Need help to find the best possible setting (Page 1 of 2)

If you want to get a really good quality with the least amount of artifacts possible use this: https://www.svp-team.com/wiki/RIFE_AI_interpolation

About halo effect around moving objects read here:
http://forum.doom9.net/showpost.php?p=1 … stcount=17
http://forum.doom9.net/showpost.php?p=1 … stcount=24

About anime read this: https://www.svp-team.com/forum/viewtopi … 308#p80308

Fast paced scenes are difficult for any model. Specifically created model for this is here:  https://github.com/google-research/frame-interpolation

If you want the best possible quality for anime, look for AI models with the best score in the ATD12K test.

If you want the best possible quality for fast paced scenes, look for AI models with the best score in the Xiph-2k test.

You can search for scores and compare them here among others:
https://arxiv.org/pdf/2204.03513.pdf
https://arxiv.org/pdf/2104.02495.pdf
https://arxiv.org/pdf/2202.04901.pdf

If you additionally want reasonably good speed and ease of use, then only SVP + RIFE filter for VapourSynth (ncnn Vulkan)

You have the same priorities as me: less halo and less artifacts during realtime interpolation.

Solution: x5 interpolation for 1920x1080 23.976 fps files in realtime using SVP + RIFE filter for VapourSynth (ncnn Vulkan)

Why x5? Better smoothess than x3 and both x3 and x5 give the best interpolation result with RIFE: https://www.svp-team.com/forum/viewtopi … 345#p80345

To do x5 interpolation for 1920x1080 23.976 fps files in realtime we need 116% more powerful graphics card than ZOTAC GAMING GeForce RTX 3070 Ti AMP Extreme Holo: https://www.svp-team.com/forum/viewtopi … 480#p80480

This will be possible soon with the NVIDIA GeForce RTX 4090.

http://forum.doom9.net/showpost.php?p=1 … tcount=312

http://forum.doom9.net/showpost.php?p=1 … tcount=323

http://forum.doom9.net/showpost.php?p=1 … tcount=324

You asked the same question almost 3 years ago and got plenty of answers: https://www.svp-team.com/forum/viewtopic.php?id=5500

You spoke up when others asked about similar problems: https://www.svp-team.com/forum/viewtopic.php?id=6139

Others have asked before you and someone even stated that he preferred FrameRateConverter over SVP for this reason: https://www.svp-team.com/forum/viewtopic.php?id=4382

I just showed You that RIFE beats both mvtools2 and FrameRateConverter in what you asked about.

It can be seen that You care a lot about this and over the years you have not found a solution that satisfies you.

Do you think there is some magical way to make the base SVP algorithm capable of approaching what artificial intelligence can achieve? 

Base SVP interpolation uses only a tiny fraction of the computing power that RIFE requires.

You can spend hours testing every possible setting in SVP yourself, or you can use that time to do extra work that will allow you to upgrade your GPU. You can also take a shortcut, although I don't recommend it: https://www.svp-team.com/forum/viewtopi … 388#p80388

You are probably writing about this: https://www.svp-team.com/forum/viewtopic.php?id=6488
Remember, there are probably 2 SVP developers and thousands of SVP users. The chance that some enthusiast finds a combination of settings that gives a slightly better visual effect is simply higher. However, this requires a great deal of time and a great deal of trying, and sometimes a little more luck.

Despite the fact that someone has found a slightly better setting than the others, you are still not completely happy. And you are not likely to be. The base algorithm has certain limitations that cannot be jumped over. It already has a huge number of setting combinations and this is a credit to the developers. Searching for optimal settings is the task of the users, who are also developers, but there are only two of them. And the search is an enormous amount of time.

That's why recently, thanks to the development of GPU computing power, a lot of machine learning-based frame interpolation models have started to appear. This makes it possible to speed up considerably to reach the ideal of artifact-free interpolation. There is no other way. Artificial intelligence is entering all areas of life.

If everyone supported AI models, we would show their developers that we care and that their work is not in vain. Thanks to them we are closer to the ideal of real-time artifact-free frame interpolation:

https://github.com/hzwer/arXiv2021-RIFE
https://github.com/hzwer/Practical-RIFE
https://github.com/nihui/rife-ncnn-vulkan
https://github.com/HomeOfVapourSynthEvo … cnn-Vulkan
https://www.svp-team.com/wiki/RIFE_AI_interpolation

Of course I wish you and myself to find someone who will find even better settings for the basic SVP algorithm, because it is still irreplaceable for 4K 10bit HDR files. However progress cannot be stopped and the best quality frame interpolation will be provided only by artificial intelligence.

Yes, as long as it is video with no more than 1080p resolution, 8-bit colour depth without HDR

Read this first:
https://www.svp-team.com/wiki/RIFE_AI_interpolation

Set the following values:

AI model: Generic (v4)
GPU threads: 2
GPU device: https://www.svp-team.com/forum/viewtopi … 170#p80170
RIFE via CUDA: Off

If you continue to have problems read what problems others had and how they dealt with them:

whole threads:
https://www.svp-team.com/forum/viewtopic.php?id=6580
https://www.svp-team.com/forum/viewtopic.php?id=6553

also read my thread from Chainik's post of April 28 to the end, everything should work from then on:
https://www.svp-team.com/forum/viewtopi … 360#p80360

skip it altogether: == RIFE / PyTorch installation == it's too slow now
https://www.svp-team.com/forum/viewtopi … 372#p80372

Did everything according to svp rife manuel. Fps stays the same as in the image i posted above.

I guess it works. Since estimated fps is 48. Can you check if it is the same on your mpv?

At the moment I don't have a RIFE compliant graphics card to run the tests.
I am relying on what the creator of Flowframes wrote a long time ago:

https://github.com/n00mkrad/flowframes/issues/40

and on the fact that already several people had problems with 10bit colour depth videos.

Did you manage to interpolate 10-bit video with RIFE in SVP?

If at x2 is a little bit smoother it means that interpolation is working and this should be the case for video with 8-bit colour depth. At the moment only one person with a GeForce 30 series graphics card has done a thorough test on a 1080p file: https://www.svp-team.com/forum/viewtopi … 477#p80477 and based on the fps we can assume that at the moment x2 for 1080p for your card is the maximum. If you want more: x3 or x4 then 720p files are left on the ground.

Perhaps there will be some optimisation of RIFE in the future. In any case, I estimate that I will need 116% more powerful graphics card than the ZOTAC GAMING GeForce RTX 3070 Ti AMP Extreme Holo to interpolate x5 1080p video files: https://www.svp-team.com/forum/viewtopi … 480#p80480 I suppose this will be more than possible for NVIDIA GeForce RTX 4090 graphics cards.

Oh well... artificial intelligence comes at a price.

I understand you perfectly. I've been saving money every month since 2019 to buy something really powerful. I had plans to upgrade my computer for 2020, but the situation at that time forced me to revise my plans: https://www.svp-team.com/forum/viewtopi … 343#p80343

...a year later, graphics card prices shot into space, and this year has brought further problems and massive inflation in most countries around the world.

Fortunately, the performance of RIFE over the past year has improved enormously https://www.svp-team.com/forum/viewtopi … 426#p78426 and there is a chance that it will also improve in the future, allowing owners of also more accessible GPUs to enjoy RIFE interpolation in better quality.

I also want, by showing the potential performance of NVIDIA GeForce RTX 4090, to encourage as many people as possible to test RIFE. The more of us there are, the greater the chance that artificial intelligence in video frame interpolation will not only be the domain of researchers, but will also have real-world applications in real-time interpolation, which SVP is famous for.