Interactive worlds that respond to typed commands and camera moves are hard to build. Existing models often forget what happened or fail to follow a series of instructions.
Hy-Nomas uses a diffusion model, a type of generative AI that creates video from random noise, with a special attention mechanism. This mechanism keeps descriptive context (like 'the room has red walls') available across turns, while action instructions (like 'move the chair left') apply only to the moment they are given.
In human evaluations, Hy-Nomas outperformed other world models on consistency and instruction following. This means smoother, more natural interactions in virtual environments for gaming, design, and training.