Selur's Little Message Board

No, afaik there is no Vapoursynth Version of that Project.

For completeness. I did mention it in the post, but for the last three, I also used CodeFormer and there is also a vs-mlrt addon.

Cu Selur

Ok download the file, use a lot of filter but i'm not good so never like the result   Dont know where I wrong...got better and fast result with topaz...
With removedirtmc can do someting good...but get a bit blurred...
Maybe need to go to build a trap for mouse..

With machine learning filters, there isn't much to configure, and it often comes down to on what content the models were trained.
So if Topaz works faster and better for you, use that.
You might be able to get better results with Hybrid with enough knowledge and tweaking, but if tool xy does the job conveniently, there is nothing forcing you to stick with Hybrid.
=> I would say: use whatever works best for you

Quote:With removedirtmc can do someting good...but get a bit blurred..

At least for the sample you shared, RemoveDirt seems like the wrong filter.

Cu Selur

Actually found this setting

mldegrain or mcdegrainsharp
knlmeanscl - yuv - gpu - strenght 2,20
finedehalo
finesharp or awarpsharp
resize realsr/srmd - gauss

Program is good, the problem is find a setting and maybe with others can be better, so its like going in the dark with a candle 
Seems scunet removes detail for example from the skin...
I keep trying...

Quote:Seems scunet removes detail, for example from the skin...

Especially skin and fine details are hard to differentiate for any algorithm (I know of) compared to compression artifacts, noise, etc.

It normally comes down to of how much time and effort you want to invest into it.
Usually you end up with some compromise since no algorithm really mirrors your quality perception and really knows a difference what are details and what are unwanted noise. (<- this is even hard for humans )

To add to the complexity, you can also:

• add a filter multiple times with different settings in a different filter order,
  
• apply a filter masked (only on portions of an image; Hybrid does come with some basic mask support),
  
• apply a filter only of portions of the clip,
  

and then settings that might work fine or one might work bad for another source.
Aside from Hybrid, you can also write your own scripts and filters and train custom models through machine learning.

=> Happy hunting/learning

Cu Selur

I discover deblock work better on luma channel in mid value (40 or more), amd maybe also degrain...denoiser dont know...
Dering do nothing for me...
The test go on...

Quote:Dering do nothing for me..

Does here, pretty obvious, but depending on the filtering before or after you might not see the effect.
You can set "FilterView mode" to "difference".

with interleave: https://imgsli.com/MzA4NTE5

Quote: I discover deblock work better on luma channel in mid value (40 or more), amd maybe also degrain...denoiser dont know...

It's up to your preferences.

Cu Selur

Find the option  "FilterView mode", very useful   So I can know what filter can be effective...
Make a fast test and make a bad clip pretty good  
But where are the tab for the filter order?

"Filtering->Vapoursynth->Misc->Filter Order/Queue" read also [INFO] About Vapoursynth Filter Order/Queue in case you want to use "Filter Queue" instead of "Filter Order".

Cu Selur