Help with better upscale on this source

[Isac.martins]

Good afternoon people.
Could you help me with a suitable script for this font?
I want to Upscale using Avisynth, I would like help with which filters and script to use. I tried to upscale using Topaz Video Enhancer, but as the file has a lot of compression artifacts, the software does a terrible job, as people end up not looking human, but rather like a doll, similar to dolls or mannequins.
Is there a better possibility with Avisynth for this task? The file has a resolution of 704x576

[johnmeyer]

What exactly do you hope to gain by upscaling? I don't think your video will look even 1% better.

The video has many other issues. It is 25 fps progressive PAL. Given Madonna's age (relatively young), this video predates HD and was therefore either shot on film (24 fps) or on video, either PAL interlaced or NTSC interlaced. I see evidence that its frame rate has been converted, so I am betting that somewhere someone converted this from 29.97 fps interlaced to 25 fps progressive. Your time would be far better spent trying to reduce those artifacts. I don't know if that can be done. Others will have to chime in.

However, upscaling is, IMHO, a waste of your time.

[davexnet]

I mostly agree with John, the source is so poor, looks like it was poorly deinterlaced or frame rate conversion,
but with some NR and chroma NR, it's already looking a little better.
Here's a quick job with the aforementioned NR and upscale to 720p using NNEDi3_rpow2.
I have not addressed any blends or other issues

[Marcio.ciconne]

What exactly do you hope to gain by upscaling? I don't think your video will look even 1% better.

The video has many other issues. It is 25 fps progressive PAL. Given Madonna's age (relatively young), this video predates HD and was therefore either shot on film (24 fps) or on video, either PAL interlaced or NTSC interlaced. I see evidence that its frame rate has been converted, so I am betting that somewhere someone converted this from 29.97 fps interlaced to 25 fps progressive. Your time would be far better spent trying to reduce those artifacts. I don't know if that can be done. Others will have to chime in.

However, upscaling is, IMHO, a waste of your time.

Well, I understand that the file presents several technical problems. However, this same show based on the same source has already been restored by a fan who left the Upscale quality acceptable and without the doll effect that I mentioned, I just don't know which techniques he used:

https://youtu.be /pEI6C_TetEU?si= EiUetA7KxZjFWKC4

The problem is that he simply couldn't agree with the colors close to the original capture of the show, as the file is a proshot with low lighting and color capture, so I made this color correction from the legitimate source (the same one he used to make the sophisticated), but I don't know of any technique to achieve the same improvement he achieved.

[Marcio.ciconne]

I mostly agree with John, the source is so poor, looks like it was poorly deinterlaced or frame rate conversion,
but with some NR and chroma NR, it's already looking a little better.
Here's a quick job with the aforementioned NR and upscale to 720p using NNEDi3_rpow2.
I have not addressed any blends or other issues

Yes, that looks better 👏🏻

[Selur]

With Avisynth, try mlrt with dpir and adding some grain afterward might help to make it more watchable, but I'm with johnmeyer and doubt that anything 'good' will come out of trying to upscale this. With Vapoursynth I would suggest BasicVSR++. https://imgsli.com/MjA4Njg2, https://imgsli.com/MjA4Njg3
Maybe after deblocking and denoising, some ai filter can add some none existent details, but to get rid of that blocking&co some heavy filtering is needed which will destroy details.

Cu Selur

Ps.: that the second link definitely didn't use the same source that you shared.

[johnmeyer]

That YouTube video has it's own issues, as can be seen in this screen capture:

[Marcio.ciconne]

With Avisynth, try mlrt with dpir and adding some grain afterward might help to make it more watchable, but I'm with johnmeyer and doubt that anything 'good' will come out of trying to upscale this. With Vapoursynth I would suggest BasicVSR++. https://imgsli.com/MjA4Njg2, https://imgsli.com/MjA4Njg3
Maybe after deblocking and denoising, some ai filter can add some none existent details, but to get rid of that blocking&co some heavy filtering is needed which will destroy details.

