Reference frames from content aware fill do not blend between keys

Pretty much as I suspected. There's not enough details (in the background screens) for the algorithm to really do much and the AI may not have enough samples for such scenarios in its database. Tons of reflections don't help, either. As is evident from your video the "dumb" version without reference frames looks a lot better. That said, the best possible workflow would probably be a classical 2D track and patching over the contestants with duplicates of the layer, masking and importing doctored stills of the studio as "clean plates". The other option might be to just accept the flaws in the CAF-based approach, but make the areas larger and then mask out those so that underneath on a duplicate layer you can recover the speaker desks. In any case you will have to put in more work. This is nothing that will be doable fully automatically.

Mylenium

Show us screenshots and provide soem info about your project and system as Rick already said. Either you are trying to do too much in one sweep or the fotage has some specifics that prevent CAF from doing its magic and you may need to use other, more traditional techniques.

Mylenium

CAF is not intuitive. Have you gone through the tutorials? You have to wait, sometimes wait a long time for the fill to be generated. Without specific workflow details, system and AE version info, and screenshots showing the modified properties of the problem layers, it is impossible even to make a wild guess.