Search papers, labs, and topics across Lattice.
The paper introduces Map2World, a framework for generating consistent and controllable 3D worlds conditioned on user-defined segment maps of arbitrary shapes and scales. A detail enhancer network is proposed to add fine-grained details while maintaining global coherence by incorporating structural information. By leveraging strong priors from asset generators, Map2World achieves robust generalization across diverse domains, outperforming existing methods in user-controllability, scale consistency, and content coherence.
Forget grid layouts: Map2World lets you generate consistent 3D worlds from arbitrary segment maps, offering unprecedented control and scalability.
3D world generation is essential for applications such as immersive content creation or autonomous driving simulation. Recent advances in 3D world generation have shown promising results; however, these methods are constrained by grid layouts and suffer from inconsistencies in object scale throughout the entire world. In this work, we introduce a novel framework, Map2World, that first enables 3D world generation conditioned on user-defined segment maps of arbitrary shapes and scales, ensuring global-scale consistency and flexibility across expansive environments. To further enhance the quality, we propose a detail enhancer network that generates fine details of the world. The detail enhancer enables the addition of fine-grained details without compromising overall scene coherence by incorporating global structure information. We design the entire pipeline to leverage strong priors from asset generators, achieving robust generalization across diverse domains, even under limited training data for scene generation. Extensive experiments demonstrate that our method significantly outperforms existing approaches in user-controllability, scale consistency, and content coherence, enabling users to generate 3D worlds under more complex conditions.
Microsoft Research