Cu Selur

Ps.: that the second link definitely didn't use the same source that you shared.

This was the legitimate source he used, the difference is that there is no color, luminosity correction

[Marcio.ciconne]

That YouTube video has it's own issues, as can be seen in this screen capture:

Is true

[Marcio.ciconne]

I mostly agree with John, the source is so poor, looks like it was poorly deinterlaced or frame rate conversion,
but with some NR and chroma NR, it's already looking a little better.
Here's a quick job with the aforementioned NR and upscale to 720p using NNEDi3_rpow2.
I have not addressed any blends or other issues

Could you share the script with which you achieved this result, please?

[jagabo]

This was the legitimate source he used

Not horrible considering the source quality...

[Marcio.ciconne]

I mostly agree with John, the source is so poor, looks like it was poorly deinterlaced or frame rate conversion,
but with some NR and chroma NR, it's already looking a little better.
Here's a quick job with the aforementioned NR and upscale to 720p using NNEDi3_rpow2.
I have not addressed any blends or other issues

I can't find these two filters (NR and chroma NR) to assemble the script you used =(

[davexnet]

I mostly agree with John, the source is so poor, looks like it was poorly deinterlaced or frame rate conversion,
but with some NR and chroma NR, it's already looking a little better.
Here's a quick job with the aforementioned NR and upscale to 720p using NNEDi3_rpow2.
I have not addressed any blends or other issues

I can't find these two filters (NR and chroma NR) to assemble the script you used =(

NR stands for noise reduction and I meant it in a general sense; that's not the name of any filter.

This is the script I used:

loadplugin("i:\plugins\cnr2.dll")
lwlibavvideosource("C:\Users\davex\Downloads\source.mp4")
ColorMatrix(mode="Rec.601->Rec.709")
mc_spudsmod(mode="low",removeblocks=true)
NNEDI3_rpow2(rfactor=2,nns=4, qual=2,cshift="spline36resize",fwidth=1280,fheight=720)
cnr2()

[Marcio.ciconne]

I mostly agree with John, the source is so poor, looks like it was poorly deinterlaced or frame rate conversion,
but with some NR and chroma NR, it's already looking a little better.
Here's a quick job with the aforementioned NR and upscale to 720p using NNEDi3_rpow2.
I have not addressed any blends or other issues

I can't find these two filters (NR and chroma NR) to assemble the script you used =(

NR stands for noise reduction and I meant it in a general sense; that's not the name of any filter.

This is the script I used:

loadplugin("i:\plugins\cnr2.dll")
lwlibavvideosource("C:\Users\davex\Downloads\source.mp4")
ColorMatrix(mode="Rec.601->Rec.709")
mc_spudsmod(mode="low",removeblocks=true)
NNEDI3_rpow2(rfactor=2,nns=4, qual=2,cshift="spline36resize",fwidth=1280,fheight=720)
cnr2()

I was racking my brain haha, thank you very much.
It just gave an error in the filter (mc_spudsmod) with the message ''there is no function called 'deblock_qed'''
I looked for the error on Google, but I can't find a solution
If you know any solution I would be very grateful and I won't bother you anymore lol

[davexnet]

Look for, download and add to the Avisynth plugin folder(s)
.dll and .avsi (scripts) will load automatically if placed in the proper place,
Otherwise you load them using Loadplugin or import respectively

[Marcio.ciconne]

Look for, download and add to the Avisynth plugin folder(s)
.dll and .avsi (scripts) will load automatically if placed in the proper place,
Otherwise you load them using Loadplugin or import respectively

All good, thank you so much !

[poisondeathray]

The source is a ConvertFPS style blend conversion from NTSC to PAL 25i . The source deblends fairly well with RestoreFPS to 29.97 after double rate deinterlacing.

The YT version is 30 FPS - but it has blends , is jerky, and it was not processed properly .

