The emergence of the metaverse has ledto the rapidly increasing demand for the generation ofextensive 3D worlds. We consider that an engagingworld is built upon a rational layout of multiple landuse areas (e.g., fores...The emergence of the metaverse has ledto the rapidly increasing demand for the generation ofextensive 3D worlds. We consider that an engagingworld is built upon a rational layout of multiple landuse areas (e.g., forest, meadow, and farmland). Tothis end, we propose a generative model of landuse distribution that learns from geographic data.The model is based on a transformer architecturethat generates a 2D map of the land-use layout,which can be conditioned on spatial and semanticcontrols, depending on whether either one or bothare provided. This model enables diverse layoutgeneration with user control and layout expansion byextending borders with partial inputs. To generatehigh-quality and satisfactory layouts, we devise ageometric objective function that supervises the modelto perceive layout shapes and regularize generationsusing geometric priors. Additionally, we devise aplanning objective function that supervises the modelto perceive progressive composition demands andsuppress generations deviating from controls. Toevaluate the spatial distribution of the generations, wetrain an autoencoder to embed land-use layouts intovectors to enable comparison between the real andgenerated data using the Wasserstein metric, which isinspired by the Fr´echet inception distance.展开更多
文摘The emergence of the metaverse has ledto the rapidly increasing demand for the generation ofextensive 3D worlds. We consider that an engagingworld is built upon a rational layout of multiple landuse areas (e.g., forest, meadow, and farmland). Tothis end, we propose a generative model of landuse distribution that learns from geographic data.The model is based on a transformer architecturethat generates a 2D map of the land-use layout,which can be conditioned on spatial and semanticcontrols, depending on whether either one or bothare provided. This model enables diverse layoutgeneration with user control and layout expansion byextending borders with partial inputs. To generatehigh-quality and satisfactory layouts, we devise ageometric objective function that supervises the modelto perceive layout shapes and regularize generationsusing geometric priors. Additionally, we devise aplanning objective function that supervises the modelto perceive progressive composition demands andsuppress generations deviating from controls. Toevaluate the spatial distribution of the generations, wetrain an autoencoder to embed land-use layouts intovectors to enable comparison between the real andgenerated data using the Wasserstein metric, which isinspired by the Fr´echet inception distance.
