MoVerse Generates Navigable 3D Worlds from a Single Image at 8 FPS

Researchers from Youku Moku-Lab and HUJING Digital Media & Entertainment Group have published MoVerse, a three-stage pipeline that converts a single narrow-field-of-view photograph into a free-roaming, photorealistic video world in real time. The system first expands the input image into a full 360° panorama via topology-aware latent diffusion, then lifts that panorama into a persistent panoramic 3D Gaussian scaffold, and finally streams user-controlled camera trajectories through a causal video renderer — all at 8 FPS on a single RTX 4090.

The approach separates world construction (stages I and II, run offline) from observation rendering (stage III, run interactively), meaning the Gaussian scaffold can be reused across multiple traversals while maintaining geometric consistency. The paper, posted to arXiv (2606.13376), demonstrates results across a wide range of scene types, from outdoor landmarks to stylized environments, and could have direct implications for virtual production pre-visualization, game environment generation, and immersive media workflows.