You can interpolate to fake 59.94p using rife or similar and it will look smoother, at the risk of artifacts. Unfortunately this source has blended scene changes and that causes extra problems. The avs version of RIFE has bad scene change detection in the first place (a lower scenechange threshold will cause some frames to duplicate and stutter in the middle of sequences because of false positive detection). One option is to use remapframes to specify frames to duplicate for the bad scenechange frames, or blend them (mix a convertfps version). Alternatively, you can interpolate 1 direction using DMVFN to fix the scene changes completely with unique frames, but there is no avs /vpy implementation yet

[Marcio.ciconne]

The source is a ConvertFPS style blend conversion from NTSC to PAL 25i . The source deblends fairly well with RestoreFPS to 29.97 after double rate deinterlacing.

The YT version is 30 FPS - but it has blends , is jerky, and it was not processed properly .

You can interpolate to fake 59.94p using rife or similar and it will look smoother, at the risk of artifacts. Unfortunately this source has blended scene changes and that causes extra problems. The avs version of RIFE has bad scene change detection in the first place (a lower scenechange threshold will cause some frames to duplicate and stutter in the middle of sequences because of false positive detection). One option is to use remapframes to specify frames to duplicate for the bad scenechange frames, or blend them (mix a convertfps version). Alternatively, you can interpolate 1 direction using DMVFN to fix the scene changes completely with unique frames, but there is no avs /vpy implementation yet

Wow, this is really incredible, I've never seen this file with this frame interpolation, the restorations always had the wrong rate, it's incredible and smooth!
I have very little experience in Avisynth, unfortunately I still don't know how to add values to the filters that would require specifying the frames, could you share the script you used?
I would be very grateful.

[poisondeathray]

RestoreFPS is x86 only. You can use it in mp_pipeline (combine and run x86 and x64 functions in a either x86 or x64 host. x64 host is usually faster does not have memory amount restriction. Or you can divide up into steps with lossless intermediate. If you were using very slow filters, maybe that would be a better option

The Trim(2,0) was because of the way the sample was cut , there are 2 leading garbage b-frames (open GOP) that screw up the cadence. On a full source, you probably don't need it

The RestoreFPS phase might need to be adjusted. In this example it was "0.8" ,but for some scenes a different value might be slightly better e.g. 0.85, or 0.9 - or you can mix/match scenes/settings for best results. 0.8 was sort of the average , where all scenes looked ok

I left scenechange for RIFE off , because it really isn't very accurate. ReplaceFrames was used to fix the scenechanges with duplicates. Some of the blends before the scene change I left in, because it's smoother and looked ok - you'd have to go through the entire concert to fix it properly. If I was spending the time fixing it up properly - I would probably use DMVFN to interpolate the scene changes. Because you don't have a reference frame "after" before the scene change, or reference frame "before" the beginning of the next scene, you cannot interpolate using RIFE or optical flow. DMVFN is 1 directional interpolation using 2 previous frames

You can play with other deniosing /deblocking/upscaling filters - the main point of this post was really RestoreFPS to 29.97 - it gives you the best starting point for that source - clearer frames and reduction of blends. It's arguably the most important step no matter what filters you choose

MP_Pipeline("""

### platform: win64
LWLibavVideoSource("VTS_01_1.mkv")
trim(2,0)
bwdif(field=3)

### ###

### platform: win32
LoadPlugin("PATH\RestoreFPS.dll")
restorefps(29.97, 0.8)
### ###

### platform: win64
temporaldegrain2()
deblock_qed_mt2()
z_convertformat(pixel_type="RGBPS", colorspace_op="170m:709:709:l=>rgb:709:709:f")
rife()
z_convertformat(pixel_type="YV12", colorspace_op="rgb:709:709:f=>709:709:709:l")
LoadPlugin("PATH\avssr64.dll")
SR()
spline16resize(1280,720)
finesharp(sstr=1)
nonlinusm(str=0.25, rad=5)

remapframes(last, mappings="165 164
167 168
509 508
511 512
1045 1044
1047 1048
1149 1148
1151 1152
1281 1280
#1503 1502
1505 1506
#1889 1888
1891 1892
2005 2004
2006 2007
2363 2362
2365 2366
2511 2510
2513 2514
2609 2608
#2611 2612
2777 2776
2779 2780
", sourceClip=last)


propclearall()

