Search options (Page 6 of 12)

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

https://github.com/hzwer/Practical-RIFE

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 are right, when evaluating efficiency for different power limits it is completely irrelevant.

I just forgot to write that with this one test I wanted to take care of 3 things at once, and that third thing I just didn't describe:

1. performance per watt at various power limits
2. coil whine at various power limits
3. estimation of the performance of future graphics cards based on the performance of current graphics cards

Until today we had no precise result on the forum for 1080p files for RIFE Model 4.0 and - RIFE filter for VapourSynth (ncnn Vulkan): https://www.svp-team.com/forum/viewtopi … 220#p80220
and a precise result in fps allows a more accurate estimation of the capabilities of future graphics cards.
Additionally more computing power is needed to interpolate 1920x1080 than for example 1920x800: https://www.svp-team.com/forum/viewtopi … 361#p80361   

Until recently I was only dreaming about 1080p x2 real-time interpolation, but after the last optimization I have gained an appetite for 1080p x5. I only need the information about the interpolation factor to estimate how close we are today to my goal:

x5 - I need 120fps (24 original frames + 96 interpolated)
x2 - I need 96 interpolated frames, so also 96 original, which gives 192 fps in total

So, if Your graphics card at 205W gives 89 fps (1080p x2) that means I need 116% more powerful graphics card!

Awesome! Later this year it should be possible!

As for coil whine You either have an exceptional unit: https://www.guru3d.com/articles-pages/z … ew,34.html or undervolting really helps here, which of course is often recommended to mute the card. In any case, it's good to know that with RIFE loading Tesor Cores can be quiet.

As I am also considering passive GPU cooling the information about coil whine is also very important to me. I know that this is a very subjective matter - one can hear it even with all the fans and HDDs humming, while the other cannot. With completely passive cooling of the entire computer and the use of SSDs, the issue of coil whine coming from the GPU is important. Especially since a constant, equal load on the Tensor Cores, is said to be a worst case scenario:

https://www.reddit.com/r/nvidia/comment … _with_rma/

So I would also ask for information on how different power limits affect coil whine when loading Tensor Cores via RIFE.

There are opinions that coil whine is particularly troublesome when the graphics card uses all its power. Hence my guess, that buying for example a GPU, which power section has the possibility to draw 600W and limiting this draw to 50% can significantly reduce the coil whine.

Of course from 1-2 meters distance coil whine may be not audible at all. In this case you can try to judge the coil whine from a closer distance, of course not so close that someone's hair gets pulled into the graphics card fan

This is not as important to me as the above efficiency test at different power limits, but if someone would check this by the way, I would really appreciate it.

Coming soon (probably September 2022) are new NVIDIA graphics cards with amazing Tensor Core capabilities, but also with unbelievable power consumption:

~900W TGP - NVIDIA GeForce RTX 4090 Ti
~600W TGP - NVIDIA GeForce RTX 4090
~350W TGP - NVIDIA GeForce RTX 4080

https://wccftech.com/nvidia-geforce-rtx … rds-rumor/

Without air conditioning, which I don't have, I can't imagine an extra 600 or 900W of heat in a small room in the summer. Of course there is a choice and we can always buy an NVIDIA GeForce RTX 4080 with 350W draw, but a more economical solution in the long run is to buy an NVIDIA GeForce RTX 4090 and drop the power down to, say, 300W or 350W. This way we can probably get an efficient graphics card in the sense of performance to energy consumed.

In addition, I wonder about passive cooling of such a card. Today it is possible to passively cool a graphics card up to 250W, although I believe that with some modifications it will be possible to reach up to 300W, especially if the card itself has a design suitable for 600W.

Until now, I believed that lowering the power consumption of a graphics card always leads to an increase in efficiency: relatively large at first then less and less.

However, I have now found some tests that confirm this, but also show that once the power consumption is lowered below a certain level, the efficiency starts to drop dramatically. The tests indicate that this depends not so much on the architecture of the specific graphics card, but mainly on the software used for testing. This can be seen in the graphs below:

Hardware: Inno3D GeForce RTX 3090 Ti X3 OC
Software: Control

https://www.hardwareluxx.de/index.php/a … l?start=23

Hardware: MSI GeForce RTX 3090 Gaming X Trio
Software: Heaven Benchmark

https://www.forum-3dcenter.org/vbulleti … st12963231

I must admit that these results worry me a little, because they indicate that lowering the power consumption to 50% does not necessarily mean an increase in efficiency. The consolation is the fact that, as we can see on the graphs, a lot depends on the software. So I am very interested to see how it looks with RIFE, which uses Tensor Cores.

1. Try to use GPU thread: 2 - this gives 3x better performance, see blackmickey1007's post:
https://www.svp-team.com/forum/viewtopi … 219#p80219

2. Try to use 8-bit video, RIFE does not natively support 10-bit video.

3. Don't use madVR upscaling, it takes away GPU resources needed for RIFE.

4. Try to use x2 interpolation first, don't go crazy with that 360Hz

5. List all GPUs you have in your system: integrated and dedicated - read what Chainik wrote here: https://www.svp-team.com/forum/viewtopi … 170#p80170

6. Screenshots are always welcome on this thread

Exactly what Chainik wrote

https://developer.nvidia.com/cuda-gpus
https://developer.nvidia.com/vulkan-driver

6912x3456 59.94fps x2

My guess is that you need faster RAM. Maybe even quad channel...

I'd like to thank everyone again for the tests and explain again why I only ask and don't test anything myself.

The answer is simple: I don't have a suitable graphics card for testing, and my temporary PC is not suitable for the latest graphics cards anyway.

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. And maybe it's even better, because in the meantime the RIFE interpolation algorithm based on machine learning (artificial intelligence) appeared and it became clear that now the graphics card will be responsible for the best interpolation of frames in real time.

I'm now looking forward to COMPUTEX 2022, which will be held at the Taipei Nangang Exhibition Hall from May 24 to 27, 2022, where we may find out more about many new products that will be launched in the second half of this year.

It is in the second half of 2022 that I now intend to build my new PC completely from scratch. Then I will share very detailed test results using SVP & RIFE on the latest hardware that will be available then.

Until then, I would be grateful to everyone for sharing their tests here on the forum, so as to provide feedback to the SVP and RIFE developers, so that we can look forward to even better frame interpolation in the future, and to other testers, so that there are even more of us.

Special thanks to those involved in development:
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

and to all those testing the implementation of RIFE in SVP here on the forum.

→

https://github.com/hzwer/Practical-RIFE/issues/12

https://paperswithcode.com/dataset/atd-12k

In summary, the 4.0 model is trained on the largest animation triplet dataset.