Heavy problem from a capture of a film

[Selur]

Baldy digitalized/scanned content that already got processed a lot. (1080p@25 is no way the original resolution)
Sample stats during a blended scene change,.. :/

Okay, what do you want to address?
Problems I see, that one can address

• not cropped => crop it manually, i.e.:
  
  clip = core.std.Crop(clip, left=20, right=30, top=4, bottom=20)
• blown out highlights => one could use for example, in a custom section:
  
  # light: # T=220, m=0.9, b=220*0.1=22 # moderate: T=800, m=0.8, b=160 # strong: # T=180, m=0.7, b=180*0.3=54 clip = core.std.Expr(clip, expr=["x 200 > x 0.8 * 40 + x ?","", ""])
  or Retinex (if the source is properly cropped) with adjusted values
  
• shaky => Stab or in this case probably better StabalizeIT and maybe add Temporalfix
  
• Tons of horizontal lines due to a bad executed transfer.
  One can try:
  • vs-hinet in 'derain'-mode with scratch enabled and a high Width divider (i.e. 8), but that will blur a lot (and is slow), also adjust the settings in Scratch,...
    
    so maybe, in combination with some masking a better result could be archived
    
  • DeScratch + SpotLess(or similar), here you probably want to use multiple Descratch calls, with different values.
    

So, my suggestion would be, assuming properly capturing/digitizing isn't an option try to get a better source or a source that is less processed.
If that isn't an option either, get accustomed to using DeScratch.
May be lower the resolution to 720x408 before calling DeScratch or HiNet. (adjust the settings accordingly)
wild try (only using 1 DeScratch call) script, clip
=> there are tons of ways to deal with such content, but it will require a lot of trying different settings and may be scene-by-scene filtering.
All possible, but a lot of work and your result won't get as good as when you had started out with a better digitalization or at least a source that hasn't be filtered before.

Cu Selur

Ps.: maybe interessting: https://forum.doom9.org/showthread.php?t=183832
PPs.: just noticed I forgot to adjust the highlight handling to 10bit, since I overlooked that the source is 10bit

scale = 1 << (clip.format.bits_per_sample - 8) clip = core.std.Expr(clip, expr=[f"x {200 * scale} > x 0.8 * {40 * scale} + x ?", "", "" ])

need to be used to scale properly, then the luma isn't mistreaded as above. DOH