This time, we have two new models for frame interpolation - Apollo A (no deblur) and Apollo B (with deblur). All Apollo models including the new ones do non-linear frame interpolation while all Chronos models do linear frame interpolation. The new Apollo B model and the existing Apollo model also reduce motion blurs along with the interpolation.
Previous alphas already introduced new Themis models for motion blur reduction. In this alpha, we have optimized their speed a little bit but still a long way to go.
In general, the new models are not optimized and could be VERY SLOW. We are working on it.
Please give them a try and share your feedback on the quality of the outputs. Thanks.
I wish larger performance gains could be had without having to go the parallel processing route but this might end up being the way…you are on to something here
And instead of the hard coding way you could do it the easy way and just start several instances of TVAI via a script, depending on the need (like with the CLI).
The video is simply divided by two, three or four (time-wise), depending on how much memory the user has available.
Just downloaded.
Having same issue using stablisation as i am having on release 3.0.6.
when running Stabilize on GoPro footage, I can preview (and it looks great) but when exporting to ProRes, 264, 265 (AMD or nVidia) the output video is a strange grid - see attached
Hello Nippon.
Process all the versions and you have the files in Dropbox.
I did not have time to write an error report and in the next few days I will send it to you,
I work as a professional in a VFX video room and I have the most recent image control elements,
The luminance undergoes severe changes in this first test.
The new interpolations in the tests outperform Cronos, avoiding serious blurring at luminance peaks.
Good job at the beginning of this Beta.
I’m concerned about Dione’s upgrade from deinterlacing to progressive.
Professional studios have thousands of hours of interlaced video waiting for improvements to this algorithm.
I’m not sure if this has been addressed before, but why even in full screen mode, all the edges can still be resized? If I put my mouse on any of the 4 edges or corners, I can see the resize cursor. And when I go to the very top right, trying to close the app, I can’t. It shows the resize cursor, so I have to move the mouse a bit.
It’s the same on Windows 10. I think they are using something called ‘QT framework’ to build the GUI and either the main window properties haven’t been set correctly (don’t allow resize when maximized), or standard windows behaviour isn’t supported by QT.
I am going to assume it’s the first case - them needing to set the main window properties to behave like a standard application in Windows. Perhaps one of the devs can comment on this. @suraj@nipun.nath
Test v3.1.0. sent DropBox Yazan Saradest.
I add three tests with original 4.2.2 10bit.
interlaced video.
For the tests it would be useful to have a version of Dione deinterlacing, to provide objective assessment.
Greetings .
“Quote:
Be careful with more than two queues if hardware-accelerated GPU scheduling is disabled. Software scheduled workloads from more than two queues (copy queue aside) may result in workload serialization.”
Wow such a speed improvement in the Themis models in this new Alpha. I went through 0.44s/f to 0.08s/f on a SD 640X480 video. And Themis A is not slower, get around the same speed maybe +0.01s/f. On RTX 3090.
I saw this new alpha release just right after I spend 2h15m on a 10 minutes video when it would have just taken 18 minutes with this new Alpha. And it seems to also produce better results.
Apollo is of course much slower with around 0.33s/f
Motion is not constant. There is motion discontinuities in real life videos. Linear tends to smooth the motion and giving sometimes unrealistic motion (sometimes called the soap-opera effect).
Non linear will better handle motion discontinuities and will produce a more realistic output.
Okay, just tried 3.1.0.2m. I tried Themis A, and well, as announced, it was very slow. But not for the reasons I expected (I expected 100% CPU load, but my CPU is hardly getting taxed at all, at only like 28%; and I don’t have it at reduced system load).
TVAI is notoriously bad at efficiently utilizing cpu/gpu. I’ve yet to ever see it use more than 60% of either, regardless of input video, model, or fine tuned settings. Normally it uses 10-30% of only one (cpu or gpu) and basically never both.
The software is PITTIFULLY optimized and the only way to bypass that is to split up your video files manually, process them in separate instances, and merge them back into on file manually again. It’s awful.
on a SD 640X480 video. And Themis A is not slower, get around the same speed maybe +0.01s/f. On RTX 3090.