Greetings everyone,
It’s been a while since I posted, and I am working on another video that helps players perform a self-analysis of their serve (using the Kovacs 8 stage model). As part of the process I did some slowmo recording on my phone (Samsung A52) and decided to try some upscaling tests using different AI algorithms (since slowmo is only available @ 720p).
Here is the video:
I haven’t gone into major detail yet on the video itself because this isn’t a scientific test or anything like that. Rather, I am curious which video(s) produce the best results on a bigger screen and am looking for feedback from people with good eyesight! I also used the YT trick of exporting and uploading at 1440p and I think it is probably best viewed at 1080p, but that may or may not be true.
I await your thoughts on the quality and not the serve itself!
TIA.
I would like to recommend my upscaling method.
There is one very good cross-platform and free program Upscayl, it works with individual frames (jpg, png, webp). It can improve images in batch mode, many frames at once. It cannot improve video files directly, you need to save the video clip as a series of individual frames, shotcut allows you to export video as many png files, after which they can be improved and loaded back into shotcut as an animation sequence.
Surprisingly, according to @MusicalBox’s assessment of the different outputs there is almost no point using Topaz AI for upscaling for tennis video.
I could see a small difference between them, but couldn’t really say which was better (due to my poor eyesight). Since the very slight improvement is negligible compared to the original 720p recording and it takes between 30 minutes and an hour to do its thing, I expected to see a better result. Especially since Shotcut natively did almost the same job with no special AI.
I will upload and link the original 720p recording shortly, since it’s possible on a TV the built in upscaling is as good or better than the software I used. I’m also going to try Upscayl and then Video2X, but iirc Video2X takes a LONG LONG LONG time to process for some reason, and I never saw much difference either.
Update with videos shortened and a several more algorithms:
Here is the original 720p, which may well play better on a TV using the built in upscaling:
Upscayl looks interesting and similar to Gigapixel Image AI. But since it was going to take longer than Video2X using the RealSmooth option and Video2X extracts every frame as a png and then tries to do its thing I just left it at that.
I personally think Videoproc is “ok” but I’m not really qualified to comment due to vision issues.
For me, watching on a 4k TV, the original 720p recording was as good, or almost as good with the built in TV upscaling, AND playing back with Kodi. Disappointing.
I have long since passed the euphoria about “smart AI upscaling”. I have a lot of digitalized video recordings from my childhood, shot on an old analog video camera, the quality of these videos by today’s standards is terrible, I tried to apply all existing methods of improvement and nothing could help me with this. I would not waste time on this now. The only content that upscalers cope with almost perfectly is drawn cartoons, their quality really becomes fantastic.
I think it also depends a lot on the TV. Some newer TVs have machine learning based super upscaling, for example. But also their high contrast, size, and default sharpness can accentuate poor quality value. If you upscale in software I guess you have more control.
