SmoothAdjust [v3.20 - Update 2015/11/15]

[Dogway]

Originally Posted by LaTo

You only need to multiply value by 256.
If I remap the value internally, it avoids fine adjustments.

I never worked with such numbers in 32bit applications, this next line is far from intuitive.

"0-4096;8704-4096;10496-8192;15616-17664;31744-31744;47872-46080;53760-53760;58880-64512;65280-60160",\

For fine adjustments we have always used decimals...

Originally Posted by LaTo

In SmoothLevels you only need to set output_low/output_high at 16/235 (or 4096/60160 in 16bits).

I see, I never use smoothlevels or smoothtweak but I thought about helping. Output_low/output_high at 16/235 is the same as clamping, and I recently read that many TVrange sources have below 16 or above 235 values which should be better leave alone.

Originally Posted by LaTo

In SmoothTweak TVrange/PCrange doesn't change the processing (you can set "limiter=true" in TVrange).

"limiter=true" readme says: Clamp the output to 16-235...
But I dont want to output/clamp TV range values,but instead work in TV range
For example, see this analogy by using smoothcurve

saturation in PC range:
"0-0;64-"+string(64-64*sat)+";128-128;192-"+string(192+64*sat)+";255-255"
saturation in TV range:
"0-16;16-16;74-"+string(74-55*sat)+";127-127;185-"+string(185+55*sat)+";240-240;255-240"

...or I am totally wrong.

Originally Posted by LaTo

[LaTo]

Originally Posted by Dogway

I never worked with such numbers in 32bit applications, this next line is far from intuitive.

"0-4096;8704-4096;10496-8192;15616-17664;31744-31744;47872-46080;53760-53760;58880-64512;65280-60160",\

For fine adjustments we have always used decimals...

No, 16bits values is the only way to go for precision. (SmoothAdjust works in 32bits internally, 16bits is just the input clip)

Example: "Invert" the 16 bits clip

SmoothCurve16("0-65535;65535-0") #Easy
SmoothCurve16("0.0-255,99609375;255,99609375-0.0") #Your proposition with decimals (65535/256 = 255,99609375)
SmoothCurve16("0-255;255-0") #Internal remap: the conversion is wrong (255*256->65280 != 65535)

Originally Posted by Dogway

I see, I never use smoothlevels or smoothtweak but I thought about helping. Output_low/output_high at 16/235 is the same as clamping, and I recently read that many TVrange sources have below 16 or above 235 values which should be better leave alone.

It's not the same as clamping (with limiter=0).

Originally Posted by Dogway

"limiter=true" readme says: Clamp the output to 16-235...
But I dont want to output/clamp TV range values,but instead work in TV range
For example, see this analogy by using smoothcurve

saturation in PC range:
"0-0;64-"+string(64-64*sat)+";128-128;192-"+string(192+64*sat)+";255-255"
saturation in TV range:
"0-16;16-16;74-"+string(74-55*sat)+";127-127;185-"+string(185+55*sat)+";240-240;255-240"

...or I am totally wrong.

Your 2nd script clamp output to TVrange (should be "0-0;16-16;...;240-240;255-255")
Also I doesn't understand why you want to leave theses parts "as-is".
If you do that, you will have strange things later.

Example: "Desaturate" a clip

SmoothCurve(UVcurve="0-128;255-128") #In "PCrange", all colors are gone
SmoothCurve(UVcurve="0-0;16-16;17-128;239-128;240-240;255-255") #In "TVrange", extreme colors are still here

Originally Posted by Dogway

I will try to improve a bit the view

[ChaosKing]

beta4 AVX test
--------------------

#smoothlevels(gamma=2.0, useOPT=0) # 159 fps
#smoothlevels(gamma=2.0, useOPT=1) # 217 fps
#smoothlevels(gamma=2.0, useOPT=2) # 92 fps

#smoothlevels(gamma=2.0, useOPT=0, useMT=0) # 66 fps
#smoothlevels(gamma=2.0, useOPT=1, useMT=0) # 140 fps
#smoothlevels(gamma=2.0, useOPT=2, useMT=0) # 33 fps

The AVX part still does not seem to work properly :/

[Dogway]

Originally Posted by LaTo

No, 16bits values is the only way to go for precision.

