Hugging Face has released a technical report on AlayaWorld, an interactive long-horizon video world model designed to generate immersive, explorable virtual environments from user inputs. Unlike traditional game development, which requires extensive manual work for assets, physics, and programming, AlayaWorld creates interactive worlds instantly from text, images, or video.
The model is built on a 15-billion-parameter video diffusion transformer and generates 24-fps video at resolutions of 540p and 720p. It operates autoregressively, producing short latent chunks guided by camera trajectories and switchable text prompts. To maintain spatiotemporal consistency over long horizons, AlayaWorld employs a bounded visual context that includes a persistent sink frame, compressed temporal history, geometry-aligned spatial memory, and recent-frame conditioning.
To reduce long-term drift, the model is trained with corrupted histories and prediction residuals collected from its own roll-outs. Additionally, a discrete autoregressive distillation formulation is introduced, combining distribution-matching distillation, self-forcing++, and consistency distillation. This reduces inference from approximately 30 sampling steps to just four steps per chunk, significantly improving efficiency.
On the iWorld-Bench benchmark, AlayaWorld achieves the best performance for long-horizon generation. The project is conceived as a full-stack, open-source, and long-term initiative, intended to provide an extensible foundation for future research on interactive video world models.