heya,
oh, i thought your branch was more about better integration of the
basecurve. to solve the color space/linearity issue i was thinking more
along the lines of doing some linear rgb curve after the Lab phase.
my branch does some fitting to raw + jpg pairs including all that i can (or
have to) to produce the look of the given jpg. i'm using levenberg
marquardt fitting on three instances of color zones (select by L,C,h),
tonecurve + ab curves, and a new module, color lut with the weirdest
unusable gui ever. maybe it'll improve.
lev/mar is very good at fitting stuff to curves where the derivatives make
sense, so you get very decent looks out of tonecurve + color zones (you
should try that fitting process on your samples if you have time).
my current problem is that canon seems to have some special sauce magic for
skin and even special for lips. the above setup without the new module
perfectly matches my sample, except for the lips. these you need to change
based on their input color + saturation, not only one of them at the time
the way color zones does it. that means our pipeline is not as powerful as
a general look up table, which is where the new module comes into play. it
allows you to remap points in the 3d LCh cube arbitrarily, making it
possible to just alter the lips, and skin tones separately from that.
unfortunately lev/mar has a very hard time fitting that module, all the
derivatives go haywire and the space is full of local minima. so i might
need to switch to a different fitting algorithm or do whatever i can with
the other modules and just use clut to fix up the lips, or any other stuff
that might come up that can't be solved with color zones.
also possible that the clut module is for some other reason still broken
and impossible to get good results with.
would be interested to hear about your images applied to my fitter.. you'll
need to install levmar-2.6/ in the film_emulation/ directory, i can assist
if you'd like to give it a try?
cheers,
jo
On Fri, Feb 7, 2014 at 11:06 AM, <ed.go...@free.fr> wrote:
> > this whole thing needs a cleanup, i'm slightly scared by the rgb
> > modules coming after the output profile, which will behave
> > differently depending on the profile (the only one i might
> > occasionally use is vignetting i think). i guess we should move
> > output/display profile to the very end and make the current colorout
> > a Lab->rgb + s-curve module.
>
> Hi Hanatos,
>
> It seems we are duplicating some efforts here :-D
>
> Your work on darktable/film-emulation and my
> darktable/tool_basecurve_improvements are heading in the same direction.
>
> As you seem to be more productive than i can be lately, i will put on
> standby my branch, but still share the results i got so far.
>
> Disabling basecurve and adding a proper tonecurve (which is well placed
> just before the color output module) i tried to simulate a late curve
> module.
>
> This works very well for the tonal curve (L curve). However the automatic
> a*b* scaling seems to hurt colors somehow, not sure what's at stake here as
> everything is intermixed by the La*b* formula.
>
> So I'm able to get nice results in many cases, but for some, colors are
> less punchy. I'd describe the colors as a bit dark and washed out.
>
> My next move would have been to not only profile the L curve, but also the
> a* and b*.
>
> I see you already started using a different color representation, LCh
> instead of La*b*, is that avoiding the color problems I have ?
>
> Cheers.
>
> PS: Do not hesitate to share results with me, i got quite interested in
> the matter lately.
>
> --
> Edouard Gomez
>
------------------------------------------------------------------------------
Managing the Performance of Cloud-Based Applications
Take advantage of what the Cloud has to offer - Avoid Common Pitfalls.
Read the Whitepaper.
http://pubads.g.doubleclick.net/gampad/clk?id=121051231&iu=/4140/ostg.clktrk
_______________________________________________
darktable-devel mailing list
darktable-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/darktable-devel
