The inputs are just photos from different positions. Then a neural net is trained so that you can ask to render a photo from any pose (including unseen). This

The inputs are just photos from different positions. Then a neural net is trained so that you can ask to render a photo from any pose (including unseen). This particular work (Zip-NeRF) builds on top of the original NeRF paper: https://www.matthewtancik.com/nerf https://www.matthewtancik.com/nerf (the website has a good explanation what NeRF aka Neural Radiance Fields are)