VHS conversion

[lostsnowboardvideos]

Selur - wondering if you can recommend some settings to enhance the quality of VHS encodes. I have attached a lossless lagarith AVI sample input and the x264 output and all the settings used. I don't typically have this issue with MKV files from DVDs but the VHS output samples look oversharpened. The only difference is I used QTGMC as a denoiser after deinterlacing and I don't usually do that for DVD files.

Debug and scripts are attached.

input/output files - https://drive.google.com/drive/folders/1...sp=sharing

[Selur]

I don't typically have this issue with MKV files from DVDs but the VHS output samples look oversharpen

Not surprised that vhs sourced content, especially when not ideally captured, looks way different than DVD content.

Looking at the sample:
a. content isn't interlaced, which vhs captures should be
b. content was badly deinterlaced (which causes the aliasing)

First thing I would do is apply DeNoise->QTGMC with 'Input Type' 2 or 3. This should help with the aliasing.
Then I would look into the chroma issues.
Will look at it some more after work.

Cu Selur

Ps: quick example of: QTGMC-mode3 + chroma only BasicVSR++ (no denoising, balance borders and similar)

[lostsnowboardvideos]

Looking at the sample:
a. content isn't interlaced, which vhs captures should be
b. content was badly deinterlaced (which causes the aliasing)

Huh. Okay. I captured the content on a gv-usb2 card from the VCR to VirtualDub64. Users in other forums insist it's interlaced even though it's scan type shows progressive. Not sure what the discrepancy is but I have deinterlace turned off on VDub. I'll have to keep messing with the settings.

Looking forward to hearing back.

[Selur]

Source isn't interlaced, otherwise bobbing would not create a duplicate frame every other frame.
You really might want to check the capturing settings. (but all those chroma artifacts&co and the deinterlacing seems like something went wrong)

About filtering the given source:
Stuff that should be done on a per-scene level:

• Fixing chroma derivations at the borders using BalanceBorders or Bore
  
• Luma adjustment (Retinex with tv scale looks fine to me, but that's just me)
  

Avoiding machine learning stuff, I try something like:

# Chroma adjustment using ChromaShiftSP
clip = chromashift.ChromaShiftSP(clip=clip, X=2.00, Y=2.00)
# contrast sharpening using CAS
clip = core.cas.CAS(clip=clip, sharpness=0.700)
# Denoising using QTGMC
clip = qtgmc.QTGMC(Input=clip, Preset="Fast", InputType=3, TR2=1, TFF=False, SourceMatch=0, Lossless=0, opencl=True, Denoiser="KNLMeansCL", TR0=2, TR1=2, Rep0=1, Rep1=0, Rep2=4, DCT=5, ThSCD1=200, ThSCD2=110, Sbb=0, NoiseProcess=2, GrainRestore=0.0, NoiseRestore=1,  StabilizeNoise=True, NoiseTR=0, NoiseDeint="bob") # probably overkill with all thse options, especiall try adjusting  ThSCD1=200, ThSCD2=110
clip = core.std.Crop(clip=clip, left=12, right=4, top=4, bottom=4)# cropping to 704x472
# denoising using HQDN3D
clip = core.hqdn3d.Hqdn3d(clip=clip, lum_spac=0.00, chrom_spac=10.00) # try tweaking the chroma denoising
# removing grain using SMDegrain
clip = smdegrain.SMDegrain(input=clip, RefineMotion=True, interlaced=False, opencl=True, device=-1)
# adjusting color space from YUV422P8 to YUV444P10 for vsRetinex
clip = core.resize.Bicubic(clip=clip, format=vs.YUV444P10, range_s="limited")
# color adjustment using Retinex
clip = core.retinex.MSRCP(input=clip, sigma=[25,80,250], fulls=False, fulld=False)
# denoising using KNLMeansCL
clip = denoise.KNLMeansCL(clip=clip, d=0)
# applying dehalo using YAHR
clip = dehalo.YAHR(clip, depth=24)
# Using FastLineDarkenMOD for line darkening
clip = linedarken.FastLineDarkenMOD(c=clip)

if machine learning is fine, I would use BasicVSR++ for the chroma cleanup instead of HQDN3D

