Topaz Video AI v3.0.9

And actually going from 25p to 50p creates these same frame blends between scene changes. This makes it unusable without babysitting and double checking every single scene change

Thanks for informing me. Iā€™m not sure why that would be the case since FFMPEG supports AviSynth as an input format. Regardless, thank you for your response. I can always fallback on using AVFS to make a dummy AVI file in real time. Iā€™m not familiar with VapourSynth, but maybe there is something similar to AVFS for it.

1 Like

Itā€™s already listed in the feature requests section: Face Reconstruction (Refinement) for small faces

1 Like

Another issue with face refinement is if they implement it, expect your processing to take weeks not days. Using face refinement of an image in Photo AI takes much longer than just noise removal, sharpening or upscaling.

Imagine how much extra work it takes to look at every frame, detect and refine faces. Even if it worked well it would be such a performance hit the software would be unusable until we have quantum computers on our desktops! Another reason to scrap this idea (in my opinion)!

1 Like

Still just sipping the GPU power. Just a 1080p upscale and clean up of a 480p clip.

Pic1 is during and Pic2 is the averages for the upscale run.

during
average

Itā€™s like my GPU just cant get out of 1st gear with this software. The usage of the GPU goes up and down in fits and starts all the time. Bizzare.

1 Like

Not surprising actually. VAI currently does a very poor job of utilizing hardware resources. In the roadmap post itā€™s mentioned they are working on improving it. Letā€™s hope!

3 Likes

With the new 3.09 release I canā€™t use the chronos fast model to convert a video file from 50fps (after deinterlacing) to 25fps. This action was possible with 3.0.6.
I get this error message:

Last FFmpeg messages:

Unable to parse option value ā€œ0ā€ as video rate
Unable to parse option value ā€œ0ā€ as video rate
Solutions?
Many thanks

Oh I hope so. I thought GPU-Z was a good metric to show the usage pattern during a run. Showing how the usage just goes up and down constantly and never ever hits the buffers. It will hit say a max of 2000Mhz for a few seconds and drop to 10MHz for a few then 300Mhz, even though I have the min clock limit set to 850MHz.

1 Like

Hi David, I donā€™t think that existing enhancement request is the same. Almost the complete opposite.

Thanks.

Andy

1 Like

No doubt quality has its price, but to resize a video quick and dirty I donā€™t need AI. That can be done quite well by traditional programs.

2 Likes

Thanks for reminding me. :slight_smile: Thereā€™s nothing TVAI can do to prevent the use of the AviSynth File System: itā€™s fully transparent.

About face refinement in video. Interestingly there is a program called HitPaw AI Video Enhancer, that can do face refinement in video, they also have a photo app, and when applied in video I think in some of the tests Iā€™ve done it takes about two times longer than simply enhancing video with no face refinement model. But its not many more times slower. While HitPaw is no where near as fast and as optimized in regular use as Topaz is, the difference between no face refinement and with it turned on, in video, does not take more than twice as long from what I was able to see. If they could do it, Iā€™m sure Topaz can optimize their AI models as well.

1 Like

I agree wholeheartedly. Your last 2 sentences were spot on. Time to put a little ā€˜iā€™ next to the ā€˜Aā€™, the way it was meant to. :slight_smile:

Ah, slowly version 3 starts making fun again. just updated directly from 3.0.6 to 3.0.9 on Mac M1. UI is now super fluid, even preview is fun again without waiting and waiting and waiting.

Feels somehow as fluid as I remember the 2.6.x versions - great work dear devs !

Sidenote: If you could consider ro bring the 4x preview from 2.6 backā€¦I loved the option to parallel display 4 different models ;o))

However - First impressionā€¦great work with this release

2 Likes

:grin: :grin: :grin:

The Reduce Noise option in Proteus seems to be reduced by about 3/4ths. I tried it out on a video or two that I had previous (3.0.4 or so) would be around 25-30 is now 90-100 in 3.0.7 and above.
edit I did want to say that the QUALITY of the noise reduction has increased, noticeably too. I like my movies and stuff without a lot of film grain, but hate any plastic look, and this is also getting better and better.

I still show hardly any real GPU usage (RTX3080) maybe like constant spikes to 30% or so, but CPU usage was consistently at like 80%+ (Ryzen 7 5850X) so much so I had to turn it on low power mode to use my pc at all. Turning it to low power mode I went from maybe 1.2fps to .5fps or so (each version fps seems to be getting better).

Have not finished a video in 3.0.9 yet, but all versions after 3.0.5 when the video finishes Video AI just hangs and I am left two two files TVAI one and Probe.

Love how quickly 3.x is improving!

2 Likes

I tested hitpaw and the other Topaz rip offs and the results were dreadful.
They are just gui frontends using BSRGAN.

5 Likes

Iā€™ve had this happen too, once. Wasnā€™t happy (nothing is more frustrating than a file failing at 99.99% progress). Had to kill the process. Good thing probe file apparently already had all its headers written, or youā€™d seen me here on the forum about it. :joy:

2 Likes

Check your Pcie-link speed. Mine RTX is maxing out (currently 8x in shared). I donā€™t know if it is a good indicator?

its x16. Latest drivers. Undervolted and will boost to 2700Mhz in games and other applications. ULPS is switched off too.

1 Like