madVR - high quality video renderer (GPU assisted)

[NeonPhoenix]

Originally Posted by Siso

Why do you need it anyway?

24 fps is too low, i wanted it for frame interpolation

[huhn]

so this wasn't good enough or what'S going on now?
https://github.com/HopperLogger/HopperRender

[Siso]

Originally Posted by NeonPhoenix

24 fps is too low, i wanted it for frame interpolation

If you want it for movies, movies are 24 fps and that's it. Any other manipulation is artificial (soap opera effect).

[strumf666]

Fell free to enjoy the original 24fps, there is a lot of us that prefer higher fps.

[strumf666]

Originally Posted by huhn

so this wasn't good enough or what'S going on now?
https://github.com/HopperLogger/HopperRender

Tried it and seems to work well enough after setting the scene change to ~1000ms. It is quite gpu intensive had to lower the chroma upscaling to NGU medium from very high to avoid frame drops (4k hdr-passthrough 23.976 source on 120Hz display amd 6900xt). I found one issue though, it conflicts with madvr display mode matching when the interpolation is set to off.
edit: the issue has been fixed.

[NeonPhoenix]

Originally Posted by huhn

so this wasn't good enough or what'S going on now?
https://github.com/HopperLogger/HopperRender

I asked about the best frame interpolation methods on reddit, but nobody mentioned this. Thanks a lot, I’ll give it a try and share my experience.

[NeonPhoenix]

Originally Posted by huhn

so this wasn't good enough or what'S going on now?
https://github.com/HopperLogger/HopperRender

I'm looking for something that does the same thing SVP + Rife does, HopperLogger seems more like low quality smooth motionish. I don't use SVP because it's not free anymore, is there any free tool that will get the job done like svp with rife?

[RealSnoopyDog]

I suggest to create a new thread because this is not related at all to madVR

[kirakami]

AMD Fluid Motion Frames and Nvidia Frame Generation/Smooth Motion and Super Resolution should be natively supported by every video player and renderer with enable/disable options to properly support High Hertz Monitors so you don't experience the slide show motion and judder on panning shots in 23.976 fps videos.

[huhn]

no they are not...
it is not meant for this anyway.

[flossy_cake]

I use RIFE in avisynth for realtime usage , not as part of SVP but just standalone .avs file played directly in MPC-HC with DirectShow LAV opening the .avs file and then LAV as the source filter in Avisynth:

PHP Code:

# Decode the source with LAV
source="C:\video.mkv"
audio = LWLibavAudioSource(source, stream_index=-1, cache=false)
video = LWLibavVideoSource(source, stream_index=-1, cache=false)
AudioDub(video, audio)

# Set multithreading mode 1 - I get much better GPU util on my 3060 with this mode, see http://avisynth.nl/index.php/SetFilterMTMode
SetFilterMTMode("RIFE", 1, true)   

# Convert to RGB 32bpc full range as required by RIFE.  Dithering not required - have tested on 8 bit grey ramp
z_ConvertFormat(pixel_type="RGBPS", colorspace_op="709:709:709:limited=>rgb:709:709:full").Prefetch(1)

# Double the frame rate using fast model 23 (rife-v4.6_ensembleFalse - https://github.com/Asd-g/AviSynthPlus-RIFE)
RIFE(factor_num=2, factor_den=1, model=23, gpu_thread=2).Prefetch(4) 

# Convert back to YUV 4:2:0 8bpc limited range
z_ConvertFormat(pixel_type="YV12", colorspace_op="rgb:709:709:full=>709:709:709:limited").Prefetch(1)

__END__ 


As huhn has mentioned, scenechanges are a real problem, especially if you do more than double frame rate interpolation such as 24->60 the scenechanges look AWFUL since there are multiple interpolated frames across the scenechange.

Just doing a frame rate doubling is not *too* terrible. I can still see it though, for 1/60th or 1/48th of a second you get a single frame like this interpolated across the scenechange:



So my next Avisynth project will be shoehorning my own better scenechange detection which I've already worked on over the last couple of years in a separate project. But it depends if RIFE plays nice with Avisynth's ScriptClip (I suspect it might not given how heavy RIFE is)

By default RIFE disables scenechange detection, and if you enable it with sc=true it appears to be just using "if Average Y changed by more than 12% then it was a scenechange frame" which is bad as it can trigger false positives on fast pans and that causes jerks because interpolation is disabled for scenechange frames. Shows like Top Gear with fast dramatic pans are often stuttering on those fast pans due to scenechange false positive. It's probably better to just leave RIFE's scenechange detection disabled unless you are doing 24->60 then you will really need it as otherwise the scenechange interpolation will occur over several frames (several warped/distored frames in a row like the one above)

RIFE interpolation quality is very good though, the lowest amount of artefacting I've ever seen in an algorithm including consumer TVs! For example it can handle complex hand gestures quite well without noticeable distortions. However you need a powerful GPU, the most my 3060 can manage is a 1440x1080p24 -> p48 scenario. For older 480p30/576p25 content it smashes through it easy due to the lower res, I'm using it on some crappy 480p30 transcode of Roseanne sitcom cause that show was shot at 60i and the end result looks faithful imo.

__________________________________________________________________

Just a comment about MadVR's "1 repeat frame every..." stat on ctrl+J - this appears to be reporting how long it would take for an entire source frame to be repeated from the perspective of the source file rate, not from the perspective of the monitor refresh rate.

eg. for a 59.94fps file played at 120.0hz, MadVR reports 1 repeat every 16.66 seconds but actually there are 2 duplicates during that 16.66 second period - one every 8.333 seconds. A minor quibble though.

