Re: [Gimp-developer] Gimp from git Gegl Gaussian Blur and Unsharp Mask useability issues

[Elle Stone <ellestone ninedegreesbelow com>]

I spent some time experimenting with 2.8 vs 2.9 and also with various 
slider values for gaussian blur and USM.

Comparing 2.8 radius and 2.9 std.dev gaussian blurs:

I made a test image composed of varying width black and white stripes 
and checked varying blur values (radius in 2.8, std. dev in 2.9) to see 
what the 2.9 equivalent std.dev (sd) was for various 2.8 blur radius (r) 
values. I used "difference" blend mode to gage which 2.9 blur values 
correspond to which 2.8 blur values.

Here's the results:

2.8     2.9     ratio
1r      0.55sd  1.82
2r      0.8sd   2.50
5r      1.7sd   2.94
20r     6.2sd   3.23
40r     12sd    3.33
100r    31sd    3.23
200r    61sd    3.28

Extrapolating, perhaps a 500r 2.8 blur would take a 155sd 2.9 blur. The 
Gegl Gaussian blur popup says "multiply by ~2 to get radius", but it 
seems that for all but the smallest values, "multiply by ~3" is closer 
to the equivalent Gimp 2.8 gaussian blur. Has the definition of "radius" 
changed? Or something in the GB algorithm?

Do you think the different values and slider labels required to get the 
same amount of blur (1r vs 0.55sd, 5r vs 1.7sd, etc) might cause a great 
deal of consternation when Gimp 2.10 is released?

http://www.hackerfactor.com/blog/index.php?/archives/289-Blurring-The-Lines.html 
gives a nice comparison between Gimp (not from git), Photoshop, Corel 
Painter regarding the "unit" of blur, which apparently Photoshop uses 
the diameter and the other editors use the radius.



Very large gaussian blurs:

Quoting Wikipedia:

"Applying multiple, successive gaussian blurs to an image has the same 
effect as applying a single, larger gaussian blur, whose radius is the 
square root of the sum of the squares of the blur radii that were 
actually applied. For example, applying successive gaussian blurs with 
radii of 6 and 8 gives the same results as applying a single gaussian 
blur of radius 10."

So it's not too hard to hit whatever target gaussian blur one has in 
mind, either visually by repeated blurring or by calculating the right 
successive blur values to use.

As the gaussian blur radius/std.dev gets large compared to the actual 
size of the image, edge and corner values necessarily start to dominate 
the result of the blur. Does anyone have a use case for a 6000-radius or 
even 2000-radius blur?


Slider values:

After some experimentation, these values seem to work nicely. Both sets 
of ranges allow for reasonably precise setting of low values with the 
sliders. Both sets of slider settings require that the user figure out 
that there really is an extended scale. I raised that last ui parameter 
from 1.5 to 3.0, to make the extended scale not be too long.


Unsharp Mask:

gegl_chant_double_ui (std_dev, _("Std. Dev."),
                      0.2, 200.0, 0.55, 0.2, 10.0, 3.0,
                      /* was 0.0, 500.0, 1.0, 0.0, 200.0, 1.5, */
                      _("Standard deviation (spatial scale factor)"))

gegl_chant_double_ui (scale, _("Scale"),
                      0.0, 200.0, 4.0, 0.0, 10.0, 3.0,
                      /* was 0.0, 500.0, 1.0, 0.0, 200.0, 1.5, */
                      _("Scale, strength of effect"))

Comments:
* These ranges allow for very precise setting of low values, critical 
for photographic images.
* On the std_dev slider there seems to be a dead spot below 0.2 on the 
slider, where no matter how high Scale is, nothing happens. So I set the 
minimum on the upper slider to 0.2.
* I set the defaults to a mild (obvious but not over the top) 
sharpening, one goal being to let the user know that the two sliders 
really are independent of each other.
* Is there an interest in creating two presets, one a mild but obvious 
sharpening and the other a mild but obvious local contrast enhance, so 
that new users would have them available as drop-down choices? It seems 
to me that with the radical change in the slider values/meanings, a 
couple of provided defaults would help users figure out where to start.


Gaussian blur:

gegl_chant_double_ui (std_dev_x, _("Size X"),
                      0.0, 1500.0, 1.5, 0.0, 100.0, 3.0,
                      /* was 0.0, 10000.0, 4.0, 0.0, 1000.0, 1.5, */
                      _("Standard deviation for the horizontal axis "
                        "(multiply by ~2 to get radius)"))

gegl_chant_double_ui (std_dev_y, _("Size Y"),
                      0.0, 1500.0, 1.5, 0.0, 100.0, 3.0,
                      /* was 0.0, 10000.0, 4.0, 0.0, 1000.0, 1.5, */
                      _("Standard deviation for the vertical axi. "
                        "(multiply by ~2 to get radius)"))

Comments:
* The Gegl Gaussian blur slider range is the equivalent of 0 to roughly 
300px radius plus 0 to roughly 4500px radius for the extended scale.
* I changed the default setting to be a visual match to the default Gimp 
2.8 "5px blur" setting.
* There's a dead spot below around 0.4, below which blur value, nothing 
seems to happen.


I'll put a patch together later today (or probably tomorrow AM), if 
these values seem suitable.


A Gegl Gaussian blur bug?

There seems to be a bug in Gimp from git's Gegl Gaussian blur. It 
produces vertical streaks, noticeable at higher value blurs in larger 
images, very obvious at 155unit blur of a 2816x2112px image, less 
obvious when the image is scaled to 1408x1056px before blurring.

Elle