Interactive, history-aware world models are a step toward controllable generated environments.
A paper introduces Hy-Nomas, an 8B diffusion video world model that follows successive free-form text commands and continuous camera control across multiple interaction turns, generating variable-length segments conditioned on cached history.