It looks cool, though I still find myself looking for Hyperlapse[0] alternatives.
Hyperlapse seemed like magic when announced, particularly, not needing to crop large parts of shaky videos
[0] https://blogs.microsoft.com/ai/hyperlapse-siggraph-2014/
[0] https://blogs.microsoft.com/ai/hyperlapse-siggraph-2014/
I think Hyperlapse Pro [1], which stitched patches from multiple frames together to create new frames after doing some 3D reconstruction and then recreating a smoother camera path was mostly replaced with Hyperlapse mobile [2]. The Pro version was really slow, but the mobile version, which used the original frames from the video, could run in a reasonable time on a phone with almost as good result. Hyperlapse Pro is still available to download [3], Hyperlapse mobile you have to find the APK elsewhere [4].
[0] https://news.ycombinator.com/item?id=36348483
[1] http://johanneskopf.de/publications/hyperlapse/first_person_...
[2] https://www.microsoft.com/en-us/research/wp-content/uploads/...
[3] https://www.microsoft.com/en-us/download/details.aspx?id=523...
[4] https://www.apkmirror.com/apk/microsoft-corporation/hyperlap...
This is fine for gopro footage or stuff like that, but for production video, the aim is to retain as much information as possible until export. I think the level of artifacting is just unacceptable for general stabilisation needs - and with today's cameras being 4k and up, framing 10-15% wider to adapt to the stabilisation crops makes much more sense.
Uses Neural Radiance Fields