An open World Model holds one minute of gameplay without drifting

Alaya Lab's open-source AlayaWorld framework produces interactive 720p environments. A method designed to eliminate visual drift.

AlayaWorld, published by Alaya Lab, presents itself as a comprehensive open-source framework for building interactive video worlds. The principle reverses the traditional creation of a game environment, where objects, animations, and interaction rules are specified one by one: here, the model synthesizes the subsequent observations from the current state and user actions, in 720p at 24 frames per second. Users can navigate freely, trigger a fight, cast a spell, or spawn a creature, with the scene holding up beyond a minute of continuous generation.

Inference starts from an initial frame image, a camera trajectory, and a prompt, then unfolds the video in blocks of about one and a half seconds, with around forty blocks sufficient to cover the minute. Real-time performance is achieved through few-step distillation and these short blocks, with prompt changes occurring at the junctions to limit latency.

The most notable technical point concerns drift, the progressive degradation that plagues long-duration generation. Where others typically seek to prevent it, the team trains the model to absorb it. A mechanism called "error bank" reinjects accumulated artifacts from already-drifted histories during training, both into the memory and the generation target, so that the model is no longer destabilized by its own past mistakes. This is complemented by a 3D cache and a history compression module, designed to maintain the consistency of a location that is left and then returned to.

Trained on gameplay recordings and real-world videos, the model also targets non-gaming use cases, particularly embodied AI and simulation for robotics. The inference code and weights are available, with the training code and datasets announced for later. Alaya Lab claims a state-of-the-art level of performance.