Does it work universally as foundational pre-trained models do?

[IoSonoMarco]

Do you need to train it on a specific dataset everytime? Or it is pre-trained so that it can accept any new out-of-the-box series of frames to do next-frame prediction on unseen videos?