New AI framework Stepper generates immersive 3D scenes from text using stepwise panoramic expansion with improved fidelity and consistency

Stepper addresses the visual fidelity vs. explorability trade-off in 3D scene synthesis by using stepwise panoramic expansion instead of autoregressive methods

Features a novel multi-view 360° diffusion model that enables high-resolution, consistent expansion of immersive scenes

Includes a geometry reconstruction pipeline that enforces geometric coherence across generated scenes

Trained on a new large-scale multi-view panorama dataset to achieve state-of-the-art performance

Applications span AR/VR and world modeling with potential for text-driven immersive environment creation