### ###

""")

[Marcio.ciconne]

RestoreFPS is x86 only. You can use it in mp_pipeline (combine and run x86 and x64 functions in a either x86 or x64 host. x64 host is usually faster does not have memory amount restriction. Or you can divide up into steps with lossless intermediate. If you were using very slow filters, maybe that would be a better option

The Trim(2,0) was because of the way the sample was cut , there are 2 leading garbage b-frames (open GOP) that screw up the cadence. On a full source, you probably don't need it

The RestoreFPS phase might need to be adjusted. In this example it was "0.8" ,but for some scenes a different value might be slightly better e.g. 0.85, or 0.9 - or you can mix/match scenes/settings for best results. 0.8 was sort of the average , where all scenes looked ok

I left scenechange for RIFE off , because it really isn't very accurate. ReplaceFrames was used to fix the scenechanges with duplicates. Some of the blends before the scene change I left in, because it's smoother and looked ok - you'd have to go through the entire concert to fix it properly. If I was spending the time fixing it up properly - I would probably use DMVFN to interpolate the scene changes. Because you don't have a reference frame "after" before the scene change, or reference frame "before" the beginning of the next scene, you cannot interpolate using RIFE or optical flow. DMVFN is 1 directional interpolation using 2 previous frames

You can play with other deniosing /deblocking/upscaling filters - the main point of this post was really RestoreFPS to 29.97 - it gives you the best starting point for that source - clearer frames and reduction of blends. It's arguably the most important step no matter what filters you choose

MP_Pipeline("""

### platform: win64
LWLibavVideoSource("VTS_01_1.mkv")
trim(2,0)
bwdif(field=3)

### ###

### platform: win32
LoadPlugin("PATH\RestoreFPS.dll")
restorefps(29.97, 0.8)
### ###

### platform: win64
temporaldegrain2()
deblock_qed_mt2()
z_convertformat(pixel_type="RGBPS", colorspace_op="170m:709:709:l=>rgb:709:709:f")
rife()
z_convertformat(pixel_type="YV12", colorspace_op="rgb:709:709:f=>709:709:709:l")
LoadPlugin("PATH\avssr64.dll")
SR()
spline16resize(1280,720)
finesharp(sstr=1)
nonlinusm(str=0.25, rad=5)

remapframes(last, mappings="165 164
167 168
509 508
511 512
1045 1044
1047 1048
1149 1148
1151 1152
1281 1280
#1503 1502
1505 1506
#1889 1888
1891 1892
2005 2004
2006 2007
2363 2362
2365 2366
2511 2510
2513 2514
2609 2608
#2611 2612
2777 2776
2779 2780
", sourceClip=last)


propclearall()

### ###

""")

Thank you very much for that, friend =)

[Marcio.ciconne]

RestoreFPS is x86 only. You can use it in mp_pipeline (combine and run x86 and x64 functions in a either x86 or x64 host. x64 host is usually faster does not have memory amount restriction. Or you can divide up into steps with lossless intermediate. If you were using very slow filters, maybe that would be a better option

The Trim(2,0) was because of the way the sample was cut , there are 2 leading garbage b-frames (open GOP) that screw up the cadence. On a full source, you probably don't need it

The RestoreFPS phase might need to be adjusted. In this example it was "0.8" ,but for some scenes a different value might be slightly better e.g. 0.85, or 0.9 - or you can mix/match scenes/settings for best results. 0.8 was sort of the average , where all scenes looked ok

