Horizontal random distortion Problem...

[dg3duy]

I need help to understand how to reduce the noise effect in the image.
I am quite confused as to how to correct the distortions of a video that I don't have the original, there is a noise of pixel movements, as if the image is constantly going from left to right. Is it possible to stabilize the pixels?
I am quite confused as to what the technical name is for what is happening, the original video was captured from a vhs apparently.

I am open to any suggestions, I tried Davinci resolve, and some AviSynth+ without success.

Thank you very much!
Example Video:
https://www.mediafire.com/file/cqi16lntywxbba3/Noise_Distort.mov/file

[johnmeyer]

I see no right/left movement of the entire image.

The video does contain some time base errors, but the quality of the video is so low that even if you fixed those errors, I don't think it would look much better.

[dg3duy]

Thank you very much for answering, technically what is the name that has that flaw in the image. I am referring to the vibration of the image.

I see no right/left movement of the entire image.

The video does contain some time base errors, but the quality of the video is so low that even if you fixed those errors, I don't think it would look much better.

[davexnet]

guess you could call it jitter, research time base correctors

[dg3duy]

Perfect! I will look into it, thank you very much.

guess you could call it jitter, research time base correctors

[themaster1]

http://www.avisynth.nl/users/vcmohan/DeJitter/DeJitter.htm

[dg3duy]

It is interesting the script, I made a test with the video that I attached in the publication and I don't see any change, maybe the video lacks information in the borders so that it works correctly, it is an assumption of mine.

http://www.avisynth.nl/users/vcmohan/DeJitter/DeJitter.htm

[jagabo]

Of course there was no change. That filter looks for the black borders at the left and right (back and front "porch") and resizes the picture between them to a standard length. That sometimes works with the original analog cap (when the active picture doesn't contain black objects at the edges of the frame) but your video has cropped away those black borders and replaced them with perfectly straight black borders. The filter has no idea what the original borders looked like.

[dg3duy]

Thank you very much for the comment, it is correct what you say, the video unfortunately has cropped edges so the script will not be able to be applied.

Of course there was no change. That filter looks for the black borders at the left and right (back and front "porch") and resizes the picture between them to a standard length. That sometimes works with the original analog cap (when the active picture doesn't contain black objects at the edges of the frame) but your video has cropped away those black borders and replaced them with perfectly straight black borders. The filter has no idea what the original borders looked like.

[Selur]

Since there are not many details to begin with, using something like BasicVSR++ might be a way to stay to stabilize the image a bit.

Cu Selur

[Alwyn]

@Selur, could you please post your videos in H264 or H265; VDub2 cannot open AV1s. Thanks.

[Selur]

attached a h264 version

[Alwyn]

@Selur thanks.

[dg3duy]

The final result is stunning! It's so much more than I was expecting!

What is the path I have to follow to reach the result you show! You left me speechless, thank you very much for your time.

Since there are not many details to begin with, using something like BasicVSR++ might be a way to stay to stabilize the image a bit.

Cu Selur

[Selur]

I used Hybrid with the torch-addon (this requires a newer NVIDIA card) to use BasicVSR++, loaded your source, enabled BasicVSR++ (Filtering->Vapoursynth->Other->BasicVSR++ with FP16 to speed things up), configured the encoder, set an output, created the jobs and let it run.

[Alwyn]

Here's Virtual Dub 2 with the Neat Video plugin applied (basic setting, no fine adjusting).

[dg3duy]

What a great workflow, I have a lot of research to do, thank you very much for everything!

I used Hybrid with the torch-addon (this requires a newer NVIDIA card) to use BasicVSR++, loaded your source, enabled BasicVSR++ (Filtering->Vapoursynth->Other->BasicVSR++ with FP16 to speed things up), configured the encoder, set an output, created the jobs and let it run.

[dg3duy]

Your copy also looks really good! What an interesting combination of programs to enhance the image.
What a great result you got, thank you very much!

Here's Virtual Dub 2 with the Neat Video plugin applied (basic setting, no fine adjusting).

[jagabo]

Not as smooth as the others but sharper. With AviSynth+ and x264 CLI:

LWLibavVideoSource("Noise Distort.mov", cache=false, prefer_hw=2) 
ConvertBits(8)
ConvertToYV12()
Crop(240,0,-240,-0)
source = last

# downscale closer to real resolution
Spline36Resize(width/4, 272)
Santiag() # antialias
MCTemporalDenoise(settings="very high") # noise reduction

# upscale back to source resolution, plus sharpening
Sharpen(0.5, 0.0)
nnedi3_rpow2(2, cshift="Spline36Resize")
aWarpSharp2(depth=5)
Sharpen(0.5)
nnedi3_rpow2(2, cshift="Spline36Resize", fwidth=source.width, fheight=source.height)
aWarpSharp2(depth=5)
Sharpen(0.5)
# use less sharpening if you want less processed-looking video

# multithreading
prefetch(8)

[dg3duy]

I like your way of clarifying the image, I am very surprised anyway by the good contribution of all the users, what a great forum! Thank you very much for your contribution and I like the way you solved the fuzzy image to improve it.