I dont know how other applications deal internally with it, and I can't tell you how but Im pretty sure is possible, that's why I tried to fill a bit your empty todo list. Its unintuitive as hell, thats all.

Originally Posted by LaTo

If you do that, you will have strange things later.

In effect I was totally wrong yey!
Some tests a few months ago I remember I could desaturate with no problems, but it makes sense that out of TVrange bounds colors follow a correlation with the curve

Originally Posted by LaTo

Your 2nd script clamp output to TVrange (should be "0-0;16-16;...;240-240;255-255")

Then I guess this example wasn't correct at all
What I expect is a curve behaviour inside the TVRange in the same fashion as PCrange, but leaving the out of bounds values practically untouched. That's why my confusion on whether "0-0;16-16... or "0-16;16-16... because the first seems to be more locked on the extremes by the knobs (thus modifying the curve weight) than PCrange curve. While it being at first glance the ideal approach I can't tell behaviour is the same, Im not mathematician.

Originally Posted by LaTo

It's not the same as clamping (with limiter=0).

Then maybe is contradictory. Because "Output_low/output_high at 16/235" in levels is not "0-0;16-16;...;235-235;255-255". Maybe you wanted to say "16,1.0,235,16,235" but this flats values in the range 0-16 to a value of 16. There is no way in levels to do "proper" TVrange like you said in curves "0-0;16-16;...;235-235;255-255" (assuming this is the correct way which I discussed in the previous paragraph.)

edit: I think interp=100 doesn't work, using beta4

[mandarinka]

Hmm, I don't know if I am doing something wrong, here...

I wanted to replicate the effect of tweak(hue=+5) using smoothtweak. After some confusion, I realized that smoothtweak's hue1 and hue2 in fact seem to do the same thing as coloryuv's off_u and off_v, respectively.

(Acording to my very limited understanding, the hue parameter as used by tweak kind of rotates the color spectrum around the center/white/black/grey (or something like that, excuse me please), and strangely enough, a dvd I am trying to color-adjust for some strange reason seem to respond well to this kind of shifting...)

1) Is my observation correct? if it is so, mybe the parameters could use some additional explanation in readme...
2) Is there a way to do the adjustment tweak does for hue with smoothadjust, with its superior precission?

[LaTo]

Originally Posted by mandarinka

Hmm, I don't know if I am doing something wrong, here...

I wanted to replicate the effect of tweak(hue=+5) using smoothtweak. After some confusion, I realized that smoothtweak's hue1 and hue2 in fact seem to do the same thing as coloryuv's off_u and off_v, respectively.

(Acording to my very limited understanding, the hue parameter as used by tweak kind of rotates the color spectrum around the center/white/black/grey (or something like that, excuse me please), and strangely enough, a dvd I am trying to color-adjust for some strange reason seem to respond well to this kind of shifting...)

1) Is my observation correct? if it is so, mybe the parameters could use some additional explanation in readme...
2) Is there a way to do the adjustment tweak does for hue with smoothadjust, with its superior precission?

1. Correct
2. No sorry

Originally Posted by Dogway

edit: I think interp=100 doesn't work, using beta4

Yes, just a bad scaling... 50 is in fact 100, so higher than 50 produces strange result (fixed in the next beta)

Originally Posted by ChaosKing

beta4 AVX test
--------------------

#smoothlevels(gamma=2.0, useOPT=0) # 159 fps
#smoothlevels(gamma=2.0, useOPT=1) # 217 fps
#smoothlevels(gamma=2.0, useOPT=2) # 92 fps

#smoothlevels(gamma=2.0, useOPT=0, useMT=0) # 66 fps
#smoothlevels(gamma=2.0, useOPT=1, useMT=0) # 140 fps
#smoothlevels(gamma=2.0, useOPT=2, useMT=0) # 33 fps

The AVX part still does not seem to work properly :/

Very very strange, I don't understand. The results was with the MSVC or the ICL build?