I left scenechange for RIFE off , because it really isn't very accurate. ReplaceFrames was used to fix the scenechanges with duplicates. Some of the blends before the scene change I left in, because it's smoother and looked ok - you'd have to go through the entire concert to fix it properly. If I was spending the time fixing it up properly - I would probably use DMVFN to interpolate the scene changes. Because you don't have a reference frame "after" before the scene change, or reference frame "before" the beginning of the next scene, you cannot interpolate using RIFE or optical flow. DMVFN is 1 directional interpolation using 2 previous frames

You can play with other deniosing /deblocking/upscaling filters - the main point of this post was really RestoreFPS to 29.97 - it gives you the best starting point for that source - clearer frames and reduction of blends. It's arguably the most important step no matter what filters you choose

MP_Pipeline("""

### platform: win64
LWLibavVideoSource("VTS_01_1.mkv")
trim(2,0)
bwdif(field=3)

### ###

### platform: win32
LoadPlugin("PATH\RestoreFPS.dll")
restorefps(29.97, 0.8)
### ###

### platform: win64
temporaldegrain2()
deblock_qed_mt2()
z_convertformat(pixel_type="RGBPS", colorspace_op="170m:709:709:l=>rgb:709:709:f")
rife()
z_convertformat(pixel_type="YV12", colorspace_op="rgb:709:709:f=>709:709:709:l")
LoadPlugin("PATH\avssr64.dll")
SR()
spline16resize(1280,720)
finesharp(sstr=1)
nonlinusm(str=0.25, rad=5)

remapframes(last, mappings="165 164
167 168
509 508
511 512
1045 1044
1047 1048
1149 1148
1151 1152
1281 1280
#1503 1502
1505 1506
#1889 1888
1891 1892
2005 2004
2006 2007
2363 2362
2365 2366
2511 2510
2513 2514
2609 2608
#2611 2612
2777 2776
2779 2780
", sourceClip=last)


propclearall()

### ###

""")

Is this ''nonlinusm'' a filter?
I can't find it for download

[poisondeathray]

Is this ''nonlinusm'' a filter?
I can't find it for download

It's a function for "non-linear Unsharp Masking"

http://forum.doom9.org/showpost.php?p=1555234&postcount=46

[Joker Young]

Hello, I tried to upscale this source by 200%.
[Attachment 74560 - Click to enlarge]

[Isac.martins]

RestoreFPS is x86 only. You can use it in mp_pipeline (combine and run x86 and x64 functions in a either x86 or x64 host. x64 host is usually faster does not have memory amount restriction. Or you can divide up into steps with lossless intermediate. If you were using very slow filters, maybe that would be a better option

The Trim(2,0) was because of the way the sample was cut , there are 2 leading garbage b-frames (open GOP) that screw up the cadence. On a full source, you probably don't need it

The RestoreFPS phase might need to be adjusted. In this example it was "0.8" ,but for some scenes a different value might be slightly better e.g. 0.85, or 0.9 - or you can mix/match scenes/settings for best results. 0.8 was sort of the average , where all scenes looked ok

I left scenechange for RIFE off , because it really isn't very accurate. ReplaceFrames was used to fix the scenechanges with duplicates. Some of the blends before the scene change I left in, because it's smoother and looked ok - you'd have to go through the entire concert to fix it properly. If I was spending the time fixing it up properly - I would probably use DMVFN to interpolate the scene changes. Because you don't have a reference frame "after" before the scene change, or reference frame "before" the beginning of the next scene, you cannot interpolate using RIFE or optical flow. DMVFN is 1 directional interpolation using 2 previous frames

You can play with other deniosing /deblocking/upscaling filters - the main point of this post was really RestoreFPS to 29.97 - it gives you the best starting point for that source - clearer frames and reduction of blends. It's arguably the most important step no matter what filters you choose

MP_Pipeline("""

