in/outsetting of the primaries. That is something I’d like to learn about. Currently, it uses a matrix I took over from EaryChow, who somehow processed the original matrices, created for Rec709, and ‘ported’ them to Rec 2020. I’d like to understand what the original matrices did, and how the port was performed. The sigmoid and primaries modules already have that, it’s nothing new in darktable.
Then come those ‘guard rails’ I’ve read about, contributed by @flannelhead , who also added the primaries to darktable, and has been patiently answering my questions and providing encouragement behind the scenes. I’ve also done some naive stuff of my own, but they are probably not practical to ever end up in darktable.
So now, the short-term plans:
- Fix the code so it actually takes the target black and white into account, instead of using 0 and 1. Not a big deal.
- Improve controls. One I really don’t like is the ‘toe/shoulder length’, for a number of reasons:
- it’s not about the length of the toe or shoulder, to the contrary: it’s about the length of the linear region, outside the toe/shoulder
- the control is not… wait for it… intuitive (and also not really relevant): each defines the length of a hypotenuse, and then what that means photographically (in terms of input exposure range) depends on the slope:
What would be a better option? Directly specify the range on the x axis (input exposure) as:- an absolute number (but what it means depends on the black and white relative exposures, and 18% mid-grey is not at the 0.18 mark on the x-axis, so, again, not intuitive
- as an exposure range, say ‘1 EV above and 3 EV below mid-grey should be linear’; that would probably have the most photographic relevance, but I’d still need to make sure input values are sanitised (e.g. if white relative exposure is 5 EV, and someone says they want to keep 6 EV above mid-grey linear, that has to be limited to 5 EV - easy to do, though). Once one changes the relative exposures, it may be best to re-adjust the exposure range of the linear part, too
- as a percentage of the distance between the black/white point and mid-grey; ‘keep 80% linear, then roll off’. There the answer to the ‘how many EV does that mean?’ question depends on the black/white relative exposure, but it may still be the most convenient one to use. I think that is worth an experiment.
- Oh, and when I say mid-grey, maybe it should really be an adjustable pivot, as raised by @Tamas_Papp. I’m not sure about this, there’s such a control in color balance rgb, and it’d mean yet another slider. But I’m just one guy here, if it’s a slider I never touch, but others find useful, why not?
Mid-term: the matrix stuff. Darktable already has it, but I want to understand it, and the best way for that is to get my hands dirty.
So that’s what I currently have in mind. Thanks everyone, for contributing your time and thoughts.
10 Likes