New version with updates to the math and cosmetic changes…
Thank you!
@priort could you do this too? It’s a version with additional temporary code for finding good defaults in a “developer” mode.
Todd, this is a build from the clone of hanno’s hl_recovery_v5 branch?
Ya I added it as a remote branch…to the main darktable one but I think its pretty up to date on in any case…
PS with all these builds I have inherited the color calibration AI edges black screen.
I should have been paying closer attention as to when it happened…
Would have been nice to catch that one
Thanks @priort for your effort, it certainly helps Iain as he can’t build dt on windows himself! Just a reminder, this is highly experimental stuff, for sure only a small part of that will find it’s way into an official dt pr. Some parameters in the module are changed / treated in a way that would be concidered “no and never”.
Thanks for the explanation. I did try cycling through the various modes in the drop down. I was not able to discern much from that exercise… perhaps I just need the right image or the differences are indeed subtle…
@hannoschwalm I dont know how to compare two branches that are not in my repo. Are the hl_recovery_V5 and gl_recovery_v4 different solutions?
Use v5 this is the most current version of it in the evolutionary process
The 5er branch won’t work after the pr 4 has been merged.
I just tested the merged PR. This is very fast and very good at recovering the highlights. The slides changed since the last time I did a build. In one image I had I wanted to feather or bloom the segment selections because it was not picking up the edges correctly. It is a very tricky image, so it might be an outlier case for recovery.
So still working on v5 ie its the working version and only 4 considered safe so merged this one??
I concider the v4 code as stable, and wanted a merge into master early for more feedback on results people get with it
The v5 code only adds the “guess values in areas with all color channels clipped” code. I have yet not decided if I will do a pr for this part, results must be pleasing and performance good. Otherwise no chance for me.
In dt github there is a pull request (currently marked as WIP) implementing the recovery code.
- As we calculate distance for every segment we can avoid processing very small/narrow segments for performance.
- the strength of the effect is now controlled via a slider
- we can select gradient attenuation for small and large segments and we have an new “generic” mode that calculates date based on maximum distance in that segment.
- noise can be added to all-clipped segments
- Performance is pretty good - here even on 45Mpix images with lots of segments the recovery mode adds less than 100msec in almost all cases.
@priort I would appreciate if you could again compile for windows to allow Iain proper testing.
Please report back your experience, current code seems to be good after a lot of testing but a) don’t know what you think of it and b) we should concider if it’s ready for open-world.
Currently in Gros Morne on vacation…can do but not for a week or so
Looked that up, have fun and happy hiking?