When to TDecimate/SRestore?

[Okiba]

Thanks everyone! Converting more videos during the weekend, and I think everything makes much more sense now. Two things worth noticing for future readers:

- For me personally, QTGMC Preset Medium was a sweet spot between good and too blurry. Specifically for cartoon. I still think Slow/Slower is better on my Camcorder videos.
- There was something that was confusing me. The cartoon I'm currently working on was TTF. With AssumeTTF() and QTGMC() it looks great. However, it also suffers from Frame-blending and required SRestore. After SRestore, I started seeing frame being skipped on pan shots. It seems like changing it from TTF to BFF, solve that issues and the panning was smooth. I'm not sure why - as the video was TFF. Perhaps it something to do with the conversion from NTSC/PAL. Something to keep an eye for when using SRestore.

Thanks again for the help everyone!

[jagabo]

There are a few possible reasons. Sometimes a video switches back and forth between TFF and BFF. Sometimes shots are sped up or slowed down by duplicating or discarding fields or frames, or with field/frame blending. Cuts/pastes can leave you with sequences that can't be perfectly restored. SRestore() takes a while to settle down and find the proper rhythm. If you random seek to a frame and step through the video from there you may get a different sequence than if you run it linearly from the start of the video (or a much earlier frame).

[Okiba]

I see this switch between TFF and BFF before. I even saw switch between Progressive and interlaced (where the intro/end-credit for example are localized). And yes, It happened to me in a specific video where settings 25FPS was perfect on a specific pan shot, but horrible on another, and when I switched to 23FPS, the first one was horrible and the 2nd one was great. Oh well, can't be perfect I guess

Thanks!

[Okiba]

Hi again everyone,

Here's another interesting example.

The video is very noisy. Probably been recorded on bad reception. Normally I wouldn't bother, but my sister wanted me to keep it as she has fond memory of it. Just looking on it, It's almost clear it's interlaced. But I Yadif it to 50FPS, and most of the panning shots I was checking were 1 movement per 2 frames (where the 2nd frame is very "noise"), like it's progressive. I finally found a section with movement per 1 frame (the video I attached).
So I tried to TFM it (and even Qtgmc.Srestore - though I don't see any blended frames) and the picture quality really getting degraded. There's things I can do to improve the quality, like degraining a bit and remove the comets but first I wanted to know - What happens here exactly? is that because the 2nd merged frame is super noisy? is that a side effect of the bad reception?

Thanks!

[johnmeyer]

Oh, it's a cartoon. All bets are off because cartoon field & frame repeats are all over the place.

[Sharc]

Yes. An attempt to improve it somewhat. Oh well ..... giving up.

[manono]

You can get rid of the noisy fields by bobbing and removing every other frame. Of course, that all goes to hell if it changes during the video, but it's okay for the sample:

QTGMC()
SelectOdd()
Crop(18,6,-18,-6)
BilinearResize(640,480)

[Okiba]

Oh, it's a cartoon. All bets are off because cartoon field & frame repeats are all over the place.

That's interlaced right? Why after bobbing 90% of the panning shot I checked were 1 movement per 2 frames? or that "blur/noise" on the 2nd frame actually point it to be interlaced but with a frame repeat issues? why was the 2nd frame always very dirty?

Yes. An attempt to improve it somewhat. Oh well ..... giving up.

That's not bad actually. If you still got the code, share it so I can learn something new

You can get rid of the noisy fields by bobbing and removing every other frame. Of course, that all goes to hell if it changes during the video, but it's okay for the sample:

Oh, nice. It's still dirtier compared to source, but much better then what TFM/QTGMC+SRestore can do.

Thanks!

[Sharc]

That's not bad actually. If you still got the code, share it so I can learn something new

AVISource("Foofur.avi").converttoYV12().assumeTFF()
TFM(mode=1,pp=6,scthresh=16.0,display=false)
dfttest(sigma=900).KNLMeansCL(h=18.0) #heavy denoising (also killing details)
LSFmod(strength=300)

You could try the denoising following manono's proposal.

[Selur]

Smooth as hell,..
QTGMC + Spotless + vsdpir with denoise strength 80

[Sharc]

Smooth as hell,..
QTGMC + Spotless + vsdpir with denoise strength 80

Wow! Amazing!

Edit: Is there an Avisynth equivalent of vsdpir? Or is it/can it be included in Hybrid?

[Alwyn]

My attempt.

[Selur]

seems like you used the wrong field order, at least here that clip is jumping back and forth the whole time,..

[Alwyn]

seems like you used the wrong field order, at least here that clip is jumping back and forth the whole time,..

Whoopsie! Fixed, I hope.

[Selur]

Yup, no more jumping around.

[Okiba]

Smooth as hell,..
QTGMC + Spotless + vsdpir with denoise strength 80

Wow, Impressive! I really need to try VapourSynth. It feels like the de-nosing makes everything about "Pastelish"? maybe even some lose of details? or It just me getting used to see pixels/noise?