# Chroma adjustment using ChromaShiftSP
clip = chromashift.ChromaShiftSP(clip=clip, X=2.00, Y=2.00)
# contrast sharpening using CAS
clip = core.cas.CAS(clip=clip, sharpness=0.700)
# Denoising using QTGMC
clip = qtgmc.QTGMC(Input=clip, Preset="Fast", InputType=3, TR2=1, TFF=False, SourceMatch=0, Lossless=0, opencl=True, Denoiser="KNLMeansCL", TR0=2, TR1=2, Rep0=1, Rep1=0, Rep2=4, DCT=5, ThSCD1=200, ThSCD2=110, Sbb=0, NoiseProcess=2, GrainRestore=0.0, NoiseRestore=1,  StabilizeNoise=True, NoiseTR=0, NoiseDeint="bob")
clip = core.std.Crop(clip=clip, left=12, right=4, top=4, bottom=4)# cropping to 704x472
clipMerge = clip
# adjusting color space from YUV422P8 to RGBH for vsBasicVSRPPFilter
clipMerge = core.resize.Bicubic(clip=clipMerge, format=vs.RGBH, matrix_in_s="470bg", range_s="limited")
# Quality enhancement using BasicVSR++
from vsbasicvsrpp import basicvsrpp as BasicVSRPP
clipMerge = BasicVSRPP(clip=clipMerge, model=4)
clipMerge = core.resize.Bicubic(clip=clipMerge, format=vs.YUV444P16, matrix_s="470bg", range_s="limited")
clip = core.resize.Bicubic(clip=clip, format=vs.YUV444P16, range_s="limited")
clip = core.std.Merge(clip,clipMerge,[0, 1, 1]) # merge filtered and unfiltered by plane
# removing grain using SMDegrain
clip = smdegrain.SMDegrain(input=clip, RefineMotion=True, interlaced=False, opencl=True, device=-1)
# adjusting color space from YUV444P16 to YUV444P10 for vsRetinex
clip = core.resize.Bicubic(clip=clip, format=vs.YUV444P10, range_s="limited", dither_type="error_diffusion")
# color adjustment using Retinex
clip = core.retinex.MSRCP(input=clip, sigma=[25,80,250], fulls=False, fulld=False)
# denoising using KNLMeansCL
clip = denoise.KNLMeansCL(clip=clip, d=0)
# applying dehalo using YAHR
clip = dehalo.YAHR(clip, depth=24)
# Using FastLineDarkenMOD for line darkening
clip = linedarken.FastLineDarkenMOD(c=clip)

Cu Selur

[lostsnowboardvideos]

Awesome thank you! I will try this out later tonight after work.

[lostsnowboardvideos]

Ran both scripts. Files attached. The one issue I see is the flickering of the frames. I ran a separate test (not included in the files here) with Retinex off and it seemed to fix the problem. Seeing some rainbowing effects too, would take recommendations on how to reduce that if possible.

Overall I think I like the HQDN3D better than the machine learning. Even with the chip and GPU upgrades I did, the machine learning stuff is still really slow.

https://drive.google.com/drive/folders/1...drive_link

[Selur]

Seeing some rainbowing effects too, would take recommendations on how to reduce that if possible.

Stronger chroma denoising is probably the only thing you can do. (or use BasicVSR++ on the chroma)
see: https://imgsli.com/MzcxNTM5

The one issue I see is the flickering of the frames. I ran a separate test (not included in the files here) with Retinex off and it seemed to fix the problem.

Then you got multiple options:
a. not use Retinex, but some other color/luma adjustment, if wanted/needed
b. try adjusting the Retinex values, maybe even per scene

Overall I think I like the HQDN3D better than the machine learning. Even with the chip and GPU upgrades I did, the machine learning stuff is still really slow.

Yes, using machine learning stuff usually is a lot slower than conventional filtering.

Cu Selur