Not as smooth as the others but sharper. With AviSynth+ and x264 CLI:

LWLibavVideoSource("Noise Distort.mov", cache=false, prefer_hw=2) 
ConvertBits(8)
ConvertToYV12()
Crop(240,0,-240,-0)
source = last

# downscale closer to real resolution
Spline36Resize(width/4, 272)
Santiag() # antialias
MCTemporalDenoise(settings="very high") # noise reduction

# upscale back to source resolution, plus sharpening
Sharpen(0.5, 0.0)
nnedi3_rpow2(2, cshift="Spline36Resize")
aWarpSharp2(depth=5)
Sharpen(0.5)
nnedi3_rpow2(2, cshift="Spline36Resize", fwidth=source.width, fheight=source.height)
aWarpSharp2(depth=5)
Sharpen(0.5)
# use less sharpening if you want less processed-looking video

# multithreading
prefetch(8)

[dg3duy]

What an excellent program Hybrid with the torch-addon, I wonder what combination you made so that the moving image would not deteriorate, that it would be perfectly preserved and the background would be stabilized, as a novice it is very easy to damage the video without knowing what filter really needs to be activated, one wrong click and I think one can get bad results.

I used Hybrid with the torch-addon (this requires a newer NVIDIA card) to use BasicVSR++, loaded your source, enabled BasicVSR++ (Filtering->Vapoursynth->Other->BasicVSR++ with FP16 to speed things up), configured the encoder, set an output, created the jobs and let it run.

[Selur]

iirc. only thing I used was crop + BasicVSR++

[dg3duy]

iirc. only thing I used was crop + BasicVSR++

It is like touching the sky with your hands, in due time you will receive a blessing by paypal.

[dg3duy]

I have a query that is quite unusual, the result obtained is too perfect, how is it possible to make an intermediate step between the original video and the final result, so that it does not look totally synthetic or too perfect, BasicVSR++ (Filtering->Vapoursynth->Other->BasicVSR++ with FP16, this way is no longer to achieve the intermediate result and it is necessary to make a totally different workflow?

iirc. only thing I used was crop + BasicVSR++

[Selur]

Sadly, No. Mchine learning model filters usually don't have many options.
You can try increasing and lowering length and try the other models of BasicVSR++, but it will likely not do much.
Since your source does not really have details, lowering the resolution before filtering might help a lot with the speed.
Also, as a general hint: those machine learning filters will behave differently when fed with a different content/resolution

Alternatively, a combination of MCTemporalDenoise and SCUNet might also work fine.

[dg3duy]

Now I have a clearer picture, thank you very much for the clear answer, it is such a big world that I have yet to understand and experience, step by step.

Sadly, No. Mchine learning model filters usually don't have many options.
You can try increasing and lowering length and try the other models of BasicVSR++, but it will likely not do much.
Since your source does not really have details, lowering the resolution before filtering might help a lot with the speed.
Also, as a general hint: those machine learning filters will behave differently when fed with a different content/resolution

Alternatively, a combination of MCTemporalDenoise and SCUNet might also work fine.

[jagabo]

You could try using a weighted Merge() to blend the original back into the filtered version.

Merge(original, filtered, weight=0.5)

Or blend with Overlay() with weighting and optionally a mask of some sort.

mymask = BuildSomeSortOfMask() # motion? edges? colors? brightness? difference?
Overlay(source, filtered, mode="blend", opacity=0.5, mask=mymask)

[dg3duy]

Interesting idea, what is the Merge script or plugin, where do you get it and what software is it for.
AviSynth+ or other software? Quite a novice question, sorry.

You could try using a weighted Merge() to blend the original back into the filtered version.

Merge(original, filtered, weight=0.5)

Or blend with Overlay() with weighting and optionally a mask of some sort.

mymask = BuildSomeSortOfMask() # motion? edges? colors? brightness? difference?
Overlay(source, filtered, mode="blend", opacity=0.5, mask=mymask)

[jagabo]

Yes, everything in that post was AviSynth or AviSynth+.

[dg3duy]

Perfect! thank you very much for the information, very valuable.

Yes, everything in that post was AviSynth or AviSynth+.