It supports spatial-content experiments where users move through a scene and expect previously visited locations to remain recognizable. Core capabilities include Image-to-world generation, Text-to-world generation, Camera-controlled exploration, and Implicit 3D-aware memory.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Viewpoint-conditioned memory retrieval, Minute-scale scene exploration, Static and dynamic scenes, and Few-step streaming generation.