Can you try the 2 dlls in this package > SmoothAdjust.AVXtest.zip < and report FPS here (with useMT=0)?

Thanks!

[Dogway]

@LaTo: I was doing some tests and investigations using vectorscope for accuracy. Indeed SmoothLevels can work with TVrange like this "SmoothLevels(16,1.0,235,16,235..." but by doing so you cant avoid out of 16 and 235 values to not be flatten. This procedure overrides limiter=0. I can post samples but dont want to make the post too bulky.

I also noticed that there is no way you can match smoothcurve and smoothlevels in a simple gamma tweak, I get different curves, and results. So I think you are using another valid approach Im not aware of or something is missing.

SmoothCurve(Ycurve="0-0;128-136.3;255-255",smooth=0,dither=0,interp=50,show=true)
!=
SmoothLevels(0,1.1,255,0,255,smooth=0,show=true,dither=0,limiter=0)

About the 16bit input numbers, could this be possible? Fit (0-65535)/256 range values into 0-255. I dont know how range mapping is done in programming so Im probably talking too much.
Also is it broken decimals input in beta4?

[LaTo]

Originally Posted by Dogway

@LaTo: I was doing some tests and investigations using vectorscope for accuracy. Indeed SmoothLevels can work with TVrange like this "SmoothLevels(16,1.0,235,16,235..." but by doing so you cant avoid out of 16 and 235 values to not be flatten. This procedure overrides limiter=0. I can post samples but dont want to make the post too bulky.

Sorry, clamping is a bug. Fixed in the next beta.

Originally Posted by Dogway

I also noticed that there is no way you can match smoothcurve and smoothlevels in a simple gamma tweak, I get different curves, and results. So I think you are using another valid approach Im not aware of or something is missing.

SmoothCurve(Ycurve="0-0;128-136.3;255-255",smooth=0,dither=0,interp=50,show=true)
!=
SmoothLevels(0,1.1,255,0,255,smooth=0,show=true,dither=0,limiter=0)

Yes, it's normal... You need to set more control points because the interpolation doesn't follow a power curve.

Originally Posted by Dogway

About the 16bit input numbers, could this be possible? Fit (0-65535)/256 range values into 0-255. I dont know how range mapping is done in programming so Im probably talking too much.
Also is it broken decimals input in beta4?

Not broken, input points are integer from the beginning. Decimals added in next beta.

[Dogway]

You are right, I get the same identical results as of levels with "mt_lut(yexpr="x 255 / 1 1.1 / ^ 255 * ",Y=3,U=2,V=2)" then what is the "correct" gamma for smoothcurves? Is there an interp value to conform the power curve?

[LaTo]

Originally Posted by Dogway

You are right, I get the same identical results as of levels with "mt_lut(yexpr="x 255 / 1 1.1 / ^ 255 * ",Y=3,U=2,V=2)" then what is the "correct" gamma for smoothcurves? Is there an interp value to conform the power curve?

No, the only solution is to add more control points to approximate the gamma curve.

[Ponder]

@Lato
I do get back a lot more contrast on Lmode1 when using very conservative setting on BrightSTR.
Varying DarkSTR is quite safe on the other hand.

Q1. If I use gamma=1, will smoothing be turn off. Q2.If set Smode=1, will it be on at all time?
At certain situation, I want it to be on at all time with gamma=1.

Q3. Same behavior between v1.74 and v2?

[LaTo]

v2.00 beta5:
tweaked AVX support
fixed clamping bug in SmoothLevels
fixed interp scaling in SmoothCurve
added decimals support in SmoothCurve

Todo-list:
- fix AVX performance?!?
- improve debug view
- add SmoothCustom
- fix all bugs

Download here: > SmoothAdjust v2.00 beta5 <

[LaTo]

Originally Posted by Ponder

@Lato
I do get back a lot more contrast on Lmode1 when using very conservative setting on BrightSTR.
Varying DarkSTR is quite safe on the other hand.