### platform: win64
LWLibavVideoSource("VTS_01_1.mkv")
trim(2,0)
bwdif(field=3)

### ###

### platform: win32
LoadPlugin("PATH\RestoreFPS.dll")
restorefps(29.97, 0.8)
### ###

### platform: win64
temporaldegrain2()
deblock_qed_mt2()
z_convertformat(pixel_type="RGBPS", colorspace_op="170m:709:709:l=>rgb:709:709:f")
rife()
z_convertformat(pixel_type="YV12", colorspace_op="rgb:709:709:f=>709:709:709:l")
LoadPlugin("PATH\avssr64.dll")
SR()
spline16resize(1280,720)
finesharp(sstr=1)
nonlinusm(str=0.25, rad=5)

remapframes(last, mappings="165 164
167 168
509 508
511 512
1045 1044
1047 1048
1149 1148
1151 1152
1281 1280
#1503 1502
1505 1506
#1889 1888
1891 1892
2005 2004
2006 2007
2363 2362
2365 2366
2511 2510
2513 2514
2609 2608
#2611 2612
2777 2776
2779 2780
", sourceClip=last)


propclearall()

### ###

""")

Friend, could you help me? In the RestoreFPS phase I need to run on the AvsPmod 32-bit platform, right?
I downloaded the RestoreFPS DLL and placed it in the 32-bit plugin folder, but I can't execute the line: LoadPlugin("PATH\RestoreFPS.dll"), I get this error in the photo.
I'm sorry, I have little time with Avisynth and I wanted to understand this part

[davexnet]

If you're using MP_Pipeline for 32 and 64 bit sections of your script,
You can certainly open it with a 64 bit host app.
Did you try that?

[Isac.martins]

If you're using MP_Pipeline for 32 and 64 bit sections of your script,
You can certainly open it with a 64 bit host app.
Did you try that?

Do I need to run the ''MP_Pipeline'' plugin at the beginning of the script for each platform?
If so, I'm not able to run it either

[davexnet]

This is poisondeathray's example:

MP_Pipeline("""

### platform: win32
MPEG2Source("F:\A Video Files\DVD Rips\VTS_01_1.d2v", CPU2="ooooxx", Info=3)
### ###

### platform: win64
AssumeTFF()
AssumeFPS(25)
QTGMC(preset="slow", matchpreset="slow", matchpreset2="slow", sourcematch=3, Lossless=2, TR2=2, EdiThreads=2, border=true)
Spline64Resize(960,720)
Crop(0,0,-6,0)
AddBorders(0,0,6,0)
### ###

### platform: win32
LoadPlugin("C:\Program Files (x86)\AviSynth\plugins\asharp.dll")##
asharp()
### ###
""")

[Isac.martins]

This is poisondeathray's example:

MP_Pipeline("""

### platform: win32
MPEG2Source("F:\A Video Files\DVD Rips\VTS_01_1.d2v", CPU2="ooooxx", Info=3)
### ###

### platform: win64
AssumeTFF()
AssumeFPS(25)
QTGMC(preset="slow", matchpreset="slow", matchpreset2="slow", sourcematch=3, Lossless=2, TR2=2, EdiThreads=2, border=true)
Spline64Resize(960,720)
Crop(0,0,-6,0)
AddBorders(0,0,6,0)
### ###

### platform: win32
LoadPlugin("C:\Program Files (x86)\AviSynth\plugins\asharp.dll")##
asharp()
### ###
""")

It still gives me this error

[davexnet]

Is that the complete script? Since you're only running one section, no need for MP_Pipeline at all,
you can remove it. Just open the sdcript in a 32 bit host
Are you running the latest version of Avisynth Plus ?

[Isac.martins]

Is that the complete script? Since you're only running one section, no need for MP_Pipeline at all,
you can remove it. Just open the sdcript in a 32 bit host
Are you running the latest version of Avisynth Plus ?

I'm using version ''AviSynthPlus 3.7.3''
MPEG2Source cannot open the file, so I opened it with LWLibavVideoSource, and this error appears when I try to run ''RestoreFPS''