The photo in article is some random stock image and has nothing to do with the research paper[1].
[1] https://vpnoverview.com/wp-content/uploads/2301.00250.pdf
[1] https://vpnoverview.com/wp-content/uploads/2301.00250.pdf
Figure 8 shows a variety of "failure cases," but even these failure cases are surprisingly accurate estimations!
The standalone wireframe images sometimes manage to get the number of people and their approximate positions correct, but everything else about them seems (IMO) believable but not necessarily accurate, like most AI output.
This doesn't surprise me, because the input data is very coarse. Look at their graphs. There's just not enough raw data there to accurately reconstruct much other than position in the room.
I'm sure it's possible to do more accurate recording and reconstruction, I just don't think this demonstrates it.