My attempt.

Thanks!

[Okiba]

Hi again everyone!

I'm working on a video that has animated intro, but the episode itself is shot using a Camcorder. The Video is at 25FPS. The animated intro, has blended frames. SRestore(25) solve most of the issues. The actual episode however, is interlaced. I'm using QTGMC to bob it. Problem is, the end results of SRestore is 25fps, and the end results of QTGMC is 50FPS. So I can't mix them into a single output. Is there anything right to do about it? or I'll just bob the intro, live with the blended frames and move to the next video?

Thanks again!

[Selur]

So I can't mix them into a single output.

Unless you create a vfr output.

Is there anything right to do about it?

Personally I would:
- keep the episode at 50fps
- use the 25fps sRetsore intro output and use frame interpolation (RIFE, FrameRateConverter, SVP,..) or add duplicate frames to get the intro to 50fps. Frame interpolation might add some artifacts but, in my opinion, those are usually preferable to blended frames.

Cu Selur

[Okiba]

Unless you create a vfr output.

Oh nice! I didn't know something that exists. I'm familiar with VBR, but not vfr. I assume it's not very supported by player? or any player that can play FFMPEG files can do that?

use the 25fps sRetsore intro output and use frame interpolation (RIFE, FrameRateConverter, SVP,..) or add duplicate frames to get the intro to 50fps. Frame interpolation might add some artifacts but, in my opinion, those are usually preferable to blended frames.

Of course! That's makes absolute sense! Is it possible to do that with AviSynth? or I have to use one of the external programs you mentioned?

Thanks!

[Selur]

vfr = variable frame rate, if you haven't used it yet, don't start.

Is it possible to do that with AviSynth? or I have to use one of the external programs you mentioned?

Aside from RIFE those are all available as AviSynth filters/scripts.

Cu Selur

[jagabo]

Motion interpolation does not work well with animation. Just use ChangeFPS(50).

[Okiba]

Thank you Sellur and Jagabo. Out of curiosity will try both, see why Motion interpolation having problems with animation

[Okiba]

Updating ChangeFPS looks indeed better for this specific Video. But I have been playing with FrameRateConverter(), and it's another tool in my box. Thank you!
Most of my videos are cropped 8, 16, -8, 16. So I set the SAR to be 12:11. I found today a animated video that the crop was larger. 6, 32, -6, 32. Is this identical for NTSC being converted to PAL? So the SAR is actually 10:11?

[Selur]

Is there an Avisynth equivalent of vsdpir? Or is it/can it be included in Hybrid?

totally missed that, vsdpir (https://github.com/HolyWu/vs-dpir), doubt there will be a Avisynth port since it used pytorch.
Got an addon for Hybird (atm. ~9 GB download; mainly vsgan models and pytorch dependency) which adds VSGAN , DPIR (https://github.com/HolyWu/vs-dpir), FFDnet (https://github.com/HolyWu/vs-ffdnet/tree/master/vsffdnet) and vs-RIFE (https://github.com/HolyWu/vs-rife). Send my a pm about it and I can send you a link.
--
Okiba: got a small sample of the animation? I'd like to see how RIFE will handle it.

Cu Selur

[Okiba]

Yea sure. I left the external Storage on my father house because he wanted to watch some Camcorder footage, so over the weekend when I'll pick it up - I'll share this specific video.

[Sharc]

Most of my videos are cropped 8, 16, -8, 16. So I set the SAR to be 12:11. I found today a animated video that the crop was larger. 6, 32, -6, 32. Is this identical for NTSC being converted to PAL? So the SAR is actually 10:11?

Cropping does not change the SAR (=Sample Aspect Ratio)

NTSC 4:3 704x480 mpeg4: SAR=10:11
NTSC 4:3 720x480 (mpeg2 DVD): SAR=8:9
PAL 4:3 704x576 mpeg4: SAR=12:11
PAL 4:3 720x576 (mpeg2 DVD): SAR=16:15

DVD do not even specify the SAR (or PAR, Pixel Aspect Ratio) directly, but specify the DAR (4:3 or 16:9) instead. The corresponding SAR (or PAR) is a derived quantity.

[Okiba]

So if a 720x576 Pal video, was being cropped to remove black bars to 656x564. That would still be SAR? I'm confused as it doesn't align with any of the above.

[jagabo]

Of course the sampling aspect ratio doesn't change when you crop. Consider a piece of graph paper full of little squares. Crop as much as you want, the squares on the remaining paper will still be square.

[Selur]

This is why I prefer using the term 'Pixel Aspect Ratio' over 'Sample Aspect Ratio'. (+ SAR sometimes is used as 'Source Aspect Ratio')

[Okiba]

Oh that PIXEL aspect Ratio! I was sure it reference the aspect ratio of final "view". Oh, OK. So yea. 12:11 it is, thank you