Across each demo, Diffusion Forcing acted as a full sequence model, a next-token prediction model, or both. According to the researchers, this versatile approach could potentially serve as a powerful backbone for a "world model" one day.

→ View original post on X — @mit_csail