Exactly, the second shot was more of a reference of what the picture should look like after corrections, rather than a subject of needed dehazing.
Here are some. I hope they are useful. Unfortunately, I did not save many that were washed out, and most of what I saved I do not want to share online.
DSCF2933.RAF (20.6 MB)
DSCF5652.RAF (23.6 MB)
DSCF8922.RAF (21.4 MB)
Wow! Are those all ‘1-click fixes’?
I do have some from my Helios 44-2 lens that were taken when the sun was low in the sky and they have a heavy orange hazing. If you provide me a link to a compiled Windows version of this I can run some tests and post screenshots with faces cropped.
Yes.
It’s still wip. I will create a branch on github when it’s more robust.
Would such a file be useful? I can try to replicate the effect if conditions are favorable.
Sure. Every example is welcome 
I tried to shoot worst case scenario, and came up with two images that have somewhat unique qualities.
Here is one that the bottom right has heavier flaring than the rest of the image.
DSCF2983.RAF (23.8 MB)
Here is one that got some rainbow flares. I’m curious to what the algorithm does to it.
DSCF2988.RAF (22.8 MB)
The first one did a nice job. I like how the rainbow flare was enhanced. But this one - not so much. The shed came out way too black on the upper left. It is really much more red in real life. Then again, I am not sure any better is possible. The image was pretty shabby to start with.
Well, the algorithm just subtracts values from each pixel in the image, which also causes a darkening of the image.
How about this?
It still has more black in it and is higher contrast than the ground truth, but it is definitely improved.
I finally got shareable orange cast images.
Here is one with recognizable subjects DSCF3463.RAF (22.7 MB) (The car color is ‘sterling gray metallic’)
And here is one with hazing cranked all the up to 11 DSCF3461.RAF (23.4 MB) (shot through a window screen which explains the grid artifact)
Heckflosse’s new tool will probably best be used in conjunction with some either Shadows / Highlights, or dynamic range compression, or good old reducing contrast in the exposure tools, to lessen the effect of too deep of shadows, resulting in a lower contrast, yet less washed out, image.
Definite improvement. It is also interesting to note that the algorithm did not change the color cast in meaningful way. It still kept the warm tones. This is exactly why I wanted to get an image like this for test. (This can be a good thing or a bad thing of course… depending on the image, the hue of the haze, etc).
This is different than the existing haze removal tool that in standard mode shifts some colors and has a tendency to make them look washed out in luminance mode. 
Any estimates on when this will be available for public testing? There is a fairly large market out there for vintage and cheap fast lenses.
If I understand correctly, RAWs with haze are somehow acceptable?
If so, I’m giving you the following two files: Hazy RAWs.zip (86.0 MB)
The permission is to use them only for open source development/testing. You probably have one more RAW with similar problem from me from GitHub issue. I don’t expect miracles from the RAW editor, so I avoid to take such photographs.
There is a noticeable color shift in the dog examples.
Thanks for the files
Here a comparison:
Top left: RT Standard Film Curve Iso Low profile.
Bottom left: As top left but + current Dehaze at default settings
Top right: As top left but + the new method
Bottom right: As top left but + the new method + current Dehaze at default settings
2 Likes
Thanks, the top right results look promising. I usually use curves to mitigate the haze. I think that my approach is a bit flawed, because the tone curves work in non-linear space.
I don’t use Dehaze and other tools with spatially varying effect, because I prefer to preserve the realism in my photos. The attempts to outsmart the nature or to deceive the viewer with spatially varying effects usually produce images with chewed up look. As I understand your method is spatially invariant (global).
1 Like
It works on raw data (linear) before demosaic und just subtracts for each channel of RGB the mininmal value of the channel. Totally simple and cheap.