A world model that remembers a scene but forgets which instruction applied when can easily make later actions leak into earlier context. Hy-Nomas addresses this by separating persistent scene description from time-bounded actions—a useful design choice for video generation that unfolds through successive user turns, rather than a single prompt.
Key Findings
- Action-aware causal attention keeps descriptive context available across segments but restricts action instructions to their annotated temporal windows. This helps later turns build on a world without treating every past command as permanently active.
- A Prompt Enhancer turns user requests and interaction history into structured conditions for the next segment. That moves some of the planning burden from the video generator to an explicit vision-language component.
- Text and camera control share one continuing world. Users can alternate between describing changes and steering the viewpoint, with cached history conditioning variable-length segments.
- Training targets the model’s own predicted history. A staged curriculum introduces causal generation, text interaction, camera control, and generated histories before diffusion distillation, aiming to reduce the mismatch between training context and inference-time continuation.
Who It’s For
A good fit for researchers studying interactive video generation, persistent visual context, and controllable world models. The paper reports stronger human-evaluation results than existing world models, but this should not be read as proof of a physically grounded simulator: it generates video, and visual continuity does not by itself guarantee reliable physics or action consequences. Look elsewhere if you need a general-purpose simulator or precise, verifiable control over every frame.