If TVAI v3 is telling you 19 days, v2 may be faster but it is not orders of magnitude faster, so I would explore other options.
I would probably use ffmpeg scale filter for the spatial downsample (you can pick the algorithm you want) and Flowframes for temporal upsample. Maybe use Flowframes first, to maximize the amount of available detail for the upsample part. The default RIFE model should be fine.
For the ffmpeg scale filter, you could experiment with lanczos, bilinear, bicubic (or others). Lanczos can add some apparent sharpness, which may be good or bad, sometimes it looks like oversharpening when downsampling. Unless the source is very bad, Topaz’ strength is really in upsampling, not downsampling, so I think using it for what you’re trying to do is overkill and a waste of compute power (especially for your spatial 5k->4k objective).
Just like ffmpeg, Flowframes is free so you have little to lose in trying. Flowframes’ optical flow is very different but much faster than what Topaz does. For different kinds of footage I have observed it produces better or worse results than Chronos or Apollo, but always much faster so I always try it first. For interpolating a single frame between each frame pair, I think the results could be quite good with optical flow. Again, this depends heavily on what the footage is.
Regarding the other 99% of the ranting in this thread, I agree V3 needs a lot of work and was not ready to come out of beta, but here we are.
A common refrain is to “just add face enhancement.” I encourage everyone to try out current ML face enhancement tech themselves if they have the ability, most of the published approaches for face enhancement work on standalone images. This has been reiterated many times by other posters, but some readers just do not hear it. Without temporal coherence integrated into the model, the facial identity fluctuates wildly with changing aspect. This is very different from single image enhancement. If you want to give it a try, look up Codeformer or GFPGAN, there are easy to use GUIs that anyone could operate. The only hurdle is following the setup instructions on github. I mention these two because they are free, but of course you can try Topaz Gigapixel if you have it.
In any of these you can input either video clips or image sequences, and if a subject’s face is not too blurred and not moving, the results look good. A single frame of motion blur or a partial face creates the worst kind of single-frame result you can imagine – blurred mangled features punctuated by sharp eyes and misplaced teeth etc. True horror, in many ways worse than what is done by any of the face-un-aware models Topaz has. I think a better short term objective for Video AI would be integrating one of the readily available facial detection models into Topaz with a tick box to disable upscaling on faces (when the model truly fails), where a small blurred mask patch of the original sequence upscaled traditionally can be overlaid on the upscaled result (blurry face is certainly better than a sharp monster face, in the short term). The latest literature has a few algorithms describing approaches with attempts at temporal coherence, but they do not appear mature enough to be applied to general footage as Topaz is targeting. If you have a tremendous amount of compute power, and upscaling only specific types of footage with specific people, there are more options there, but again, that is outside what Topaz is aiming for so you’re basically asking them move beyond the envelope of cutting edge basic research. Don’t hold your breath.
By the way, I think the comments about “the team” are bordering on personal attacks, not cool. Settle down and have some sense of decorum, this behavior reflects poorly on the poster. From the outside you have little insight into what is going on with the team. They are not obligated to let you behind the curtain, as much as you wish for it. Besides, if you’re ranting about how you think the team is this that and the other thing, why would they be inclined to share MORE with you? You can catch more flies with honey than with vinegar, as they say. Anyway, that is just my two cents. I have been a long time lurker and it’s unfortunate to see the hostility growing in a few of the regular posters in these threads.
Back to the software, hopefully we can see some good enhancements in the next months, but I wouldn’t hold your breath. The next 6-7 weeks are very dense with holidays in the US where the development goes on AFAIK. This is obviously a good explanation for why there will be no new release next week. The developers are humans too and they have families and lives outside of feverishly trying to address every scatterbrained screed on the forums.