Capture sharpen, soft edges test

Capture sharpen is really working well but I’ve noticed a dither effect on smooth transitions. It’s most visible when you have a light and a dark area with a “long” gradient between. This happens with out of focus transitions and soft shadows. I’ve previously noticed it in the shadows on faces but these test images try out the oof scenario.

Raw files pp3’s and the crops below can be found at:

dual demosaic amaze-vng is used below but I found only a tiny difference with other demosaicers.

Look carefully at the transition between out of focus petals and the white background. Lower right corner and above the central diagonal stem are good areas to view.

Capture sharpening automatic Contrast threshold

Capture sharpening off

Roughly manually tuned Contrast threshold to get rid of dither.

My guess is that it’s just sharpening the noise. Was surprised to see it act on such a low detail area.

Can you show a screenshot of the sharpening mask from your example please? I know, because you provided the pp3, I could do that as well, but let’s exclude errors made by assumptions.

Sharpening mask

This is with version 5.7-113-g443b02e43 but observed it before.
I keep hoping there’s a baby in there :wink: looking a bit ultra soundish.

i just realized I did the tests with a pretty ancient build. Got my builds mixed up… same result with build from today though.

I inspected your raw file. The mask is based on microcontrast. There is some microcontrast in the region you showed (screenshot is without sharpening)

Maybe this can be improved (not sure about that, though)

I was surprised to see those soft gradients get picked up as detail. The mask visualises it well the gradients are thick solid areas of sharpening. The in focus areas have less detail it seems.

Can’t say I can really see the micro contrast in that screenshot. Unless the contrast is very micro. Do you think most gradients will be detail to the algo if the contrast and length of transition is such that adjacent pixels differ enough. I mean even a synthetic Gaussian type transition?

I think we’re talking about different areas. That’s the actual stem right? My concern was more with the petals in the background. see hightlighted areas. They are even more feathered than the stems.

The flat areas that aren’t picked up by the mask appear to have more microcontrast (noise) than the oof gradients. Super zoomed in examples below.

no cs

cs (auto)

mask (edit: oops ignore the markup below, means nothing right now)

Comparing two areas I can’t see more micro contrast in the (red) area picked up by the mask that in the blue area omitted by the mask. What am I missing?

It seems to me that somehow even large features such as transitions get picked up not because of internal detail but large scale appearance.

This isn’t just with capture sharpening. The normal sharpening shows the same behaviour.
Interestingly the areas highlighted by @nosle are kept in the mask even with noise reduction - Luminance set to maximum. Noise should be smoothed out completely then.

Does the noise reduction come before capture sharpening in the pipe? Normal should be after I guess but when is the mask calculated? Synthetic raw would be nice for these kinds of tests.


So quick test with normal sharpening mask.
Detail destroyed with these settings which atleast affect the appearance of the mask. Sorry about the language (at work…)
result of blur

Mask at 20
(issues uploading will try later “permission denied”)
Not sure what this means i anything, the transitions get picked up however.

Yes, currently the processing is not stopped in mask view, means tools after capture sharpening will destroy the appearance of the mask as well, but not the mask itself. That’s a thing I should fix… processing in preview mode should skip the steps after mask generation when in mask view mode.

The above was meant to test the masking of normal sharpening but i can’t upload the mask image regardless of what I do. drag’n drop, browse for file, new filename, re-saved image. “Permission Denied”

Copy paste did it! Here it comes.

I’ll try to find/produce a very large oof transition and see what happens. Presumably there’s a point (the gradient of a sky, vignetting, synthetic full size gradient?) that doesn’t get picked up as detail? If that’s the case there should be a knob one can tweak right?

I thought the whole point of capture sharpening (as opposed to sharpening for effect/detail) is that capture sharpening should be very subtle/limited as all it does is compensate for the inherent softness of the digital capture process. It comes first in the processing chain and any artifacts it introduces are then magnified by subsequent operations.

Sharpening to bring out detail should be very much later (almost the last thing) in the processing chain and does not need to be subtle/limited.