Topic: AI interpolation implementation in SVP in the future?

May be off topic but recently I saw this AI interpolation methods
https://hackaday.com/2020/09/20/boost-y … -using-ai/

It seems to produce way better results than SVP currently, but the method is very time consumers and generate huge temporary files (basically extract every single frame as picture first).  I wonder if there is a plan for SVP to adopt some of the methodology in the future?

Re: AI interpolation implementation in SVP in the future?

DAIN's free, you can just use it instead if SVP isn't producing the results you want for transcodes, BUT...
Bro DAIN's SSLLLOOOOWWWWW.

You can use it to make amazing short clips, but whole movies, whole episodes, whole SERIES?  I can't even imagine how long that'd take, but we're talking weeks & months, here.
If SVP doesn't work on an AI implementation maybe wait for Topaz Labs to come up with something.  Their Video Enhance AI software is bonkers, and I get 0.40-0.80fps upscales (could be better, but not bad) with an aging 980Ti, so I bet they could work some magic and get AI interpretation up to speed.

Also minor complaint but DAIN really should support more FFmpeg codecs than it does.  Converting to freaking APNG, then slowly unpacking the APNGs to PNGs...  SVP supports ProRes, FFV1, etc.  All the good intermediate codecs.
And DAIN only does exact multiples, so if you want 60fps and the video is 23.976 you have to waste way more time converting to 95.904, and DAIN is slow enough already without being forced to go overboard.

I use SVP + ReelSmart Motion Blur and my transcodes look smooth enough for my tastes, with no artifacts.

3 (edited by getitall 31-12-2020 21:44:57)

Re: AI interpolation implementation in SVP in the future?

Thanks, finally someone knows what i am talking about,  exactly! DAIN doesn't even utilize the tools already available like ffmpeg, extremely slow (more like a academic experiment for now than a usable software).  And SVP devs seems not interested in this direction, while current methods always lead to bad artefacts for fast moving scenes.

How do you combine  ReelSmart Motion Blur  with SVP?  i looked up realsmartmotion blur and it looks like a very expensive plugin for premier pro?

4 (edited by MichaelR 05-01-2021 21:05:30)

Re: AI interpolation implementation in SVP in the future?

You just import the output from SVP into Premiere Pro or After Effects or whatever other program RSMB works with, then render that out. (I like the free Voukoder plugin that lets you output video to FFmpeg)

SVP, even with strong masking, tends to just create motion ghosts rather than actually compensate artifacts with realistic motion blur.  Only DAIN offers a third option, that being no artifacts AND no motion blur, but we just talked about the drawbacks to DAIN.

Admittedly $75 for NeatVideo Noise Reduction is a good deal considering how great it is.  And although I got Video Enhance for $160, it's honestly worth at least $200 IMO...  RSMB feels like it's worth $20-$50, so it's too bad they charge $100 for it.  RSMB does a good job turning the masking from SVP into realistic motion blur.  Only mistake it might make is if you have motion behind a watermark, it can blur into the watermark, but you can use workarounds to fix that.

The company that makes RSMB also makes Twixtor, which is their take on SVP and Adobe's Optical Flow, but honestly, SVP, Optical Flow, Twixtor, they're all about the same quality, whereas at least SVP lets you tweek it to your preference, plus SVP's WAY cheaper, and even works on Linux, and does realtime, etc.

Re: AI interpolation implementation in SVP in the future?

Thanks, i found new apps that is faster than DAINE and great results, it's RIFE,
https://github.com/hzwer/arXiv2020-RIFE

and FLowFrames
https://nmkd.itch.io/flowframes

Try it out.

AI interpolation is way much better than vector based on like SVP, I like SVP for it's simple usage so i hope SVP quickly shift to the AI interpolation methods before it becomes obsolete.

Re: AI interpolation implementation in SVP in the future?

> RIFE

interesting, but still only "30+FPS for 2X 720p interpolation on a 2080Ti GPU"

7 (edited by Blackfyre 25-01-2021 20:38:16)

Re: AI interpolation implementation in SVP in the future?

Chainik wrote:

> RIFE

interesting, but still only "30+FPS for 2X 720p interpolation on a 2080Ti GPU"

Yeah I laughed when I saw that, still a long way to go. Even nVidia Optical Flow still produces more artifacts than SVP in real time. Hopefully nVidia Optical Flow 3.0 has big improvements in terms of artifact masking, even then we'll still use it with SVP + MPV.

EDIT:

Just to make it clear, I am in support of AI Interpolation and it will likely be the biggest advancement in motion interpolation. But we're still not close to it becoming a mainstream in regular systems. Even some TV manufacturers are using their own methods starting this year I believe.

8 (edited by UHD 24-03-2021 21:13:31)

Re: AI interpolation implementation in SVP in the future?

I can pay $100 for SVP 5.0 beta with REAL TIME RIFE support. Even though I already paid for the current version.

You could even develop two versions at once: a standard SVP and an SVP AI version for super powerful GPUs.

2X 720p or 2x SD with the best interpolation algorithm real time is enough for $100. I'm serious. I hope I'm not the only one who thinks so.

If not the creators of the SVP then who can do it? Just a rhetorical question, of course.

If already 2080Ti can interpolate 2X 720p with 30+FPS, what can do such monsters as 3090? And this is just the beginning. After all, the algorithm will be optimized all the time: „2021.2.9 News: We have updated the RIFEv2 model, faster and much better!” https://github.com/hzwer/arXiv2020-RIFE

Also, there will be a strong emphasis on AI algorithm support in the new GPU models.

And finally, the most important: SVP with real-time RIFE support can provide better results and less artifacts than the motion interpolation algorithms in the best TVs!

9 (edited by yongbin 24-03-2021 23:39:47)

Re: AI interpolation implementation in SVP in the future?

If there are two versions, I am willing to pay for more advanced version, as long as the perfect transcoding effect is good, willing to spend more time

10 (edited by MichaelR 10-07-2021 04:53:39)

Re: AI interpolation implementation in SVP in the future?

yongbin wrote:

If there are two versions, I am willing to pay for more advanced version, as long as the perfect transcoding effect is good, willing to spend more time

I called it earlier that Topaz Labs would probably make their own frame interpolation model, and just recently they did!
Considering how much better Video Enhance AI is compared to RealSR and Waifu2X (For non-cartoons), It's worth comparing it to RIFE or DAIN.

Neither of these 3 solutions are real-time; only SVP does that.  Video Enhance AI has a trial, and you can just uninstall it, make a new account, and keep using the trial as long as you want.
I did actually buy a license, but that was only after like 10 trials, lol.

I did a quick test and it's slow, but not DAIN slow.  IDK if it's as fast as RIFE, but It might be better quality.  One thing I did notice is it doesn't seem to require crazy amounts of VRAM to do 1080p, which I know DAIN did and required you to run it in tiles/rows.  Since I called it they've added tons of models to Video Enhance AI, like an AI model for interlaced footage, and a new one that you can use for anti-aliasing.  Still no replacement for Waifu2X for cartoons and anime, though; all of these models create weird artifacts with cartoons and anime upscaling.