These were old photos, many taken with older digital cameras and early smartphones, and they worked very well. I believe as long as they're of a reasonably high resolution (so, any camera from the last 20 years or even more), it'll work. Of course, it's ML-based attempting to simulate 3D from a single image, so it's not always perfect, but I didn't run into any significant failures.
Btw, to capture spatial videos, the easiest way is to use an iPhone 15 Pro (with its camera in "spatial video" mode) or the Apple Vision Pro itself. I'd guess there are ways to convert videos from other formats into this, or to use other apps to view photos/videos in non-native formats, but I'm not personally familiar.
Photos are more up in the air in part because the use and support of the Apple-pushed HEIC is much more limited than HEVC. But internally it's basically the same thing as MV-HEVC except with just the two still frames encoded instead of full videos, so any codebase that supports MV-HEVC should be able to support the 3D variant of HEIC as well without a lot of extra work over the long term.