Q1. If I use gamma=1, will smoothing be turn off. Q2.If set Smode=1, will it be on at all time?
At certain situation, I want it to be on at all time with gamma=1.

Q3. Same behavior between v1.74 and v2?

Q1: No
Q2: Smode=0 -> always OFF // Smode=1/2 -> always ON
Q3: Yes but Smode is removed ( smooth>0 -> always ON // smooth=0 -> always OFF )




Here is a fast comparison of v1.74 & v2.xx with strength/smooth=100 (very strong):

Source image
Smoothing OFF
v1.74 (Smode=2)
v2.00beta5 (More effective, more details & sharpness, no artifact...)

SmoothCurve(Ycurve="0-16;16-16;64-128;128-224;235-235;255-235",Ucurve="0-0;64-192;128-128;192-64;255-255",Vcurve="0-255;64-64;128-128;192-192;255-0")

[leeperry]

Originally Posted by LaTo

Here is a fast comparison of v1.74 & v2.xx with strength/smooth=100 (very strong):

Source image
Smoothing OFF
v1.74 (Smode=2)
v2.00beta5 (More effective, more details & sharpness, no artifact...)

very impressive. for the new build, much appreciated

[Dogway]

LaTo, thanks a lot! This plugin feels profesional, I wonder what SmoothCustom is... : D
btw how do you get that fancy colors in the UV debug, the Cb Cr colors...

[Ponder]

@Lato, Thanks, it clears up what I suspected on Q1&2 since DarkSTR was set to some value ,even
when gamma was 1.

[ChaosKing]

beta5 AVX test
------------------

#smoothlevels(gamma=2.0) # 232.56 FPS
#smoothlevels(gamma=2.0, useOPT=0) # 173.86 FPS
#smoothlevels(gamma=2.0, useOPT=1) # 232.89 FPS
#smoothlevels(gamma=2.0, useOPT=2) # 240.10 FPS

#smoothlevels(gamma=2.0, useOPT=0, useMT=0) # 83.33 FPS
#smoothlevels(gamma=2.0, useOPT=1, useMT=0) # 165.28 FPS
#smoothlevels(gamma=2.0, useOPT=2, useMT=0) # 178.57 FPS

I noticed that smoothlevels(gamma=2.0) doesn't seem to use the AVX part. With useOPT=2 it's always faster!

[LaTo]

Originally Posted by ChaosKing

beta5 AVX test
------------------

I noticed that smoothlevels(gamma=2.0) doesn't seem to use the AVX part. With useOPT=2 it's always faster!

Thanks for testing! Can you try these 2 dlls (AVXtest.zip) and report the speed for each?

[ChaosKing]

AVXtest (test)
----------------

##avx 1
#smoothlevels(gamma=2.0) # 233.33 FPS
#smoothlevels(gamma=2.0, useOPT=2) # 232.56 FPS

#smoothlevels(gamma=2.0, useMT=0) # 129.03 FPS
#smoothlevels(gamma=2.0, useOPT=2, useMT=0) # 128.05


##avx2
#smoothlevels(gamma=2.0) # 241.95 FPS
#smoothlevels(gamma=2.0, useOPT=0) # 105.12 FPS
#smoothlevels(gamma=2.0, useOPT=1) # 227.27 FPS
#smoothlevels(gamma=2.0, useOPT=2) # 243.89 FPS

#smoothlevels(gamma=2.0, useMT=0) # 129.03 FPS
#smoothlevels(gamma=2.0, useOPT=2, useMT=0) # 127.39 FPS

avx1 seems to be slower with useMT=-1. Each test was run with 3-5 runs.

[LaTo]

Originally Posted by ChaosKing

beta5 AVX test
------------------

[...]

I noticed that smoothlevels(gamma=2.0) doesn't seem to use the AVX part. With useOPT=2 it's always faster!

@ChaosKing: Please redownload beta5, now it will autoselect AVX correctly