[NeonPhoenix]

Originally Posted by flossy_cake

I use RIFE in avisynth for realtime usage , not as part of SVP but just standalone .avs file played directly in MPC-HC with DirectShow LAV opening the .avs file and then LAV as the source filter in Avisynth:

PHP Code:

# Decode the source with LAV
audio = LWLibavAudioSource(source, stream_index=-1, cache=false)
video = LWLibavVideoSource(source, stream_index=-1, cache=false)
AudioDub(video, audio)

# Set multithreading mode 1 - I get much better GPU util on my 3060 with this mode, see http://avisynth.nl/index.php/SetFilterMTMode
SetFilterMTMode("RIFE", 1, true)   

# Convert to RGB 32bpc full range as required by RIFE.  Dithering not required - have tested on 8 bit grey ramp
z_ConvertFormat(pixel_type="RGBPS", colorspace_op="709:709:709:limited=>rgb:709:709:full").Prefetch(1)

# Double the frame rate using fast model 23 (rife-v4.6_ensembleFalse - https://github.com/Asd-g/AviSynthPlus-RIFE)
RIFE(factor_num=2, factor_den=1, model=23, gpu_thread=2).Prefetch(4) 

# Convert back to YUV 4:2:0 8bpc limited range
z_ConvertFormat(pixel_type="YV12", colorspace_op="rgb:709:709:full=>709:709:709:limited").Prefetch(1)

__END__ 


As huhn has mentioned, scenechanges are a real problem, especially if you do more than double frame rate interpolation such as 24->60 the scenechanges look AWFUL since there are multiple interpolated frames across the scenechange.

Just doing a frame rate doubling is not *too* terrible. I can still see it though, for 1/60th or 1/48th of a second you get a single frame like this interpolated across the scenechange:



So my next Avisynth project will be shoehorning my own better scenechange detection which I've already worked on over the last couple of years in a separate project. But it depends if RIFE plays nice with Avisynth's ScriptClip (I suspect it might not given how heavy RIFE is)

By default RIFE disables scenechange detection, and if you enable it with sc=true it appears to be just using "if Average Y changed by more than 12% then it was a scenechange frame" which is bad as it can trigger false positives on fast pans and that causes jerks because interpolation is disabled for scenechange frames. Shows like Top Gear with fast dramatic pans are often stuttering on those fast pans due to scenechange false positive. It's probably better to just leave RIFE's scenechange detection disabled unless you are doing 24->60 then you will really need it as otherwise the scenechange interpolation will occur over several frames (several warped/distored frames in a row like the one above)

RIFE interpolation quality is very good though, the lowest amount of artefacting I've ever seen in an algorithm including consumer TVs! For example it can handle complex hand gestures quite well without noticeable distortions. However you need a powerful GPU, the most my 3060 can manage is a 1440x1080p24 -> p48 scenario. For older 480p30/576p25 content it smashes through it easy due to the lower res, I'm using it on some crappy 480p30 transcode of Roseanne sitcom cause that show was shot at 60i and the end result looks faithful imo.

__________________________________________________________________

Just a comment about MadVR's "1 repeat frame every..." stat on ctrl+J - this appears to be reporting how long it would take for an entire source frame to be repeated from the perspective of the source file rate, not from the perspective of the monitor refresh rate.

eg. for a 59.94fps file played at 120.0hz, MadVR reports 1 repeat every 16.66 seconds but actually there are 2 duplicates during that 16.66 second period - one every 8.333 seconds. A minor quibble though.

Could you provide an installation guide please? I failed to make it work after 2 hours of chatting with claude.

[flossy_cake]

Originally Posted by NeonPhoenix

Could you provide an installation guide please? I failed to make it work after 2 hours of chatting with claude.

I don't know if it's worth it for you because I also use an ahk script to automatically generate an avisynth script for each video file otherwise you'd have to manually make a script for every video file which is not really practical

It might be better in your case to use CrendKing's avisynth directshow filter which inserts avisynth between LAV Video Decoder and MadVR in the directshow graph. In that case you can omit the first 3 lines of the script cause it's already getting the source from directshow LAV

CrendKing's filter can be added manually in MPC-HC under external filters -> add filter. Then to test its working you can set the script to Text("hello world") to overlay those words on top of the video.

But if you don't want to use CrendKing's filter (cause it's kind of a hack and is not really the orthodox way that avisynth is intended to run) instead get klite "full" (which is MPC-HC + MadVR + LAV decoders) and then install avisynth (get the one with _vcredist.exe otherwise it wont work if you dont have that dependency...that one cost me a few hours of hair pulling recently )

Then make a new text file called test.avs and inside it write Version() and that will create a small video clip that prints the version of avisynth installed to confirm its working

[kirakami]

I wish there was a version of motion interpolation for mpv and mpc-be that simply detects panning shots in videos and smoothens it while rest of the video plays as if no interpolation has been applied.

[Ilovetv9]

Using mpc-hc or mpc-be, updated Test Build 208 madvr settings window now uses widescreen rectangle instead of previous v0.92.17 square traditional settings window. This - https://imgbox.com/ph2NMZrP - screenshot is how new madvr settings window displays using 4:3 CRT pc monitor. See how the sides are cut off? Dragging window edges to resize seems disabled. Is it possible to edit/change madvr settings window size/ratio in any way possible? Would madvr author consider in future update adding the ability to edit/change settings window size or changing settings window to square 4:3 ratio when crt monitor is selected - ironically in the settings window?

[huhn]

your resolution is just below the minimum the option screen needs that's it 4:3 has nothing to do with that.

if tone mapping should ever get toned in it will go back to the normal factor anyway.

it also can not just switch it needs that size.