Simpsons - TFM vs QTGMC

[protectron]

Done with Futurama and working through The Simpsons now.

In the deinterlacing process, TFM does well at removing ghosting artifacts, but leaves jagged deinterlaced frames every 5 frames or so.
Conversely, QTGMC has clean deinterlacing, but leaves many ghosting artifacts. I found it works better with TDecimate Mode 7, but not perfect.

Is there some magical third option that can eliminate both problems?

TFM bad deinterlace


QTGMC good deinterlace


TFM good deinterlace


QTGMC bad deinterlace

[Selur]

You could:
a. use TIVTC with QTGMC
b. use QTGMC (Filter) afterwards for the clean-up
c. try vinverse(2)
d. some aa filter

[protectron]

a. Activating TFM at all whether through TIVTC or from Misc -> TFM produces the bad deinterlacing. I've tried all of the deinterlacers, modes, and PP in TIVTC and the output is the same. It's a TFM problem. I confirmed this by running QTGMC as the deinterlacer and toggling TFM (filter).
b. Do you mean the filter under DeNoise? I'm already running DGDenoise. It reduces the ghosting in some cases but makes it worse in other cases.
c. This does not appear to have an impact on either TIVTC or QTGMC.
d. I'm running AA in the first example image. It reduces the effect but still makes a noticeable flicker when animated.

[Selur]

b. Do you mean under DeNoise? I'm already running DGDenoise. It reduces the ghosting in some cases but makes it worse in other cases.

Yes, QTGMC under DeNoise allows using QTGMCs with InputType !=0:

InputType:
Default = 0 for interlaced input.
Settings 1, 2 & 3 accept progressive input for deshimmer or repair.
Frame rate of progressive source is not doubled.
Mode 1 is for general progressive material.
Modes 2 & 3 are designed for badly deinterlaced material.

Alternative one could use BasicVSR++ (maybe with masking if the general effect of BasicVSR++ is too strong), or a machine learning trained model (https://openmodeldb.info/).

Cu Selur

[protectron]

Is there an AI/ML style deinterlacer? All of the filters available today seem to be for upscaling and fixing the effects of deinterlacing.
I would prefer to fix the problem at the source.

---

Playing around with resizers: RealCUGAN is NUTS. My goodness. I might use that instead of DGDenoise if I can stomach 4 fps.
Still doesn't solve my original issue, though.

[Selur]

Hybrid does support:

• DeepDeinterlacer: https://github.com/pifroggi/vs_deepdeinterlace
  
• MFDIN: https://github.com/HolyWu/vs-mfdin
  

I don't know of any other machine learning deinterlacer that is available for Vapoursynth.

Cu Selur

[protectron]

If I wanted to try deepdeinterlacer is that in the Hybrid GUI or do I need to save the py filter and add custom code? If so, how do I do that?
I'm guessing you solved your CUDA issue here?

[Selur]

DDD and MFDIN are available in the deinterlacer selection when the torch add-on is installed. (I doubt they will help with your issue,...)

I'm guessing you solved your CUDA issue here?

not really

Cu Selur

[protectron]

Lmao. Found the addon in your Google Drive. 
I'll give DDD a try for giggles.

---

Do I also need vs-mlrt, onnx_models, and Avisynth32bit if I only want Torch?

[Selur]

No, those are different addons, also best use the latest dev and the corresponding torch addon