This seems impossible: creating a detailed 3D scene from just a few photos. I wonder how landscapes would come out; the trees look good so landscapes should also work (I think).
https://www.youtube.com/watch?v=dZ_5TPWGPQI
Looks incredible, almost like a hoax.
I've thought about that too but it seems well founded. There is even a link to a "paper". How an AI can conjure up so much in such a correct way is beyond me.
Maybe it is a hoax and they've done it the other way round; first make a detailed movie, then take out some photo's to fool us all ;D
Maybe less a hoax and more just impractical. Like the infinite detail voxel engine that has only been used a handful of times despite being very powerful.
That's pretty cool! They share their code as well, but the setup looks pretty laborious. Linux only too. Almost makes me want to go back to dual-booting, but not quite.
https://github.com/darglein/ADOP
Come to think of it, to draw an object, 6 photos as a reference are usually enough to cover all sides, as opposed to hundreds needed for photogrammetry. Of course, there are situations where 6 photos are not enough, such as objects with holes and weird bulges. Like trees for example. However the trees in the video look good as far as I can tell. Reflections in windows on the other hand do not work, which is logical since there is hardly any information available about them.