AlayaWorld: Long-Horizon and Playable Video World Generation

AlayaWorld is an open-source, full-stack framework designed to move beyond labor-intensive game development by using generative video models to synthesize interactive, explorable virtual worlds. Unlike traditional game engines that require explicit authoring of assets and rules, AlayaWorld autoregressively predicts future visual states based on user navigation and prompt-driven actions, allowing for real-time, open-ended interaction.

The framework addresses the core challenges of interactive world generation—control, consistency, stability, and runtime—through a modular architecture. To manage camera control, it uses a hybrid approach: it maintains an explicit 3D cache that is rendered along the target trajectory to provide grounded visual evidence, while simultaneously using AdaLN-style modulation to inject camera pose information into the transformer backbone. For consistency, it combines this spatial cache with a compressed temporal history, ensuring that revisited locations remain visually coherent. To ensure stability over long rollouts, the model is trained with an error-bank mechanism that exposes it to drifted histories, teaching it to correct artifacts rather than compounding them. Finally, it achieves real-time performance through few-step distillation and a chunk-based generation strategy that supports on-the-fly prompt switching.

AlayaWorld provides a practical, reproducible foundation for researchers to explore generative world models. By unifying data preparation, training, and deployment into an extensible architecture, it lowers the barrier to entry for developing interactive systems that can be applied not just to gaming, but also to robotics simulation, embodied AI, and human decision-making studies. Its ability to generate persistent, explorable environments from text prompts or simple user inputs represents a significant step toward scalable, AI-driven virtual world creation.