This is a Plain English Papers summary of a research paper called Hy-Nomas Uses Causal Attention for Multi-Turn Video World Control. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.
Hy-Nomas turns video generation into multi-turn world control
Hy-Nomas is an 8B-parameter diffusion world model that generates variable-length video segments from free-form text, camera controls, and cached interaction history. Unlike a conventional video generator that receives one prompt for a fixed clip, it accepts commands across successive turns and aims to preserve their visible consequences as the scene continues.
The system’s central mechanism is action-aware causal attention aligned with structured video captions. Persistent descriptions of subjects and scenes remain available across segments, while action instructions are confined to their intended temporal windows; a vision-language Prompt Enhancer converts the latest request and observed history into a structured prompt and segment duration.
The interaction loop combines semantic direction with camera navigation: users can request subject actions, object interactions, scene changes, or camera motion, then continue from the generated result. Hy-Nomas carries visual and semantic context through an autoregressive key-value cache, with the Prompt Enhancer preparing conditions for each next segment....
