best way to sharpen this?

[taigi]

What could be the best way to sharpen this?
Of course face is the target.
Thanks.

[pm-s]

What could be the best way to sharpen this?
Of course face is the target.
Thanks.

Sharpen(0.2)

[Selur]

# sharpening using DetailSharpen
clip = G41Fun.DetailSharpen(clip=clip, sstr=5.000, power=100.000)

also looks fine:

I also tried how it would look like when using BSRGAN (had to div the width&height by 4 to be able to apply it without running out of memory).


Cu Selur

[taigi]

Looks great.
Any other solution?
(BTW, I sharpened one still with Gimps unsharp mask,
result was OK, need to say).
Because DetailSharpen creates those small bubbles
BSRGAN reveals what Topaz is made of.

[pm-s]

BSRGAN is so sharp

[taigi]

BSRGAN is so sharp

Plasticy too.

[Selur]

Plasticy too.

To be fair some of the pastic effect in BSGRAN probably comes from me having to downscale by a factor of 4 to run it.

Any other solution?

LimitedSharpenFaster and SeeSaw should also work fine, you probably just have to use high strength values since due to the image resolution.

[pm-s]

Ok how do I install BSRGAN

[Selur]

Ok how do I install BSRGAN

Depends on how you want to use it.
https://forum.selur.net/thread-1858.html describes how I use it through VSGAN.
But you could also use it directly, see: https://github.com/cszn/BSRGAN

[pm-s]

Thank you Selur

[jagabo]

A simple fft3dfilter and unsharpmask creates halos but doesn't look too bad.

function UnSharpMask(clip v, float radius, float strength)
{
  blurry = v.BinomialBlur(VarY=radius, VarC=radius, Y=3, U=3, V=3) # or GaussianBlur
  edges = Subtract(v, blurry).ColorYUV(off_y=2).ColorYUV(cont_y=(int(strength*radius*256.0)-256.0))
  Overlay(v, edges.ColorYUV(off_y=-128), mode="Add")
  Overlay(last, edges.Invert().ColorYUV(off_y=-128), mode="Subtract")
  ColorYUV(off_u=-1, off_v=-1) #overlay is causing U and V to increase by 1
}

LWLibavVideoSource("howToSharpenThis.mp4", cache=false, prefer_hw=2) 
ConvertBits(8)
ConvertToYV12()
fft3dFilter(sigma=4.0, plane=4, bt=3)
UnsharpMask(4.0, 0.7)

[taigi]

thanks, @jagabo

[Selur]

"mcdegrainsharp(clip=clip, bblur=0.22, csharp=1.00, opencl=True)"
also works:


Cu Selur

[taigi]

What about all other sharpeners listed here?:
http://avisynth.nl/index.php/External_filters#Sharpeners

[Selur]

You are welcome to try them, most of them will give some improvement assuming their strength is just set high enough.

[poisondeathray]

What could be the best way to sharpen this?
Of course face is the target.

There are many "AI" face upscalers - but beware many of them add fake details or change characteristics.

I would describe the issue as softness, or general blurriness. It could even be described "out of focus". Most available machine learning deblur algorithms attempt to address motion blur, but there are a few at attempt to address out of focus. I was testing some deblurring algorithms; and your video fits the bill for the "out of focus" category

apng should animate in most browsers (I might be concerned about the teeth on the last frame, not sure how accurate that part is...it might be a misinterpolation)



Downscale by half (spline16), deblurred with restormer and deblurgan-v2, smoothed over with QTGMCp and oversharpened a bit in avisynth (cas and adaptive_sharpen), nnedi3_rpow2 upscaling

https://github.com/swz30/Restormer
https://github.com/VITA-Group/DeblurGANv2
(I just used the python code examples in the instructions ; these ones have not been ported to vapoursynth yet. Some projects can be difficult to use - but some of them have google colab notebook or some webpage demos that make it easier to use)

In the attached slowed down video you can see some issues with the right sideburn/temple "wobbling" or "pulsing" - normally something like that would be attributed to a mvtools2 issue such as QTGMC, or mdegrain derivatives. But in this case it was the restormer deblurring algorithm, which is a single image deblurring, not multimage/video algorithm. Those underlying artifacts predisposed QTGMC to exacerbate those artifacts more. ( maybe you can see it slightly in the source too, you don't notice it as much when blurred)

[lollo]

There are many "AI" face upscalers - but beware many of them add fake details or change characteristics.

Impressive result anyhow. Thanks pdr!

[taigi]

poisondeathray, wow, what a great, detailed analysis!
Your reply offers me only one solution -
to load up all these plugins and revive the work I had put off..
Your reply after more than a year is the best example that narcissism has not yet taken over the world;
there is still hope for the world to survive.
More than - thank you!

[Selur]

Downscale by half (spline16), deblurred with restormer and deblurgan-v2, smoothed over with QTGMCp and oversharpened a bit in avisynth (cas and adaptive_sharpen), nnedi3_rpow2 upscaling

That might be usable for one frame, but how long does that take? (how many minutes per frame are we talking here?)

[poisondeathray]

On that cropped area, it's pretty fast. On a UHD frame most people would be very slow because of memory. Restormer has tiling options in the original demo, deblurgan-v2 does not. I didn't measure it, but those deblurring algorithms are not as slow as some of the upscaling algos. You can batch them, but you waste time saving images and using them as input into the next (ie. it would be nicer if they were native vapoursynth ports, or if there was some GUI like chainner that could handle piping different architectures) . There is probably a way to do it directly in the vpy script if you're handy enough with python

[Selur]

Ah, okay, you only looked ad the cropped area and didn't process the whole image.

[taigi]

@ poisondeathray, I can't use restormer and deblurgan-v2 because of Python.
What could you suggest as an alternative?
Thank you.

[poisondeathray]

What is the problem specifically? One of the requirements or python in general? Did you want to try debugging it ? Or too much hassle ?

There are alternative /traditional methods of deblurring , such as deconvolution methods . For example, focus magic, smartdeblur. Smartdeblur actually has a free older version. They can do a great job on some types of blur, but one issue tends to be oversharpening, halos. GUI based and you can make adjustments and/or post filter with other such as avisynth, dehaloing. The pro version of smartdeblur has CLI options so you can batch it

[taigi]

@poisondeathray, too much hassle could be an answer.
Cos I need to start learning all the Py
and even have no idea how to start
(AvsP mod probably).
That could take ages,
but I'm sure results could be simply the best.
So trying to escape new studies lazily.

[Selur]

For the fun of it, I downscaled by 4 and then upscaled using vs-femasr:
https://imgsli.com/MTQzNTQ4

Cu Selur

[taigi]

I decided to extract the video to TIFF
and pass it through a commercial program
because I can't handle PY.
The process is terribly slow,
the entire video processing will probably take half a year
but me happy.